This action might not be possible to undo. Are you sure you want to continue?
Seongju Chang School of Architecture Carnegie Mellon University
Ph.D. Committee Prof. Ardeshir Mahdavi, Ph.D.(Chair) Prof. Volker Hartkopf, Ph.D. Prof. Sebastian Thrun, Ph.D. Prof. Bernd Bruegge, Ph.D.
Table of contents
Table of contents..................................................................................................................................................i List of Figures.....................................................................................................................................................ii List of Tables......................................................................................................................................................vi Acknowledgment................................................................................................................................................vii Copyright Declaration......................................................................................................................................viii Abstract...............................................................................................................................................................ix
1.1 1.2 1.3 Motivation............................................................................................................................................1 Thesis goals & hypotheses...................................................................................................................3 Thesis outline.......................................................................................................................................4
2.1 Building process and its control...........................................................................................................6 2.1.1 2.1.2 2.1.3 2.2 Overview..............................................................................................................................6 Daylight and electrical lighting process ..............................................................................7 New perspective on building control process ......................................................................9
Models for building process control ..................................................................................................10 2.2.1 2.2.2 Simulation model...............................................................................................................10 Machine learning model ....................................................................................................16
HISSTO: a hybrid system for environmental control ..................................................21
3.1 Problem identification........................................................................................................................21 3.1.1 3.2 Issues in building control models ......................................................................................21
Requirement elicitation......................................................................................................................24 3.2.1 3.2.2 Functional requirements ....................................................................................................24 Non-Functional requirements ............................................................................................30
System analysis..................................................................................................................................32 3.3.1 3.3.2 3.3.3 3.3.4 Analysis of hybridization...................................................................................................32 System analysis object model ............................................................................................34 System sequence model .....................................................................................................35 User interface mock-up......................................................................................................37
System design ....................................................................................................................................39 3.4.1 3.4.2 3.4.3 System architecture............................................................................................................39 System component design .................................................................................................40 Concurrency identification ................................................................................................52
3.4.4 3.4.5 3.4.6 3.4.7 3.4.8 3.4.9 3.4.10 3.4.11 3.5
Hardware/software mapping..............................................................................................52 Data management ..............................................................................................................54 Global resources management ...........................................................................................55 Software control implementation ......................................................................................56 Boundary conditions ..........................................................................................................60 Adaptation to the system changes......................................................................................62 Design priorities.................................................................................................................63 Design trade-off .................................................................................................................63
Object design .....................................................................................................................................64 3.5.1 3.5.2 3.5.3 3.5.4 3.5.5 3.5.6 Object design of the Predictor ...........................................................................................64 Object design of the Tester ................................................................................................83 Object design of the Commander ......................................................................................88 Object design of the Sensor ...............................................................................................89 Object design of the Actuator ............................................................................................90 Object design of the daylight responsive lighting control .................................................90
System test & evaluation .................................................................................................96
4.1 4.2 Test bay & control target systems......................................................................................................96 Data acquisition for system test .........................................................................................................99 4.2.1 4.2.2 4.3 Analysis of the measurement...........................................................................................100 Data preparation for testing the predictors ......................................................................102
Performance test of the predictors ...................................................................................................104 4.3.1 4.3.2 Prediction performance of the LUMINA.........................................................................104 Prediction performances of the hybrid predictors............................................................106
Test of predictive control scenarios .................................................................................................121 4.4.1 4.4.2 4.4.3 Test of simulation-based control......................................................................................121 Test of hybrid predictor-based control.............................................................................124 Test of the convergence time to reach a control action decision .....................................126
Overall evaluation of HISSTO ........................................................................................................128
5.1 5.2 Contributions ...................................................................................................................................130 Future Research ...............................................................................................................................131
Appendix A: HISSTO’s test instrumentation Appendix B: HISSTO’s control performance evaluation
.........................54 Figure 3................................................................................5:Sequence diagram illustrating interactions among the key objects in HISSTO: UML notation ...............38 Figure 3............................9:Top level logical model describing HISSTO’s subsystem breakdown: UML notation .........................21: Physical connectivity for HISSTO’s system instrumentation ...........12 Figure 2.........14:Sequence diagram showing HISSTO’s event handling process: UML notation.....23: HISSTO’s event notification process based on the Observer design pattern: UML notation.....13 Figure 3.............3: Feed-forward control process example ....................List of Figures Figure 2........................1: Daylight and electrical light process in a space .......15:Logical model of HISSTO’s PredictionManager component: UML notation ....................................................35 Figure 3................. 1997)...................................................37 Figure 3.........13:Logical model of HISSTO’s EventManager component: UML notation ...............................24: HISSTO’s alternative control strategies based on the Strategy design pattern: UML notation ..................51 Figure 3...............................4:: Use of BDI and GAT for simulation-assisted building control (Mahadavi..............47 Figure 3......36 Figure 3.....18:Logical model of HISSTO’s CommunicationManager component: UML notation...27: Construction of a View based on the Builder design pattern: UML notation .........49 Figure 3.................................................32 Figure 3......48 Figure 3............................................................................12:Logical model of HISSTO’s MonitoringManager component: UML notation ......16:Logical model of HISSTO’s ControlManager component: UML notation ...26: HISSTO’s handling of control command based on the Command design pattern: UML notation60 Figure 3..................45 Figure 3..22: Access privilege management in HISSTO: UML notation..............................9 Figure 2..........................................59 Figure 3...................................42 Figure 3............................................50 Figure 3...................................2: Structure of a virtual building and its relationship with other entities...............30 Figure 3...10:Logical model of HISSTO’s SystemManager component: UML notation...................58 Figure 3..25:State transition diagram describing HISSTO’s control mode changes: UML notation ....57 Figure 3......55 Figure 3...........2:Feed-forward (a) and feed-back (b) control process..........19:Logical model of HISSTO’s UserInterfaceManager component: UML notation....................................4: System analysis class diagram of HISSTO: UML notation.....62 iii ...........................54 Figure 3.......................41 Figure 3.7: HISSTO’s Control Window on the web.....11:Logical model of HISSTO’s EnvironmentDescriptor component: UML notation .............44 Figure 3......40 Figure 3..6: HISSTO’s Monitoring Window on the web...............................34 Figure 3........................17:Logical model of HISSTO’s DatabaseManager component: UML notation .......................43 Figure 3............8:Component diagram describing HISSTO’s system architecture: UML notation ......1:Use case diagram of HISSTO: UML notation ..............................20:Deployment diagram describing HISSTO’s hardware topology: UML notation.....................46 Figure 3...........12 Figure 2..................3:Typical hybridization schemes between a simulator and a machine learner .........................
....2)................9:Indoor illuminance profiles depending on the sensor positions and louver angles (3/13/98).114 iv ..........113 Figure 4..........................17: Average relative distance and standard deviation of hybrid predictor’s reference parameters and preference indicators prediction from the simulations (with D1 and D2 data sets in Table 4.....1).83 Figure 3.....110 Figure 4.....65 Figure 3.36: Exemplary preference functions for HISSTO’s performance indicators .37:Activity diagram of HISSTO’s daylight responsive lighting control process: UML notation...12:Histogram showing LUMINA’s relative indoor illuminance prediction error distribution ..........6:Sky illuminance and irradiance measurements on a typical clear day (3/13/98).73 Figure 3....1).......................92 Figure 4...........................................15:Different hybrid predictors’ indoor illuminance prediction accuracy expressed in RMS of relative errors (Test case 2 with D4 and D5 data sets in Table 4..............................7:Sky illuminance measurement showing daily and seasonal variations . luminaires.....................87 Figure 3... and indoor sensors......33:Collaboration diagram describing SM12CNN’s prediction process: UML notation.......102 Figure 4.............97 Figure 4.....100 Figure 4..................71 Figure 3...75 Figure 3...105 Figure 4................11:Relative error of LUMINA’s indoor illuminance prediction .............................28: Viewer information necessary for LUMINA’s glare calculation ............................107 Figure 4........30: Generic architecture of the neural networks design in HISSTO ........................................98 Figure 4...............10:LUMINA’s indoor illuminance prediction performance in the test bay .....................16: Different hybrid predictors’ performance in predicting both reference parameters and performance indicators represented as the averaged relative distances from the simulator’s prediction............99 Figure 4.........................34:Exemplary collaborative process between WFNN(2) and BRIDGENN in WFC1NN...................................13:Learning curves of the neural networks in HISSTO.......................................8:Indoor illuminance profile on all reference points influenced by the louver angle changes ............79 Figure 3.................2: HISSTO system test bay at the Intelligent Workplace...106 Figure 4.........................18: MS1NN’s indoor illuminance prediction accuracy with the selected training sensor position...14:Different hybrid predictors’ indoor illuminance prediction accuracy expressed in RMS of relative errors (Test case 1with D1 and D2 data sets in Table 4........31: Different neural network training schemes based on the progression of time ............81 Figure 3.............32:Collaboration diagram describing SM12CNN’s training process: UML notation..List of Figures Figure 3............................................101 Figure 4.........................4: Plan and elevation view of the test bay with louver...............97 Figure 4.96 Figure 4............................109 Figure 4..........................................106 Figure 4........100 Figure 4............................29:Iconic representation scheme of the neural networks designed for HISSTO.111 Figure 4..35: Training and operation processes of the different hybrid predictors in HISSTO........3: Illustrative photos showing exterior and interior of the test bay ......5: Daylight station on the top of the IW measuring sky condition parameters......................................................................1: Intelligent Workplace at Carnegie Mellon University as the system test bed ..................80 Figure 3.....................
.............122 Figure 4.....22: Change adaptation performance of each hybrid predictor in indoor illuminance prediction.21: Decay curves of the hybrid predictors’ predictions for the reference parameters and performance indicators expressed in terms of their relative distances from the simulation output... 116 Figure 4..120 Figure 4......................127 v .....................115 Figure 4...........25: Mapping between ranked louver positions and control quality index .......119 Figure 4...................................26: Time requirements comparison needed for a single control time window ........List of Figures Figure 4............................19:WFC1NN’s indoor illuminance prediction performance (a) and RMS of the predicted performance indicators’ relative deviations from the simulations (b) with different calibration sensor points............................24:Overall evaluation of the hybrid predictors in a graphical form ..........20: Indoor illuminance prediction decay curves of the hybrid predictors in HISSTO ............. 118 Figure 4........117 Figure 4.....23: Change adaptation profile calculated as the closeness of each hybrid predictor’s reference parameters and performance indicators prediction to the simulation output .
......12:Order of the best louver positions suggested by the MSC0NN predictor .........................4: Design priorities of a typical building control system............103 Table 4....................103 Table 4..................89 Table 3....................................17: Comparison of the required time for HISSTO’s different controllers..123 Table 4...................................................104 Table 4...................................25 Table 3..............................................................8:Design characteristics of the neural networks to construct HISSTO’s predictors.........18:Assessment of the research outcome including HISSTO implementation..........2:Visionary scenarios of HISSTO ......9:Diagrammatic description of each hybrid predictor’s prediction process .............................10: Various sensors used in HISSTO............................................................126 Table 4........5:Overall evaluation of the hybrid predictors designed in HISSTO...........22 Table 3....................123 Table 4..................................9:10 best control actions suggested by the pure simulator ........................5: Input and output parameters for LUMINA.............................1) .............124 Table 4......7 Table 3....................................................11:10 best control actions suggested by the simulator + luminaire matrix predictor ...........................125 Table 4..............................82 Table 3........125 Table 4...........13:10 best control actions suggested by the MSC0NN predictor......2: Test data sets within ± 20 % deviation range from the Perez model prediction ....................................................................................120 Table 4..................................2..............................4:Example of typical simulation validation methods ..........................................70 Table 3.......................1:Test data sets filtered by the validation criteria (see A.........123 Table 4............................129 vi .......................66 Table 3........................................29 Table 3.............................11: Louver operation schedule example .....1: Examples of the challenges in energy simulation validation process .......6: Input and output parameters for the neural networks designed in HISSTO...................................................................1............6:Overall evaluation of the daylight control test ....................122 Table 4..1: Factors influencing thermal and lighting process in a building..............................124 Table 4........3:HISSTO’s use cases.................128 Table 4................126 Table 4.16:Final systems control states (with utility values) suggested by the tested predictors...................List of Tables Table 2.............................................................63 Table 3........................................................................................8:Initial state for the inputs to the predictor......................7:Percentage of instances with a specific control quality index ....................103 Table 4...15:10 best control actions suggested by the WFC1NN predictor................................................................7:Selected input (I) and output (O) parameters for each neural network designed for HISSTO ...14:Order of the best louver positions suggested by the WFC1NN predictor ............69 Table 3.................................................................................................94 Table 4.....................72 Table 3......................................3: Neural network training data sets generated by the weather file based simulations .........................10:Order of the best louver positions suggested by the simulator................................................125 Table 4........
letting me perform necessary experiments at the Intelligent Workplace. Professor Bernd Bruegge has been passionate and resourceful in bringing object-orientation and component technology into my thesis system development effort. I treasure the time spent with him and his ambitious students at the software engineering laboratory. It is she who has been always willing to be victimized by my run-time humors even when I broke her “no-more-than-ten-minutes” rule on purpose. and Jay Shankavaram for their friendly help and knowledge exchanges with unexpected occasional laughters. He has been an ever-present source of guidance and encouragement throughout my doctoral program residency. vii . I owe many thanks to my family members.Acknowledgment It has been a long journey through which I learned that life has many hidden layers to be discovered. Robert Ries. lending me their shoulders to lean on. I must add Rohini Brame. I want to thank Vineeta Pal for sparing her time to discuss the issues involved in her daylight simulation program with me. I also thank Jamieson Shulte at the School of Computer Science for his help in the instrumentation of my thesis experiments as well as the neural network implementation. To those named persons. It would certainly have been much difficult if I had not have my colleagues around. providing invaluable moments of friendship. commenting and answering my machine learning related questions with unsurpassed clarity. especially to my mother whose unconditional support and trust made this thesis finally see the outside world. I feel sincere gratitude to Professor Ardeshir Mahdavi who brought a great range of expertise and interest to my work. Paul Mathew. I should like to thank Professor Volker Hartkopf for his visionary inspirations and support. an essential part of contribution came from Professor Sebastian Thrun through his continuous enthusiasm. Surely. I dedicate this thesis to her with love and respect. Among those deserve being named.
Pennsylvania to lend this thesis to other institutions or individuals for the purpose of scholarly research. Pennsylvania to reproduce this thesis by photocopying or by other means. I authorize Carnegie Mellon University. I authorize Carnegie Mellon University. Copyright c 1999 by Seongju Chang viii . Pittsburgh.Copyright Declaration I hereby declare that I am the sole author of this thesis. in total or in part. at the request of other institutions or individuals for the purpose of scholarly research. Pittsburgh.
e. which makes the technique inevitably sensor-dependent. implemented in neural networks. However. On the other hand. ix . enhancing the accuracy of predictions necessary for the identification of the best control option. In this process. The hybrid constructs designed to explore diverse synergistic effects have been tested in daylight responsive lighting control domain. trained either to replace a simulator or to predict the simulation error. A machine learner can address this problem. This thesis has computationally implemented and tested seamless control scenarios based on the different hybridization schemes in order to identify desirable processes for such hybridization. A machine learner. A wide distribution of the direct digital control approach has inspired researchers and engineers to pay more attention to the computational solutions in building control. some simulation programs are computationally too intensive to be effective for real-time control purpose. the resulting pilot hybrid control system is capable of regulating target lighting systems effectively by covering a wider portion of the control state space with increased agility and precision even in the environments subject to dynamic changes. simulation) and inductive learning methods (i. it often requires large amount of data for training or retraining. is trained for faster predictions. HISSTO demonstrates that the various modes of user-system interactions can be facilitated by using component-based object-oriented software engineering techniques along with internet-activated control and monitoring interfaces. and maximizing the searches in the lighting system control state space within a limited time. serves as the pivotal component for a better predictor through a hybridization process. The enhanced knowledge about building systems behaviors as well as the increased computing power make attractive the potential use of a computational building model as the agent to carry on building control task. the role of a simulator is augmented as the source of the system knowledge by which a supervised learner. Multiple hybrid controllers are designed to meet four control goals: enriching the informational repertoire of systems control operations for lighting (by inclusion of performance indicators for glare and solar gain). neural networks) can cooperate to facilitate building systems control.Abstract The increased complexity in building systems integration provides a new challenge for building operation process.e. This thesis intends to demonstrate that analytical approaches (i. reducing the number of sensing units necessary for capturing the states of building’s visual performance indicators in real time. HISSTO’s software design pattern driven modular system architecture also allows it to be easily extended even to the other building control domains. HISSTO (Hybrid Intelligence for System State Transition Operation).
) that are essentially guided by temperature sensing in a space. as their past impact on the building’s dynamic behavior is reflected in the collected information by the sensing system. In this scenario. These dynamics make control action take place on a time scale: on the order of seconds for actions involving local control loops. addressed 1 . to a time scale on the order of minutes for changes in zone conditions (Kelly et al. There have been many hybridization efforts in control domain. thermostatic control of HVAC components involves typical operations (on/off. Developing a hybrid computational model for building control is the major focus of this thesis. For example.1 Introduction 1. yet some of them still remain to be constantly monitored. building control systems have operated based on a homeostatic short-term feed back mechanism. the effect of external environmental changes on building interior conditions takes place over a time scale on the order of hours. change in volume and/or temperature of heating/cooling media. Most of the building processes can be quite accurately predicted using a set of well-established models.1 Motivation Traditionally. This view has been significantly challenged by the growing complexity of various building systems and components as well as the dynamic interactions among them. because of the thermal inertia associated with a building. building control systems have become increasingly sophisticated. control options can be improved ("optimized"). but few of them. A building used to be considered as a static and a passive physical construct. One of the approaches has been to utilize various methods and tools (including neural networks) to accurately capture a building’s dynamic characteristics so as to provide a more reliable basis for the control of its behavior (Curtis 1996. etc. Capturing building’s dynamic behavior toward enhanced control strategies can be supported using advanced computational performance simulation routines. Such dynamism makes a building systems control task a complex and sophisticated process. 1984). This research starts from the idea that a well organized hybrid construct out of both simulation and machine learning technique can support the control of a dynamic building process such as the visual environment in a building. For example. if any. Generate-and-test as well as bi-directional inference methods can be used to derive preferable control schemes and required values for control variables based on parametric and iterative simulations (Mahdavi 1997b). More recently. Mistry and Nair 1993).
this algorithm can yield unstable control outputs unless it is tuned correctly. PID control algorithm using a feed-back loop with three mathematical terms: proportional. highly desired as long as the control system can deal with resulting complexity in both generation and assessment of control options. retrofit) using simulation data • Reduction of the number of sensing units necessary for capturing building's real-time operational status 2 . the role of building simulations can be augmented as the source of the system knowledge. Hybridization of different models or components is not an entirely new approach. It has been argued that prior knowledge can help significantly in learning (Mitchell 1997) and simulation can provide such knowledge to a machine learner. Beyond enhancing the effectiveness of dynamic control systems. the evaluation criteria of the visual environment in a building has been limited to very few quantitative indicators such as illuminance and uniformity. since there have been such attempts as neuro-fuzzy systems and neural network assisted PID controllers (Curtis 1996). the additional benefits of the suggested approach include: • • Prediction of the effects of the changes to building hardware and its control systems Beta-testing of building control system hardware based on the simulated actions in the context of a virtual building • Pre-training of machine learning systems simulation data on building behaviors prior to their field utilization • Retraining of machine learning systems to account for the effects of abrupt modifications to building configuration (i. Enriching the performance evaluation criteria is. For example. cumulative. a well-defined evaluation scheme for different control system designs and implementation options is also important. The mapping between possible states of a system and their outcomes useful for the training or retraining of a machine learner can also be provided by simulations. In this scheme. The major motivation for the hybridization between simulation and machine learning in this thesis lies in the importance of prior knowledge in learning process. A neural network could serve for this tuning purpose.the issue of combining model with machine learning techniques. Both data dependency and sensor dependency in neural network implementation can be dramatically reduced in this process. and time-variant processes. especially for the visual environment control in a building. therefore. Normally. Simulation can assist the identification of control-sensitive variables. Identifying potential synergistic effects between simulation and machine learning for an enhanced building systems control can be attempted through a hybridization process. integral.e. At the same time. But under non-linear conditions. and derivative works with well-behaved linear.
A machine learner can also copy the simulator’s knowledge so that it can reduce the time and computational load significantly in predicting the outcome of control actions after the training is completed.On the other hand. By combining a simulation model and a machine learning technique such as neural networks. The primary motivation for the thesis system development effort is to be able to dynamically change the values of control parameters on-line while the system allows implementationindependence and flexible future extension. some of the critical issues in model-based control can be resolved. the simulation-based virtual model allows for additional operations: i) the virtual model can move backward in time so as to analyze the building’s past behavior and/or to calibrate the model toward improved predictive potency. 1999). While the real building can only react to the actual contextual conditions (i. occupancy) and building control actions. To summarize. Interactive and modular building control software development is the other driving force of this thesis. The expected benefits that simulation model can have when it is combined with machine learners are as follows: • • Calibration of simulation outcomes by compensating for the simulator’s prediction error Reducing computational load and time in predictions by copying and replacing a simulator In order to realize this idea. This benefit could lead to a major improvement in building control performance as long as the proper design and implementation of the hybridization are achieved. a conventional building automation system must be supplemented with a multi-aspect virtual model of the building that runs parallel to the building's actual operation. The engineering goal is to identify a prototypical building control software architecture allowing flexible adaptation to dynamic environmental and user preference changes. This tuning process can improve the performance of simulation predictions. ii) the virtual model can move forward in time so as to predict the building’s responses to the various alternative control scenarios (Mahdavi et al.e. sky luminance distribution pattern.2 Thesis goals & hypotheses The scientific goal of this thesis is to verify the benefits of the hybridization between simulation and machine learning technique in building control domain. a machine learner can be utilized for on-line simulation calibration. 1. the major goals of this thesis are to identify: TG1) Hybridization architecture between a simulator and a machine learner for high-performance building control operations with increased speed and accuracy. 3 .
and interoperability while being successfully integrated with the developed hybrid computational models The fulfillment of the scientific thesis goal (TG1) will be tested against the following hypotheses: TH1) Hybrid prediction model can increase the speed of convergence to a control decision. change adaptability. Different controllers based on the different hybrid prediction models are then evaluated in the visual environment control operation of an actual building.3 Thesis outline This thesis involves a system development effort and the experiments for the developed system validation. demonstrating the performance enhancements of the hybrid prediction models in building control domain is critical. As is stated. and tested. Both the predictors and the controllers using those predictors are tested according to the defined evaluation criteria. The prototypical control software architecture designed for HISSTO. TH5) System flexibility can be maximized by using design patterns and component-based objectoriented programing methodology. multiple hybridization schemes in the various forms of computational models are designed. 1. namely use cases. TH2) Hybrid prediction model can increase the accuracy of predictions for more precise control. This thesis is progressed by the following steps: 4 . The performance of each controller is evaluated against measurement.TG2) Well organized building control software architecture for enhanced accessibility. is engineered to provide all necessary functionalities. TH3) Hybrid prediction model can decrease the sensor dependency of control operations while enriching the performance evaluation criteria involved in control action decision process. These prediction models are combined with a testing algorithm which performs control options evaluation task to produce the most appropriate control decision. system requirement elicitation section. within the target control domain. For this reason. implemented. The engineering thesis goal (TG2) will be tested against the system’s use cases and non-functional requirements described in Chapter 3. the thesis control system. TH4) Hybrid prediction model can provide better chance for the control system to adapt to building configuration or environmental changes. The success of this prototype control software is evaluated based on those use cases described in the system requirement elicitation section.
and evaluation of the prototype system: Chapter 4 Step 6) Validation of the thesis hypotheses and the system use cases: Chapter 4 Step 7) Summary of the achievement and future work: Chapter 5 5 . execution.Step 1) Outline of contextual knowledge: Chapter 2 Step 2) Identification of the requirements for the prototype system development: Chapter 3 Step 3) Identification and resolution of the prototype system design issues: Chapter 3 Step 4) Detail design of the hybrid prediction models and other system components: Chapter 3 Step 5) Instrumentation. planning.
Both of them increase internal cooling load in summer and decrease heating load in winter. A HVAC system consumes certain amount of energy based on the system type. From the control point of view.1. and shades interact with outside weather condition and influence indoor thermal condition. The building factors and the system factors are to be controlled to satisfy occupants’ requirements under the different states of the natural factors. in turn. Occupants not only impose control goals. but also influence various building processes as is shown in Table 2.1 2. Both the natural and the occupancy factors are dynamic but normally not controllable. Build6 . There have been significant amount of modeling efforts to describe each physical process in a building as well as the inter-dependencies among various such processes. an adequate indoor illuminance level and an acceptable PMV (Predicted Mean Vote)). Electrical lighting increases energy consumption whereas daylight is basically free of charge. Occupants and equipment also have impact on thermal environment by introducing heat and moisture inside. a building is a system which has multi-variant dynamic subsystems showing various linear or mostly non-linear behaviors. Building envelope components such as walls. the building and the system factors need to be interactively adjusted to these factors.e.1. operation mode and state. Daylight and electrical lighting can be complimentary in terms of satisfying the required indoor illuminance level. therefore. The building factors are normally determined during design and construction phase of a building life cycle while the states of the system factors are practically decided in the building’s operation phase. The control goal of the system factors is to minimize cost while maximizing occupants’ comfort (i. A HVAC (Heating Ventilation & Air Conditioning) system controls temperature. Both environmental and occupancy changes in a building increase the complexity of control operations. humidity. influences PMV(Predicted Mean Vote) as the index of thermal comfort. windows.2 Background 2. The factors being impacted by such phenomena as lighting and thermal process in a building include both human comfort and energy implication. and air flow which.1 Building process and its control Overview A building has diversified physical processes interacting with each other.
internal shade. and building orientation are.1) . ground surface reflectance number of occupants. interior surface reflectance. occupancy schedule site latitude & longitude. 1992). determining even a single control target system state requires a proper understanding of the interactions among all related factors (Table 2. Incoming sunlight is also diffused by the clouds before reaching a building envelope.2 Daylight and electrical lighting process Visual environment in a building is determined by numerous factors. Modeling of this modified sunlight by the atmosphere and clouds is normally based on the turbidity applied to the entire sky. shades. clothing Building factors building geometry. overhang. solar altitude. activity type. glazing reflectance.1. glazing transmittance. Solar azimuth. overhang.1 shows how light interacts with a built environment. wind. building geometry. glazing area. orientation. orientation. humidity Occupancy factors number of occupants. surface area Lighting process solar irradiance. Figure 2. outdoor temperature. For electrical lighting calculation. glazing area. blinds. only direct and indirect internal components are to be considered. occupant schedule. Such performance criteria as illuminance level and energy consumption are the typical example of those few factors. Angular relationship between the sun and the target space is another factor creating daily and seasonal variations of the indoor daylight profile. sky condition. room volume.ing control normally focuses on the manipulations of the target systems to make up for the gap between the desired state and the current one. The sun is a single source of daylight. interior finish. critical factors influencing this 7 . therefore. There are three components to be considered in daylight calculation: • Sky component: The rays from the sun are partially absorbed and modified by the atmosphere. luminaires 2. and early 1970s (Flynn et al. wall material. 1960s. glazing material. occupant’s location. wall construction. Because of the complex inter-dependency of all factors involved in a building process. interior surface transmittance System factors external shade. Few performance factors have been used for selection of lighting equipment and its location throughout 1950s.1: Factors influencing thermal and lighting process in a building Classification Natural factors Thermal process solar insolation. activity type. interior surface area. Table 2. interior finish. ceiling & floor materials. HVAC system characteristics light redirection louver.
Solar radiation brings not just light but heat into the space. In daylight responsive lighting control. Glare is caused by unbalanced perception of light on the human retina when there is a significant luminance difference in his/her visual field. absorption. ground.phenomenon. Uniformity factor is another index to evaluate how evenly illuminance values are distributed throughout a space. There have been many new daylight modification systems introduced to the construction of a building. Glare is another phenomenon which has impact on visual comfort. • Externally reflected component: The sky component sometimes can directly hit the indoor reference point but part of it is either diffused or reflected by the external objects such as neighboring buildings. flexibility in controlling our visual environment can increase significantly even after a building is designed and constructed. Surface properties such as transmittance. the radiosity method is used for diffuse surfaces whereas the ray tracing is applied to specular surfaces. the internally reflected light component is also important part of indoor illuminance calculation. Dimming is basically a control scheme to compensate daylight with electrical lighting so that both human comfort and energy saving can be achieved. 8 . a very delicate issue (Mahdavi et al. therefore. Controlling glare is a non-trivial task because glare is also dependent on such factors as the occupant’s position and the mode of activities in a space which are normally dynamic in their nature (Li 1997). determining how much daylight should be introduced to the inside is. or overhang before being introduced to the indoor space. Average illuminance can be derived to represent the illuminance of a certain area having multiple illuminance reference points. • Internally reflected component: Along with the externally reflected component. Device such as light redirection louver is adding more sophistication to our visual environment control capability by turning a traditionally static building facade into a dynamic one. Glare can come from either the daylight source or an electrical light source. By having this type of device. This phenomenon has energy implication especially during summer and winter because it increases cooling load in summer and decreases heating load in winter (Reynolds 1992). 1995). For more precise illuminance calculation. reflectance as well as specularity are important to calculate the amount of light bouncing back from a surface. Especially the glare on a computer monitor is becoming an important factor to be considered as the typical office work evolves toward a computer-supported information processing task.
Four major issues are identified for this research considering the new trends in building control domain: • • Extending the number of control criteria beyond the limited traditional ones Enhancing prediction capability of a building control model 9 . has been used for thermal load prediction and HVAC system control and is sometimes combined with PID controller to enhance its adaptation capability. Integral. Conventional PID (Proportional.Luminaire External reflection Direct component Internal reflection Figure 2. Neural network.1: Daylight and electrical light process in a space 2. Furthermore. as one of the machine learning techniques. and air flow environment (Jeannette et al. DDC (Direct Digital Control) has greatly influenced the overall building control practice by innovating both control hardware and software (Brandt and Shavit 1984).1.3 New perspective on building control process New building systems. visual. The goal of building operation is to maximize occupants’ satisfaction while minimizing operational cost. 1998). but new approaches using advanced process models or AI (Artificial Intelligence) techniques are now available to control various building processes including thermal. and derivative) controller is still being widely used. In recent times. and assemblies have been introduced along with increasing attention to the innovative concepts for energy conservation and environmental quality. 1998). a more intelligent and reliable building control approach is indispensable. In this context. internet-based building monitoring and control systems have been introduced which enable location-independent building operation with proper set of software considering security and various levels of access privilege (So et al. efficiently. the integration of building systems for an enhanced performance is becoming more and more important to increase user satisfaction. components.
The possible outcome of this process could be an option minimizing energy consumption and guaranteeing stable environmental control with reliable equipment at the same time.• • Agile and precise adaptation to the environmental and system changes Providing a well-established interaction scheme between users and building control system 2.2. Those computational models are being used to supplement or sometimes even replace costly equipment testing to develop more energy-saving building systems and components. In building control domain. • By establishing a model having a core of formulations describing already explained mechanisms in the process followed by numeric manipulations to decide the unknown coefficients.1. Simulations in real-time helps to build the control logic for complicated building systems to evaluate the apparently explosive number of potential options. as the virtual building. By applying numeric method to fit unknown coefficients to the statistically defined dynamic relations. Some of the ways to build a process model are discussed as follows: • By acquiring a cross-correlation function out of input and output data and use it to derive one of multiple candidate model forms. can provide an useful tool for generating and evaluating multiple control alternatives based on the given set of preference criteria without being restricted by the constraints that actual 10 . A model can be gained in a mathematical form describing interactions among its components.1 Simulator to support building process control Simulation is the transparent model of a system constructed through examining pertinent objects and processes in the target system. allows one to make system parameter changes and to study their effect without relying on the actual system which is sometimes inaccessible or costly. Simulation. on-line simulation models can be used for predicting actual system’s dynamic behavior by having input data from the sensors or other sources. The increasing need for more energy-sensitive and adaptive systems for building control has been encouraging the use of more precise and delicate computational models of various building systems.2.2 2.1 Models for building process control Simulation model 2. as a virtual system. • • By using regression method to determine the magnitudes coupled with hypothesized relations. Simulation.
the outputs of the simulation can be analyzed to see the outcome of that assumed action taken as assumption. to satisfy target indoor illuminance level. A series of experiments could be done to achieve this goal by systematically increasing or decreasing a subset of the independent variables. For example. When a target output is defined.e. This “goal-seeking” process needs a reasonably precise system model which is an essential part of feed-forward (open-loop) control architecture. • Goal seeking: In goal seeking. one can either change the luminaire dimming level or simply manipulate the blind. exceptional environmental conditions (i. a system model identifies necessary set of inputs to the system by predicting the system outputs with different input sets and comparing them with the target output. both sensitivity analysis and what-if analysis can help to guide which input set better be modified to produce the target output. lighting). to observe the impact of such operations. Special techniques are necessary to get to the desired input because there can be multiple input sets generating the same target output. and the level of equipment use can be provided to the simulator to increase its predictive potential.2. 11 . This process is also useful for system fault detection or for identifying an adaptation strategy when a building component system or the building configuration is changed. Contextual data such as the fluctuation of indoor occupancy. • What-if analysis: What-if analysis is to identify the consequence of a certain assumed control action. while others remain constant. Typical feed-forward control (model-based) process as opposed to the feed-back control (sensor-based) process is illustrated in Figure 2. The way of using a system model in an actual control practice is another issue to be explored. the desired target output of a system should be defined first to identify set of inputs which possibly yield that target output.building and its systems have. The following techniques are often used to achieve different goals by using simulation as a model for control: • Sensitivity analysis: Identification of control-sensitive variables is important. By setting simulation input values the virtual action. Usually.
a) Feed-forward control process r(t) system model u(t) system y(t) b) Feed-back control process r(t) + - c(t) system model u(t) system y(t) sensor r(t): model input. c(t): calibrated model input. Loop a) is the model calibration cycle by which the simulator is calibrated based on the detected model error.deviation b) goal seeking loop input system model model output - control error a) calibration loop system model error system output Figure 2.3 shows an example of how an analytical system model such as a simulator can be used for the control purpose. serve as the basis for comparative and/or parametric simulation runs. Loop b) is the goal seeking loop by which a sufficiently calibrated simulator is used to identify a proper set of system inputs for generating the given target output.3: Feed-forward control process example A critical task toward the realization of a simulation-assisted building control system lies in the development of a strategy to create a well-defined set of control options that. target output . u(t): system input.2: Feed-forward (a) and feed-back (b) control process Figure 2. While there may be numerous methods to derive at a structured set of such control options. two principal approaches are briefly described below (Mahdavi and Berberidou 1994): 12 . together with the projected contextual conditions. and y(t): system output Figure 2. The results of these simulations enable the control system to anticipate the impact of a control action choice on the selected performance indicators.
Such options may involve. attributes preference processing virtual building Bi-Directional Inference (BDI) control variable attribute sensing control Generate-And-Test (GAT) rule-based option generator control schemes matrix parametric simulation control scheme selection Figure 2. target perf. emissions. These schemes are then evaluated and ranked (possibly in view of multiple criteria involving power consumption. or various on/off timing schemes for intermittent heating/cooling. the average cumulative deviation of the maintained space temperature from the set-point temperature. cost. etc.4: : Use of BDI and GAT for simulation-assisted building control (Mahadavi 1997b) 13 . Figure 2. or the deviation of heating/cooling set-point temperature from the space target temperature. various positions of a movable external shading device. the bidirectional inference mechanism facilitates the derivation of required changes in the control variables based on desired changes in the performance indicators. or the average cumulative PPD (Predicted Percentage of Dissatisfied) in a space. positions of operable windows. illuminance distribution uniformity.4 illustrates the use of BDI and GAT for simulation-assisted building control. annual building energy demand.• The Generate-and-Test Method (GAT): This method involves the rule-based generation of a finite number of discrete control options. visual and thermal comfort. Starting from an initial operational state. This derivation can be accomplished via the investigative projection technique. An example of a control variable would be the position of a moveable external shading and/or light-redirection device. Examples of the performance indicators are the average illuminance on a desk surface. • The Bi-directional Inference Method (BDI): This method involves the explicit definition of control variables and performance indicators. for example.) based on the results of multiple simulation runs.
2.) occupancy model daylight process model electrical light process model Lighting process model can be coupled with thermal process model to assess the impact of heat gain from solar radiation and luminaires. In a radiosity based algorithm.2 LUMINA as a lighting simulator To be able to use a simulator as the prediction tool for both daylight and electrical lighting.2. A two-pass radiosity algorithm using extended form factors which include ray traced specular transfers is given by Tellier and Bouatouch 1994. a predominantly radiosity-based method which enters into a recursive ray tracing mode whenever a specular surface is encountered is preferred for the development of LUMINA.e. Ray tracing and radiosity based simulation techniques have overcome many of the limitations a simplified table-based hand calculation method has. shade. The computational storage is somewhat more of an issue in the case of radiosity (especially for diffuse surfaces) while the computation time is critical in the case of ray tracing (especially for specular surfaces). i. the most computationally expensive part of the calculation. making a trade-off between computational storage and computational time (Sillion 1994). The major concerns of the ray tracing or the radiosity based methods are computational storage and time. LUMINA uses both radiosity and ray tracing as its core algorithms.1. As a lighting simulation program. Since most architectural spaces include diffuse surfaces with occasional specular surfaces. needs to be performed just once and these results can be used repeatedly for multiple iterations. This is extended to include transmitting surfaces and is implemented in the following sub-processes (Pal 1999): 14 . LUMINA (Pal and Mahdavi 1999) is capable of modeling both daylight and electrical lighting process even in the spaces having a complex geometry. Most global illumination simulation techniques may use a combination of radiosity and ray tracing method. the calculation of form factors. luminaire etc. it must have: • • • • • • building model sky model system model (light redirection louver.
4. Read external surfaces from a file. y. and a phong exponent for transmittance. Visibility of window or light patch is taken account. 5. transmittance. 9. Calculate the visibility matrix for internal surface elements. A bigger grid is used for surrounding building surfaces. Similarly. 3. b) Compute the direct component of illuminance on the external surface elements. For each time step of the simulation period: a) Build the sky luminance distribution using SOLARIS and Perez sky algorithm. Surface information includes vertex coordinates (x. The surface information is the same as for internal surfaces. 2. Divide windows and lights into smaller elements by binary subdivision. calculate the extended form factors for internal surface elements. z) for each vertex followed by reflectance. specularity of reflectance. diffuseness of transmittance. d) Compute the direct component of daylight and also the initial direct lighting from the electric lights for the internal surface elements. Calculate the visibility matrix for external surface elements. phong exponent for reflectance. Read windows and lights. specularity of transmittance. Read internal surfaces from a file. A smaller grid is used for internal surfaces and louvers and overhangs. Repeat the process except that lights don't reflect or transmit but have an intensity distribution which is read from IES formatted luminaire files. 8.1. Calculate the extended form factor matrix for external surface elements. 6. Divide internal and external surfaces into patches by overlaying a grid and determining which grid elements lie inside the surface. diffuseness of reflectance. c) Compute the radiosity solution for the external surfaces using progressive radiosity. The sky is divided into patches and if a sky patch is obstructed by an external surface element. The extended form factors use ray-tracing to determine the specular component of the extended form factor. 10. the lumi- 15 . 7.
A machine learner can be tuned for optimal system operation even in an uncertain and dynamic environment. this model uses a clearness index and a sky brightness factor to compute the sky luminance distribution (Pal and Mahdavi 1999). Most of machine learning modules need training and cannot handle untrained conditions beyond their capability of generalization. the CIE standard overcast sky. e) Compute the radiosity solution for the internal surfaces using progressive radiosity.2. The environment is partitioned into exterior and interior spaces. These have limited applicability for predominantly intermediate skies. Most of the time needed is taken up by the visibility and extended form factor calculations. high-dimensional non-participant regression methods) available for dealing with data mining tasks with a weaker model and large data sets (Banks et al.nance of the sky patch is replaced by the luminance of the obstruction. Running LUMINA takes about a minute for a normal sized room (1. Earlier daylight simulation techniques relied on simplified sky models such as the uniform sky. Light redirection louver is considered as a dynamic external object having multiple surfaces modifying incoming daylight before it is introduced into the interior space. The radiosity solution for the exterior space is calculated first and then its contribution to the interior space is computed. a typical machine learning technique relies on inductive reasoning based on the given data sets.e. LUMINA uses Perez model for generating sky conditions necessary for the simulation. If the obstruction is specular. 2. 1999) Machine learning technique is capable of on-line learning from the real system operation. With the more advanced sky luminance distribution models available today. backward ray tracing is used to compute the specular component of light transfer from the external element to the internal element. On-going researches show that there are many alternative statistical methods (i. Depending on the source of knowledge and the way a machine learner captures it. most of these models have either been validated for the data from which they are empirically derived or have differential performance results when applied to different climatic conditions (Littlefair 1994).000 patches). since it is easier to solve two smaller systems of equations rather than one large one. This is more efficient than treating them as one space. This is based on the assumption that the interior space has an insignificant contribution to the exterior space. (1993) on the basis of sky luminance data collected for predominantly clear skies. a machine learning method can be classified as one of the following three categories: 16 . However.2 Machine learning model While simulation involves the explicit description of real systems and their processes. the applicability and accuracy of daylight prediction tools are expected to increase. All weather sky model has been developed by Perez et al. or the clear sky.
An unsupervised learner self-adjusts its parameters and structure to capture the regularities of input vectors without receiving explicit information from the external environment. This self-organizing learning scheme is useful especially when there is no prior knowledge either explicit or implicit except for the data set which only includes inputs. Machine learning module for a system control has strength especially in the following situations (Narendra 1990): • • When a controller is theoretically definable. Clustering attempts to classify objects based on feature descriptions. Clustering and discovery are two major application domains of unsupervised learning technique. The basic motivation of using machine learning technique is to capture the pattern of dynamic system behaviors (system identification) and to project its knowledge into the near future to predict possible states of a system or a controlled process (parameter estimation). machine learning techniques are useful in the control of unidentified complex and dynamic systems by learning directly from the multiple instances of the system input-output mappings called patterns. Supervised learning requires the correct output signal for each input vector to be specified. Neural network is one of the algorithms for implementing this type of learning. and Joule’s law. a program named BACON. • Unsupervised learning: Unsupervised learning clusters and characterizes input values without any target output or reward guiding its learning process. the learner only gets reward (either positive or negative) as a state is changed into another by an action. • Reinforcement learning: Reinforcement learning is suitable when there is no “condition-outcome” example to be trained with. Discovery learning is more likely to be used in theoretical environment. This learning scheme achieves its goal through identifying a series of actions that maximize rewards. Unlike rule-based systems. Ohm’s law. Reinforcement learning progresses by receiving an implicit scalar evaluation of an action and is capable of policy identification for choosing actions to achieve its goals (Mitchell 1997). Instead. as a model of data-driven scientific discovery. Supervised learning is capable of function approximation and system identification. the identification of the system over a long period of time followed by the use 17 . discovered a number of scientific laws such as ideal gas theory.• Supervised learning: A supervised learner is given both inputs and corresponding outputs so that it can predict right “condition-outcome” mapping after training. In this case. For example. but too complex for implementation When the controlled process has a poor transient response for initial conditions in a domain of interest.
is useful in building a controller capable of responsively dealing with a complicated and dynamic environment. Neural network is robust to noise in training data (Mitchell 1997). the number of training examples considered. Neural network is one of the most popular supervised learning techniques. It has been engineered as a computational system modeled after the learning abilities and parallelism of biological nervous systems. airline seat allocation systems. which requires both learning and prediction capabilities. the optimal input corresponding to each region is required One the advantages of having a machine learner as the major control agent is its adaptation capability to dynamic conditions. The backpropogation algorithm is the most common network learning method. neural network architectures and training techniques for building dynamic models are required. discrete-valued. Neural network consists of input and output layer and hidden layers in between.of nonlinear controller is needed • When stabilization of a controlled process is needed for the efficient operation of a system at several equilibrium states • When a control should be carried out on the basis of output patterns and the state space is partitioned into disjoint regions which are equivalent for purposes of control. They can successfully be integrated with the overall control system. and vector-valued target functions. Adaptive control techniques are normally designed to deal with uncertainty in the controlled process. When noticeable deviations are expected due to. explosives detectors. Backpropogation searches the space of possible hypotheses using gradient descent to iteratively reduce the error in the network fit to training examples. and the settings of various learning algorithm parameters. Adaptive control allows a control system to observe its own behavior and adjust accordingly. adaptive control systems is supposed to compensate through backing-up the model based on the available information. for example. therefore. Dynamic systems are generally history-dependent and cannot be modeled by static input-output maps. A process model for control can be refined by a certain additional program which observes the behaviors of the control process model in operation. Training time of neural network depends on the number of weights in the network. Adaptive system. Learning to interpret complex real-world sensor data is one of the best fits of neural network. Neural networks are suitable for non-linear modeling for dynamic system control. The applications of this technique include speech recognizers. Gradient descent converges to a local minima in the training error with respect to the 18 . therefore. Neural network learning methods provide a robust approach to approximating real-valued. an incomplete model. Therefore. Each layer contains nodes and links and those weights associated with the links are to be adjusted to fit the input-output relationships. and loan risk assessors.
For each network output. t〉 in training examples. t j is the target output for unit j .e. For each hidden unit number. calculate its error term δ h δh ← oh ( 1 – oh ) k ∈ outputs ∑ ω kh δ k (2. Do For each 〈 x. propagate the errors backward through the network: 3-2. Until the termination condition is met. n hidden hidden units. Initialize all network weights to small random numbers (i.0. The input from unit i into unit j is denoted x ji .3) 3-3. where x is the vector of network input values. Learning is the process through which neural networks are acquiring the knowledge embedded in the training data. Gradient descent is useful for searching continuously parameterized hypothesis spaces where the training error is a differentiable function of hypothesis parameters. and the output error is represented as E d .network weights. sys- 19 . 3.05). calculate its error term δ k δk ← ok ( 1 – ok ) ( tk – o k ) (2. and n out output units 2. Update each network weight ω ji ∆ω ji = – η ∂E d = η ( t j – o j )o j ( 1 – o j )x ji ∂ w ji ∆ω ji = ηδ j x ji for output unit weights (2. Input the instance x to the network and compute the output o u of every unit u in the network. o j is the output computed by unit j . between -0. The application of neural network typically comprises two phases. Create a feed-forward network with n in input units. A trained network subsequently represents a static knowledge base which can be recalled during its operation phase. and t is the vector of target network output values. η is the learning rate. A learning phase and an operation phase. Following is the backpropagation algorithm used for capturing knowledge embedded in training examples (cp. Do propagate the input forward through the network: 3-1.6) for hidden unit weights Each training example is a pair of the form 〈 x.5 and 0. the weight from unit i to unit j is denoted wji .4) 3-4.5) (2. The capabilities of neural networks in control domain can be summarized as follows (Barto 1990): • Copying existing controller: Neural network can be trained from already existing controller. Mitchell 1997): 1. t〉 .
tem model, or a human expert who has knowledge about effective control rules. • Adaptive prediction: The input to the neural network consists of delayed values of the signals, and the target output consists of the current values of the signals. The network tries to match the current signal values by adjusting a function of their past values. When the input to the network bypasses the delay units, the output of the network is a prediction of the values the signal will have in the future. • System identification: Training information can be obtained by observing the input-output behavior of a system where the network receives the same input as the system, and the system output is the target neural network output. If one selects a class of model to design an effective controller, a control law can be expressed in terms of the parameter estimates produced by the system identification procedure. • Identification of system inverse: In this method, the input to a network is the output of the system, and the target output of the neural network is the system input. If the neural network is a system inverse, by feeding desired system output to the neural network, it can predict what the system input should be. For error learning, the input to the neural network should be actual system output and a feedback controller needs to be used during training. • Differentiating a model: This method is for training a controller relying on backpropogation between layered neural networks such as forward model (system model) and controller. The propagation of the errors between actual and target system outputs back through the system model generates the error in control signal useful to train another neural network for controller.
HISSTO: a hybrid system for environmental control
Issues in building control models
With the advances in computation and DDC (Direct Digital Control), the model-based building control becomes an attractive option. A model may not always precisely capture the actual system behavior due, in part, to the difficulty of acquiring exact descriptions of building system properties, such as materials and construction features. Some simulation programs are computationally too intensive to be effective for real-time control purpose. An example is a lighting simulator that uses ray tracing in its modeling process. Such simulation programs cannot be used for control purposes unless the control state search space is dramatically reduced. Machine learning can address this problem. However, it often requires large amount of data for training. For instance, before a neural network is trained, or if it encounters unexpected conditions, it is not able to predict accurately (Curtiss 1996). The need for retraining makes it difficult for a machine learner to respond quickly to the system retrofit and/or seasonal weather pattern changes. Its sensor-dependency represents an additional difficulty, especially when placing and/or maintaining sensors are costly or otherwise not desirable. Conventional lighting control systems do not necessarily allow multiple control modes such as manual mode, scheduled mode, and predictive mode. Limited accessibility due to the location-dependency of building systems control also needs to be addressed.
22.214.171.124 Challenges in simulation Accuracy of prediction
Simulation is a generic analytical model for a specific process and can not always describe “actual world”. In many cases, validation of a simulator based on empirical data is expensive and time-consuming. It is normally pursued at a national or an international level, although many individual research groups undertake small monitoring schemes which focus on specific areas of building performance (Clark 1984). It usually takes much time to build and validate a simulation. Simulation can be used for
model-based control, training and testing of other control models (i.e. machine learning), and by generating the pairs of control action and predicted outputs so that the most promising control option can be identified. It also can be used to train or test other control models (i.e. machine learner). The feed-forward (model-based) control scheme requires high resolution system behavior description of the system model. Ultimately, it is only by comparing model predictions with the corresponding results from actual building in operation that a model’s usefulness as a predictive agent must be ascertained. Unfortunately, this task is difficult due to the lack of comprehensive data relating to the performance of real buildings and the shortcomings inherent in even the most sophisticated technique, which make it difficult to model reality exactly. Commonly encountered problems in validating an energy simulation program are shown in the Table 3.1 (Clark 1984).
Table 3.1: Exemplary challenges in energy simulation validation process Complex process Data validation Monitoring air movement and specifying the results in a form meaningful to modelers On-line validation, organization and management of large data sets in a manner which allows estimation of the data reliability Occupants behavior Ill-understood actions of occupants with regard to window opening, blind operation, and control manipulation Design/construction discrepancy Cost of development Climatic data Cost associated with the data logging system, sensor installation, and quality technical staff Availability of simultaneously recorded climatic data relating to the test site (This is especially problematic with many historical performance data sets.) Instrumentation Necessity to ensure, by constant checking, that all instruments remain as calibrated for the duration of the monitoring period Accurate determination of thermo-physical properties as required by the models
Increased computational load
Some simulation programs are too computation-intensive for real-time or quasi real-time control purpose. A typical example is any simulator using FEM (Finite Element Method) or ray tracing as the primary modeling algorithm. Normally, model predictions and the control decision based on them should be done within a certain time frame. For this reason, those simulation programs demanding heavy computation are not suitable for the control purpose unless the control state search space is dramatically reduced to limit the number of simulation sessions necessary for testing different control options. This problem has been significantly eased by the dramatic increase of available computing power, yet a careful consideration needs to be made when developing of a building product or process model for the control operation demanding heavy computation.
For example. For example. 23 . This sometimes causes problem. a machine learning module needs retraining for adaptation. in case of building configuration or system changes. neural network training process is normally data-intensive and time consuming even though a trained neural network quickly produces outputs for the given inputs. In fact. This sensor dependency can easily limit the boundary of the adoptable performance criteria by which a desired control option is selected. This problem leads to the following requirements: 1) Provision of seamless transition from the training mode to the operation mode 2) Creation of an algorithm whereby the training occurs with minimum memory requirement 3) Inclusion of a substitute algorithm to generate values for the controller output while the neural network is trained Sensor dependency Typical neural networks require the inputs/target outputs to be measured.3. especially when placing or maintaining sensors is difficult or costly. glare in an office environment can not easily be measured because it again is the function of the viewer’s location and view angle which are dynamic in their nature. which makes it inevitably sensor dependent. This makes the neural network slow in responding not just during the initial training phase but also when it needs a update through retraining.1. and this poses problems in using neural networks for control purpose. Also. There are certain cases no sensor could be placed because it simply doesn’t exist.1.2 Challenges in machine learning Data dependency Machine Learning technique often requires large amount of data for training. lots of control variables in building operation can only be calculated instead of being measured.
a new system development effort.2. HISSTO must minimize the lighting and thermal energy consumption and maximize occupants’ visual comfort in the target control zone. A scenario could be one of four types (Carroll 1995). elucidating visionary scenarios is the best choice to identify HISSTO’s potential requirements. 1999) for developing HISSTO. human comfort and cost minimization are the primary driving forces behind any building systems control efforts. spatial retrofit).1. informal description of a single feature of the system from the viewpoint of a single actor (Bruegge and Tutoit 1999). additional performance indicators need to be introduced. as an instance of a use case. Table 3.1 Visionary scenarios of HISSTO A scenario.e. 3. therefore. Since developing HISSTO is a greenfield engineering.g. There are also some non-functional requirements (Bruegge and Dutoit. HISSTO needs to capture seasonal changes of the sky condition while adapting itself to the changes of building configuration (e. focused. 10 to 15 minutes). HISSTO makes use of the simulator-assisted machine learner training as well as the machine learner-supported simulator calibration and tuning. Evaluation scenarios describe user tasks against which the system is to be evaluated. Visual comfort in a space is not sufficiently captured just by identifying indoor illuminance distribution. is a concrete. Predicting and evaluating the values of the performance indicators for all candidate control options should be done within a certain time frame (i.2 Requirement elicitation HISSTO is a pilot control system applied to an intelligent daylight responsive lighting control task. In general. Therefore. which starts form the scratch. the sensors required during HISSTO’s operation stage need to be confined to a small number (i.3. one or two).1 Functional requirements 3.2.2 shows an exemplary set of HISSTO’s visionary scenarios describing possible system usages of the different users. It is not feasible to use a large number of sensors in the target space beyond a certain period of data collection.e. Visionary scenarios focus on future system to refine ideas on the system needs to be developed. 24 . As-is scenarios reflect current situation regarding the current system understood by observing and describing user actions. For a successful daylight responsive lighting control. Training scenarios teach users how to use the system.
connects to HISSTO web interface and be authorized. Mark selects ControlZone 1 and changes control mode from predictive to scheduled. David checks out the new partition configuration. 4. David opens up the web-browser on his laptop computer. 1. David selects ControlZone 1 and changes simulation and machine learner training parameters to speed up the control response time. 2. notifying that the partitions in their workplace have been reconfigured. David gets a phone call from Mark. Julia presses exit button and closes the web-browser. connects to HISSTO web interface and be authorized. HISSTO updates simulator and machine learner for a faster response. 3. Kate presses exit button and closes the web-browser. David presses stop button and closes the web-browser. 3. Mark presses exit button and closes the web-browser. David presses monitoring button and select ControlZone 1 4. 5. David connects to HISSTO server interface and be authorized. HISSTO updates simulator and machine learner for change adaptation. the team A leader. Team A members decide to have a brief outdoor picnic. 1. While working at her office. 3. 3. All sensor values and the values of control performance indicators for ControlZone1 are displayed on the screen. 3. 1. 4. 2. Kate selects ControlZone 1 and change control mode from scheduled to predictive. 1. 4. connects to HISSTO server interface and be authorized. 3. 3.2: Visionary scenarios of HISSTO ID VS1 Scenario name managerOn Travel Actor instance David: FacilityManager Flow of events 1. the team A leader. Julia selects ControlZone 1 and change control preference profile based on her personal visual comfort priority. David gets a phone call from Mark. 2. about team members’ complaints on the delayed dimming response in ControlZone1. Mark. Kate opens up the web-browser on her desktop computer. 5. 2. opens up the web-browser on his desktop computer. 2. VS2 repeatedLightening Kate: Occupant VS3 teamAOnPicnic Mark: Occupant in team A VS4 workOnWeekend Kate: Occupant VS5 partitionChange David: FacilityManager VS6 workspace Re-assignment Julia: Occupant VS7 slowLampDimming David: FacilityManager 25 . Kate sets a comfortable ambient light level by manipulating luminaire slide bars on the screen. Julia moves to Kate’s workspace after the recent work group change. 5. connects to HISSTO web interface and be authorized. connects to HISSTO web interface and be authorized. Kate notices that on-going lightening distracts ambient light level too often due to the predictive control mode in operation. connects to HISSTO web interface and be authorized. Kate presses exit button and closes the web-browser. HISSTO sets ambient light level and louver angle based on the nonoccupancy schedule. David presses exit button and terminates the HISSTO server interface. David selects ControlZone 1 and changes building configuration file according to the new configuration. David presses exit button and terminates the HISSTO server interface. 1.Table 3. 2. Kate comes to her workspace on Sunday to finish the remaining work. David is out of Pittsburgh participating in the COMDEX show in Atlanta. 1. 4. the team leader. Kate selects ControlZone 1 and change control mode from predictive to manual. 4. 5. Kate opens up the web-browser on her desktop computer. 2. 5. 4. Julia opens up the web-browser on her desktop computer.
This is not just for providing users with the information on what is happening in their work space. always closed in summer. This process can be repeated based on a fixed time frame until the control mode is changed. and always open in winter. The following statement are the additional descriptions of the HISSTO’s major use cases . current sky condition is captured by outdoor sensors.illustrated in Table 3. and logged for future reference.3. The variables that can’t be measured also need to be calculated.initiated by only human actors . 26 . A use case represents a complete flow of events through the system depicting series of interactions caused by the initiation of the use case (Bruegge and Tutoit. This scheduled control mode requires the analysis of occupancy patterns. displayed. but also for accumulating history data to train machine learner or to calibrate simulation whenever it is necessary. Use case 1: Monitor Building Operation All environment factors. Global optimization should be activated in this mode so that the overall visual environment of a control zone is not dominated by only a subset of the occupants. Use case 3: Scheduled Control Scheduled control is another operation mode with which a predefined control schedule dictates the systems states at any point of time. environmental changes throughout a year. For example. luminaires can be scheduled to be all off between 7 PM and 7 AM and maintains the dimming mode for the rest of the day.1. Then the selected predictor predicts the possible control outcomes based on different candidate control actions.2. occupant factors.3. Those predicted control outcomes are then evaluated according to the defined performance criteria. Use case 2: Manual Control: In manual control mode. Use case 4: Predictive Control In predictive control mode. a selected control action is executed to change systems states. and system factors need to be monitored during building operation. Louvers can be scheduled separately in such a way that they are in open position early in the morning and late in the afternoon during spring and fall season. 1999).2 Use cases of HISSTO Use case generalizes all possible scenarios for a given piece of functionality. Eventually. accordingly. and possible impacts of the system state changes. the control is basically up to the user who is directly operating a louver or luminaires through a well-established interactive control interface.
the detected condition and the prescribed action for that need to be recorded. If the change is either observed or predicted. at least monthly on-line retraining with updated data set is necessary for the neural networks used for building systems control. This is a temporary condition which normally lasts only for some period of time before getting back to normal. When a space layout change. The control system should be able to identify the best control actions for those unexpected situations and maintain the selected control state until the exceptional environmental condition is back to normal. Pure simulator as a predictor doesn’t need any of this efforts because it can always deal with any time instance in a year to generate predictions over the outcome of the performance indicators. elimination.e. the control system should update its knowledge on its target environment to reflect those changes while the building operation is maintained until the updated predictor is ready to take the control back under its regulation. Use case 5: Update Predictor There are two types of change events upon which the predictor needs update by retraining. Even when a system fault is detected. a typical feed-forward (model-based) controller using updated simulator can be activated. The following process is a possible scenario which can handle this situation: 1) Once a system or building configuration change occurs. 2) While the updated simulator is temporarily regulating control operation. in this case. The simulation. It is assumed that the change is made explicitly but the impact of it is unknown. In the mean time. or an interior surface finish change occurs. the control system should also be able to adapt itself to the resulting new building configuration. Daily solar pattern change is relatively easy to capture because it can be done by including multiple days of either measured or simulated data in the neural network training patterns. The daily and seasonal solar insolation pattern change are mainly due to the relative position and movement of the earth to the sun. Such an adaptation is the explicit one in the sense that the change can be traced and be informed to the simulator anytime it occurs. lightening) should be carefully handled not to distract occupant’s visual field. For tracking seasonal solar insolation changes. this scenario might work as long as the faulty system is identified. or replacement. For the future reference. accordingly.Sudden unexpected weather changes (i. a furniture system rearrangement. the neural networks used in the current hybrid predictor is in the retraining process by having either the simulator-generated data or 27 . Upon encountering these situations. can predict the impact of such changes even without waiting for the actual control output showing deviations. the Facility Manager changes simulation inputs. This also needs to be done in cases of system addition. Environment change is slower and predictable. the neural networks in a hybrid predictor can always be retrained by using either the simulation generated or the measured data reflecting those changes.
‘ Use case 6: Set Control Preference A User should be able to modify his/her control preference by assigning different weight to each performance indicator. Actuator actor receives the control decision and actuates the corresponding system to change its state. simulation. or machine learner training while a specific predictor is engaged in the prediction of control outputs. As a part of data acquisition hardware. They can also trigger any of three control-specific use cases. The Unified Modeling Language (UML) is a graphical language for visualizing. constructing and documenting the artifacts of software systems. the hybrid predictor in command needs retraining after either the simulation or the machine learner parameters are changed. the control decisions after the change automatically reflects the user’s new control preference. Based on the control output prediction and current control preference functions. Sensor actor participates in data collection. 3) Once the retraining of the neural network is finished. Both an Occupant and a Facility Manager are capable of either monitoring systems control states or changing control preference functions for control customization. The weight of a performance indicator is relative to each other and has impact on the selection of the best control option at a given time. the evaluation of multiple control options need to be done to yield the final control decision. the control system brings the updated hybrid predictor back on for the normal operation. The UML was originally derived from the object modeling languages of three leading object-oriented methods: Booch. the modification of those predictor specific parameters can be done. Use case 7: Set Control Parameters Changing parameters for simulation and machine learner training has impact on the performance of each predictor. Collected data is used for monitoring. Once the user preference profile is changed. Object Modeling Technique (OMT) and Object-Oriented Software Engineering (OOSE). includes “Test Control Options” and “Set Actuator State” use case. For example. 28 . as the main focus of the system development. “Predictive Control” use case. Whenever HISSTO needs performance tune-up. Figure 3. accordingly. an Occupant can directly specify desirable control actions in “Manual Control” mode. specifying.the measurement data retrieved after the change. UML became the industry standard for modeling objects and components (OMG 1977).1 illustrates HISSTO’s functional requirements represented as a use case diagram in UML notation. To take this change in effect.
The Occupant selects target control zone and changes control preference functions and/or the weights of performance indicators based on his/her personal visual comfort priority. Exit: The User presses exit button and closes the web-browser. 2. Entry: The FacilityManager is informed of a certain building configuration change in a workspace. The Facility Manager selects target control zone and changes control mode to the scheduled mode. connects to HISSTO web interface and be authorized. 2. 2. Exit: The Occupant presses exit button and closes the web-browser. The FacilityManager presses exit button and terminates HISSTO server interface. 2. 2. 3. Exit: HISSTO updates the simulator and the machine learner for making change adaptation possible. 3. The Occupant selects a control zone and change control mode from the current one to the manual mode. Exit: The Facility Manager presses exit button and closes the web-browser. The User presses monitoring button and selects a control zone to monitor. 1. The User opens up the web-browser on her desktop computer. connects to HISSTO web interface and be authorized. The FacilityManager checks out the changed building configuration. The FacilityManager selects the target control zone and changes building configuration according to the identified new configuration. Exit: The Occupant presses exit button and closes the web-browser. The Facility Manager opens up the web-browser on his/her computer. 1. Exit: HISSTO updates the simulator and machine learner to finish tune-up process. 3. The FacilityManager presses exit button and terminates the HISSTO server interface. connects to HISSTO server interface and be authorized. connects to HISSTO web interface and be authorized. 2. The User selects target control zone and change current control mode to the predictive mode. The User opens up the web-browser on the computer. Entry: A User is in his/her workspace while the current mode of the target zone control is not the predictive control mode. The Occupant opens up the web-browser on his/her desktop computer. The FacilityManager selects the target control zone and changes simulator and machine learner specific parameters for adequate tune-up. connects to HISSTO web interface and be authorized. connects to HISSTO web interface and be authorized. The Occupant sets comfortable ambient light level by manipulating luminaire slide bars and louver angle knob on the screen. 1. 1. The FacilityManager connects to HISSTO server interface and be authorized. 1. Entry: An Occupant wants to re-specify his/her control preference. Entry: The Facility Manager is either in the work space or in a remote place while the control mode of the target zone is not the scheduled mode. Entry: An Occupant is in his/her work space while the control mode is either the scheduled or the predictive mode. 3.3: HISSTO’s use cases ID UC1 Use case name Monitor Building Operation Actor User Entry/Flow of events/Exit Entry: A User is in his/her work space or in a remote place. 1. The Occupant opens up the web-browser on his/her desktop computer. 3. Entry: The FacilityManager decides to tune-up the predictors for better predictive control performance in the target control zone. The current sensor readings and control performance indicators for the selected control zone are displayed on the screen. 2. Exit: The User presses exit button and closes the web-browser.Table 3. UC2 Manual Control Occupant UC3 Scheduled Control Facility Manager UC4 Predictive Control User UC5 Update Predictors Facility Manager UC6 Set Control Preference Occupant UC7 Set Control Parameters Facility Manager 29 . HISSTO sets ambient light level and the louver angle based on the schedule defined by the FacilityManager. 1.
<<include>> User Monitor Building Operation <<include>> Manual Control <<inc lude>> <<inc lude>> Get Sensor Value <<extend>> Sensor Get Event <<include>> Predic tive Control <<include>> Set Actuator State Actuator <<include>> Occupant Set Control Preference <<include>> Test Control Options <<include>> Scheduled Control Set Access Previliege Set Schedule <<include>> <<include>> Set Control Parameters <<include>> <<extend>> Set Building Configuration Update Predi ctor <<include>> Generate Performance Predictions <<include>> Fac il ity Manager Limit Device State Range Figure 3. and security issues (Bruegge and Tutoit 1999). performance issues. hardware considerations. error handling.1: Use case diagram of HISSTO: UML notation 3.2 Non-Functional requirements Non-functional requirements of HISSTO describe user-visible aspects of the system that are not related to the system’s functional behavior. These include user interface factors.2. HISSTO has the following non-functional requirements: 30 .
NFR4) The system needs to be remotely accessible while providing necessary security by a tightened authentication and authorization scheme. The response to this exception needs to be quick and pre-programmed since there is not enough time to go through the normal predictive control process. NFR5) The system needs to provide a way to accomplish cross-building performance domain evaluation of a candidate control action. Preferable time window for a control decision is 10 to 15 minutes. This control time window can be overridden when an exceptional environmental change is detected (i. NFR2) Evaluation of each control option requires to be agile enough to cover a sufficiently wide range of systems state search space within a limited time window.e. NFR3) Number of sensors in operation needs to be minimized to reduce the cost and maintenance effort.NFR1) Predictions should be precise enough to make an adequate system state change based on the given control performance indicators. 31 . sudden hit of direct sunlight or lightening). Allowable relative prediction error is within ± 25 % for any predictor.
The predictor receives building description and the real-time sensor values (for measuring environment and systems states) to generate control options and the predicted values of the performance indicators for each of those control options. They use both simulator and machine learner components. sky condition is a typical independent variable whereas indoor illuminance profile is a dependent variable which is determined by the interactions between the environment and the real building including the systems to be controlled.3. The focus here is on the virtual building which interacts with a real building and its environment. Virtual building has both independent and dependent variables. and tester. Tester evaluates the generated control options to identify the most desirable control action.2: Structure of a virtual building and its relationship with other entities 32 .3 3. sensor.1 System analysis Analysis of hybridization Figure 3.2 shows how HISSTO’s model components are combined and related to each other. Hybrid predictors are special type of predictors developed for HISSTO.3. but also virtually tests and self-regulates systems actions for the purpose of satisfying visual comfort of the occupants with minimized operational cost. virtual building building descriptor predictor simulator machine learner tester actuator sensor real building environment time system wall window louver dimmer blind clouds soalr radiation Figure 3. The virtual building consists of building descriptor. For example. The virtual building’s behavior is more than just a replica of the real building in the sense that it does not only copy the real building’s behavior in a passive way. predictor (a simulator and/or a machine learner).
2) Every time simulation performs prediction. 3) By observing prediction error. The type B predictor is basically targeting a calibrated simulator supported by a simulator-error prediction machine learner. the machine learner compensates the simulator output based on its own prediction for the possible simulation error. the main focus of this research. the difference between the calibrated simulator’s prediction and the actual system output. 2) The magnitude of error between the system output and the trained machine learner’s prediction can serve as the indicator whether or not the machine learner should be updated by retraining. or a system. 1) the simulator generates training patterns to provide prior knowledge for the machine learner. or a system. is designed primarily for the predictor. it can be identified when the machine learning-based simulation calibrator needs retraining to reflect the possible change in the environment. In this scheme.The actual hybridization process. this will likely increase prediction error of the machine learner which triggers simulationassisted retraining process. the simulator is the major player in control and the machine learner assists simulation calibration for its better prediction.3 demonstrates two distinctive structures of such hybridization process focused on the machine learner’s role in the hybrid construct. the building. so that a machine learner can have initial training. Figure 3. 33 . the building configuration. 1) A machine learner is trained for predicting the simulator’s prediction error by being trained with simulation error patterns. The type A predictor utilizes the “trainer-trainee” relationship between a simulator and a machine learner. 3) If any change happens in the environment.
3: Typical hybridization schemes between a simulator and a machine learner 3.Predictor type A.3. The Simulator or the MachineLearner accesses a View through the EnvironmentDescriptor to have necessary information about a Zone or ZoneComponents. the PredictorManager. and gets their current states. The SystemManager is the major component regulating most of the system’s monitoring/control behaviors. The MonitoringManager accesses Transducers: either Sensors or Actuators. The PredictorManager creates a Predictor in a combined form of Simulators and MachineLearners. An User is authorized by the SecurityManager and can access the SystemManager. The Tester in the ControManager evaluates possible control options identified by 34 . measurement training prediction update trigger prediction error - sensor simulation simulation error + machine learner training operation calibration Figure 3. measurement training prediction update trigger simulation sensor training pattern prediction error machine learner Predictor type B. It also gets Sensor or Actuator data through the MonitoringManager. and the EnvironmentDescriptor to come up with a desirable control decision depending on the current ControlStategy. The ControlManager uses the MonitorManager.4 is HISSTO system analysis object model showing the logical structure of the pilot control system to be developed.2 System analysis object model Figure 3.
and other multi-modal user interfaces across the entire system scope. windows.the Predictor based on a defined set of performance indicators.3 System sequence model Figure 3. a certain interaction is either prior or posterior to the other. The UserInterfaceManager is responsible for creating and maintaining browsers. creates Events.3. The EventManager screens exceptional states of the devices and environmental parameters. The interactions described in this diagram are order-dependent. and handles those Events. The control decision is then passed to the Commander which actuates the corresponding system for its state change. Us erInterfac eCom ponent W indow S y s tem M anager G laba lDataMo nitor Us erInterfac eM anager Us er S ec urity M anager Databas eM anager Connec tion Databas eTenan t rans a c ti on ControlM anager Control M onitoringM anager Com m unic ationM anager ControlS trategy Comm ande r S c heduledControl M anualControl P redic tiveControl Devic e Stat e P r edic ti on Tes ter en s or A c tuat or E ventM anager P redic tionM anager E nvir onm entDes c ri ptor E vent E ve ntHandler P redic tor Calibrator V iew V iewE lem ent A lertE vent ChangeE vent F ailureE vent S im ulator M ac hineLearner ZoneV iew Zone Co ntrolV iew Figure 3. The DatabaseManager handles persistent data requirements within the system boundary. The diagram depicts: a) a typical sensor value monitoring process. c) the predictor update process triggered by a building configuration change event.4: System analysis class diagram of HISSTO: UML notation 3. properly.5 shows the interactions among the key objects in HISSTO. The CommunicationManager mediates all necessary communications among any subset of the system components. 35 . d) and a typical daylight responsive lighting control process based on a hybrid predictor using a combined lighting simulator and machine learner. for example. b) an event detection and notification process.
displayChange() activateControl() startPredictiveControl() *getSensor() get Traine dNNWeig hts() createNN() getState() performPrediction() getResult() evaluateLouverOptions() rquestV ie w() cre at eV ie w() bui ldLight in gS imul at or() e tView() performLightSi mulation() returnResul t() evaluateControlOptions() setState() notifyChange() di splaySy s emStat e t logControlHistory() d.5: Sequence diagram illustrating interactions among the key objects in HISSTO: UML notation 36 .Occupant: User Bro wser: UserInterface DatabaseManager SkySensor: Sensor ZoneView: Vi ew Lu mi na : Simulator WF1CNN: MachineLearner LouverControl: Control EventManager LouverAc Actua selectMon itoring() *requestSensorVal() saveSensorVal() *getSensorVal() displaySensorVal() a. Figure 3. detectBldgChange() editView() updateView() saveUpdatedView() e tView() *performSimulation() saveResult() *getT ra in ingPat ten () performTraining() notifyChange() saveWe ights() c. requestEventReport() requestState() sendState() detectEvent() saveEvent() sendEvent() di aplayEvent() b.
indoor illuminance distribution. Figure 3.3. and data sampling rate along with the integration time interval of the sampled sensor values. As shown in Figure 3. An user can specify the control zone he/she wants to monitor.6. The other sections include the sub-windows for displaying outdoor daylight condition. monitoring schedule.3. and real-time values of the defined control performance indicators based on the deployed control actions as well as the given evaluation algorithm. The web-based user interface components on the client side have two types of windows: the Monitoring Window and the Control Window.6: HISSTO’s Monitoring Window on the web 37 .4 User interface mock-up HISSTO’s overall structure is based on the “thin-client-thick-server” architecture. the Monitoring Window has a real-time video display section to deliver visual information about what is happening in the target control zone.
7: HISSTO’s Control Window on the web 38 . momentum as well as the maximum prediction error could be specified along with the type of machine learner to be deployed.7 shows a design of HISSTO’s Control Window on the web. maximum training steps. Figure 3. password. control mode and schedule can be selected. Another section of the Control Window allows automation agent setup. For a simulator. learning rate. a user is required to input user ID. and the luminaire dimming change interval can be specified.Figure 3. To be authorized for certain control actions. System Control Station section includes user interface elements for manually controlling the louvers and the luminaires in the target control zone. and the target system he/she wants to control. the louver position change interval. building surface discretization level. For a machine learner. Depending on the user’s access privilege.
The HISSTORemoteCommunicationServer is designed primarily for the web-based system operations and has two sub-components: the DataSocketServer and the WebServer. There are other system components connected to the core of HISSTO making it fully realize a distributed building control system. loaded on a remote web browser can pull the values and display them on the screen.4 3. For example. there are nine subsystems such as SystemManager. EventManager. The Hardware/SoftwareInterface provides a bridge to the BuildingMonitoringControlHardware which contains Sensors and Actuators and communicates with HISSTO for data exchanges. In terms of implementation. The SystemManager regulates user access privileges while globally managing the system operations in the way it can avoid potential conflicts. 1996) provided useful resources to structure HISSTO’s system framework. et al. MonitoringManager. The DatabaseManager provides a single database interface for various data transactions involving all persistent data types for the system. The PredictionManager builds a proper predictor and performs predictions on the possible outcomes of candidate control options. HISSTOLocalInterface is mainly for the Facility Manager to perform necessary system monitoring and operation tasks. ControlManager performs control tasks based on the performance predictions and evaluations of the selected control options. Finally.8 describes HISSTO’s system software architecture. In HISSTO. The MonitoringManager polls sensors and actuators and creates exceptions when the captured state of a device shows an abnormality. PredictionManager. then notifies those events to the observers. and UserInterfaceManager. this communication is achieved through library function calls (HISSTO’s core package is complied as a DLL: Dynamic Link Library). The EnvironementDescriptor in HISSTO assembles a necessary building product model on-demand which typically includes a ControlZone with multiple ControlZoneComponents serving as the inputs to the simulator.1 System design System architecture HISSTO’s system design is conceptually based on the object-oriented modeling and design methodology (Rumbaugh et al. DatabaseManager. CommunicationManager. The EventManager generates and handles various events. The DataSocketServer sends or receives run-time data across remote processes. ControlManager.3. The CommunicationManager primarily controls most of the communications among HISSTO’s subsystems either locally or in a remotely distributed context. Figure 3. sensor values captured through the Hardware/SoftwareInterface can be posted on the DataSocketServer so that the monitoring user interface elements such as ActiveXTM controls. The WebServer is to publish HTML files in which various ActiveX controls are embed- 39 . OWL project (Bruegge B. 1995) along with the UML (Unified Modeling Language) representation (OMG 1997). the UserInterfaceManager is responsible for creating and maintaining multiple instances of UserInterfaces across the system boundary.4. EnvironmentDescriptor.
In the system modeling process. The major focus of this task is to increase the system modularity and extensibility.4. 1995) are widely used to provide reusable modular solutions for organizing various system software components.8: Component diagram describing HISSTO’s system architecture: UML notation 3. the entire object hierarchy is organized into generic model components. Hardware/Software Interface Building Monitoring/Control Hardware Sensor Act uator ignal rocessor A/D Converter HISSTOLocalInt erface HISSTODatabase Control Interfac e Monitoring Interface HISSTO ontrol anager Environment Descriptor Database Manager HISSTORemoteCommunicationServer Event Manager Communication Manager Datasocket Server eb erver System Manager Prediction Manager UserInterface Manager Monitoring Manager HISSTOWebInterface WebBrowser ActiveX Control HTML Figure 3. Decoupling implementations-specific products or processes from the generic representation is useful to enhance system modularity.ded. which could be flexibly adapted to the various control situations possibly demanding flexible system extensions. software design patterns (Gamma et al. 40 . The MonitoringWebInterface provides current systems states browsing whereas the ControlWebInterface is used for remotely changing systems states or various control parameters. The HISSTOWebInterface is a remote client accessing the system through the web and has two types of user interfaces.2 System component design The system analysis object model is further developed to accommodate various system design decisions. Since HISSTO needs to handle more than one control zone in possibly different control domains.
1995). Every User needs to be authenticated and authorized before accessing the SystemManager object services.9: Top level logical model describing HISSTO’s subsystem breakdown: UML notation The SystemManager component (Figure 3. Other methods of the SystemManager include the initialization of device control 41 . which makes the subsystem easier to use (Gamma et al.Every subsystem in HISSTO has its own API (Application Programming Interface) designed as the combination of the Singleton pattern and the Facade pattern.10) is designed in a Singleton pattern so that the only one global system instance is activated and maintained across potentially multiple instances of control/monitoring operations at any point of time. This control of access is achieved through the SecurityManagerInterface by checking an user’s personal information. bi-directional associations are established between the CommunicationManager and any of the other subsystems as is described in Figure 3. The Facade pattern defines a unified higher level interface to a set of interfaces. For the necessary interactions among subsystems.9. This is important especially to guarantee that all control actions are monitored and regulated on a globally optimized basis while maintaining tight system security. This highest-level interface makes it easier to modularize the system and to simplify the use and maintenance of each subsystem’s services. Singleton pattern ensures a class only have one instance allowing a global point of access to it. SystemManager UserInterfaceManager Co ntrolMan ager Database Ma nager CommunicationManager Environment Descriptor MonitoringManager PredictionManager EventManager Figure 3.
actions.n logOn() operateSystem() logOut() 1 1 1 SecurityManagerInterface <<virtual>> authenticate() <<virtual>> authorize() <<virtual>> editAccessList() 1 GlobalDataMonitor globalData getGlobalData() detectConflict() 1 SecurityManager Occupant FacilityMa nager se tA ccessPrevil iege () checkGl obal Da ta() checkAccessList() authenticate() 1 authorize() editAccessList() AccessList add() delete() update() 1 1.n User name : String password : String ip_address : String accessLevel : Interger 1. monitoring current systems states on the global level as well as resolving potential or detected conflicts among different threads of control actions.. By constructing Views in an ad-hoc based manner. Before performing necessary simulations.n Figure 3.11) primarily creates a building product model serving mainly as a part of the Simulator’s input data... 1995). The Builder design pattern used in this subsystem separates the construction of a complex object from its representation so that the same construction process can create different representations (Gamma et al. multiple control actions having different target systems or control zones can share a significant portion of otherwise redundant data.n ConflictResolver 1. SystemManager <<static>> _instance : SystemManager* communicationManager : CommunicationManager* userInterfaceManager : UserInterfaceManager* databaseManager : DatabaseManager* monitoringManager : MonitoringManager* eventManager : EventManager* predictionManager : PredictionManager* environmentDescriptor : EnviromentDescriptor* controlManager : ControlManager* 1 <<static>> instance() getSubSystem() controlSystemAccess() initializeSystemOperation() performSystemOperation() terminateSystemOperation() detectConflictEvent() resolveConflictEvent() performDeviceControl() performMonitoring() monitoringGlobalData() 1 1 1 1.n 1 GlobalConflictTable addItem () reso lveConfl ict () rem oveItem () 1 1..10: Logical model of HISSTO’s SystemManager component: UML notation The EnvironmentDescriptor subsystem in HISSTO (Figure 3. a 42 ..
n Space 1 Occupant System Equipment Furniture 1 Partition 1 0. If a Zone has multiple spaces in it.. A View is defined as the aggregation of ZoneComponents. an Equipment. this is described by a Composite design pattern. A Site can contain multiple Buildings as a Building consists of multiple Spaces. a System.n 1 AD_Converter Figure 3.11: Logical model of HISSTO’s EnvironmentDescriptor component: UML notation 43 ...n 0.. a View object can have ViewElements such as Schedule.n vie wT ype : String vie wList : Int eger* atta chVi ewElem ent() dett ach ViewElement () upd ate ViewElement () 1 1. an Occupant. or multiple Spaces. A ZoneComponent is either.. a Space..n 0. or a Partition. or Actuator objects.. a control operation focuses on a ZoneComponent rather than a Zone representing aggregated spaces potentially up to an entire building.n Site 1 0. a Building. n Building 1 1. A Zone could be either a Site. so is a Partition having Doors or Openings on it. EnvironmentDescriptor <<static>> _instance : EnvironmentDescriptor* communicationManager : CommunicationManager* systemManager : SystemManager* databaseManager : DatabaseManager* monitoringManager : MonitoringManager* eventManager : EventManager* controlManager : ControlManager* predictionManager : PredictionManager* userInterfaceManager : UserInterfaceManager* <<static>> instance() registerToT heSystem() getSubSystem() createView() updateView() deleteView() 1 0. Sometimes.n ViewBuildDirector 1 constructView() ViewBuilderInterface <<virtual>> buildView() 1 View ViewBuilder currentView : View* buildView() getView() 1 1. especially Sensors and Actuators. Any View object can become a dynamic entity by being attaching a subset of those elements. Throughout all levels of aggregations.n Sen sor value : Double getValue() accept() 1.. a FurnitureElement.. Sensor..View object describing the target control Zone or the ZoneComponents can be built on demand through the Builder design pattern.n ViewElement viewElementID : Integer Schedule setSc hedul e() getSc hedul e() cha ngeSche dule () Geometry setGeometry() getGeometry() changeGeometry() Actuator setValue : Double setValue() getValue() accept() ZoneView zo neID : In tege r zo neType : String zo neCo mp onentList : ZoneCom pone ntV iew* 1 ZoneComponentView zoneComponentID : Integer zoneComponentT ype : String Opening 0. Geometry. whenever it is necessary.
12) is to monitor the states of Sensors and Actuators by using the Visitor design pattern. and disseminating events within the system boundary (Figure 3.. An Event object is created by referencing the ExceptionTable generated by the MonitoringManager.. n 1 Actuator setVa lue : Doubl e setVa lue() getVa lue () accep t() Figure 3.12: Logical model of HISSTO’s MonitoringManager component: UML notation The EventManager subsystem in HISSTO is responsible for creating. handling. it is handled by an 44 .n ExceptionTable addException() removeException() 1 detectEvent() ExceptionDetectionVisit or iterationPeriod : Integer 1 visitSensor() visitActuator() createExceptionTable() 1.. defining a new operation is possible without changing the classes of the View elements on which it operates (Gamma et al.The MonitoringManager subsystem in HISSTO (Figure 3.. DeviceSateTable MonitoringManager <<static>> _instance : MonitoringManager* communicationManager : CommunicationManager* systemManager : SystemManager* userInterfaceManager : UserInterfaceManager* eventManager : EventManager* environmentDescriptor : EnvironmentDescriptor* predictionManager : PredictionManager* <<static>> instance() registerToTheSystem() getSubSystem() getTargetView() monitorException() monitorDeviceState() notifyResult() 1 0.. 1995).n visitSensor() visitActuator() createDeviceStateTable() DeviceStateVisitor iterationPeriod : Integer 1 setDeviceState() getDeviceState() 1 1 0. n TargetElement <<virtu al> > ac cept() Sensor value : Double getValue() accept() 0.n DeviceVisitor <<virtual>> visitSensor() <<virtual>> visitActuator() 1 0. n TargetView 1 1 .13). the Visitor pattern activates the activation of those operations designed for detecting device state as well as any abnormality on a TargetElement. Once the entire detection process is completed.. With the Visitor pattern. The Visitor design pattern represents the operations to be performed on the elements of a View object structure. While traversing down through the TargetView object hierarchy. Once an Event is identified. each Visitor object creates a DeviceStateTable or an ExceptionTable to be passed to the other subsystems such as the EventManager or the PredictionManager.
..n attach() dettach() notify() handl eEvent() 1 1 1 n <<vi rtual>> handleEvent() Al ertEventHandler handleEvent() EventObserver observerState : Double update() 1. 1995). or a FailureEvent.n EventRecord 1 1 Excepti onT able addExcept ion() rem ov eExc ept ion() detect Even t() 1 EventHandler Gl obalEv entHandler 0. an Event can be classified as an AlertEvent.n Event eventSourceState : Double 1.n Fa il ureEvent fail ureEventCode : String 1 getEventSourceState() setEventSourceState() Figure 3.EventHandler object the process of which is regulated by the Chain of Responsibility design pattern.. Every Event detected and handled is logged and saved in the database for future reference. This design is to avoid coupling an Event to its handler by giving more than one EventHandler object a chance to handle that Event through chaining the EventHandlers and pass the Event along the chain until an EventHandler handles it (Gamma et al. EventM anager <<static>> _instance : EventM anager* com m uni cationM anager : Com m uni cati onM anager* system M anager : System M anager* userInterfaceM anager : UserInterfaceM anager* databaseM anager : DatabaseM anager* m onitoringM anager : M onitoringM anager* environm entDescriptor : Environm entDescri ptor* 1 <<static>> i nstance() registerT oT heSystem () getSubSystem () generateEvent() handl eEvent() notifyEvent() updateObserver() logEvent() 1 1 1 EventLog 1 1 1 0..13: Logical model of HISSTO’s EventManager component: UML notation 45 . 1995).. n 1 Al ertEvent alertEv entCode : St ring getEve ntSource Stat e() setEven tSource Stat e() Change Event changeEventCode : String 1 getEventSourceState() setEventSourceState() ChangeEventHandler handl eEven t() FailureEventHandler handleEvent() 1.. all its EventObservers are notified and updated automatically (Gamma et al. n Ev entO bse rverInterf ace <<virtual>> update() 1 1 0. Based on its characteristics. Event notification process in HISSTO is designed in the Observer pattern which defines one-to-many dependency between the Event and the EventObserver objects so that when one Event object changes state. a ChangeEvent.
14: Sequence diagram showing HISSTO’s event handling process: UML notation The PredictionManager subsystem in HISST (Figure 3. but it can be the other domains as well depending on the control target devices and systems. For the implementation of the MachineLearner. there is no need to distribute inter-task calls for those objects. MonitoringManager LouverActuator: Actuator AlertEvent: Event EventManager AlertEventHandler: EventHandler AlertEventLog: EventLog AlertEventObserver: EventObserver requestState() reportState() checkState() createException() createEvent() identifyHandler() handleEvent() checkMemory() checkSens or() confirm() confirm() createLog() notifyEvent() getEventState() update() Figure 3. A Predictor can be composed of multiple PredictorComponents such as Simulators and MachineLearners collaborating based on a specific hybridization scheme. The Builder design pattern is used to structure the Predictor creation process. Both the Simulator and the MachineLearner are internally using procedure calls. 46 .14 shows a typical sequence of interactions among participating objects in HISSTO’s event generation and handling process. HISSTO uses NeuralNetworks which need training after being created but before it is activated for the actual prediction tasks. The Simulator in HISSTO currently is focused on the lighting domain.15) provides an interface for assembling the Predictors and performing predictions for the control outputs needed to evaluate various candidate control options. The design of this process is based on the Exception Monitor pattern dealing with exceptions within the scope of an individual collaboration (Douglass 1999). The design of the PredictionManager in HISSTO allows dynamic assemblies of the PredcitorComponents and their services depending on the control request within various contexts in a building.Figure 3.
n MachineLearner 1..n 1 1 NeuralNetwork Encoder Lighti ngS im ul atorB uiler EnergySim ula torB uilder AirFlowSimulatorBuilder 1 1 compute() learn() 1 n Training Pat tern 1 1.n Simulator 1 predictorID : Integer 1.n getResult() MachinLearnerBuilder buildPredictorComponent() getResult() 1....n Layer 1 1..PredictionManager <<static>> _instance : PredictionManager* communicationManager : CommunicationManager* systemManager : SystemManager* userInterfaceManager : UserInterfaceManager* databaseManager : DatabaseManager* monitorManager : MonitorManager* environmentDescriptor : EnvironmentDescriptor* <<static>> instance() registerToThe System() getSubSystem() getSimulator() getView() getMachineLearnerTrainingPattern() buildPredictor() updatePredictor() performPrediction() 1 SimulatorInput 1 0...n Predictor 1 PredictorDirector predictorID : Integer constructPredictorComponent() 1 MachineLearnerTrainer machineLearnerTrainerID : Integer getTrainingPatterns() peformTraining() saveTrainedMachineLearner() 1 assemblePredictor() runPredictor() calibratePrediction() 1 1 1.. n 1 Si mula torBuil der buildPredictorComponent() 0.n Link we ight : Do uble updat eWeig ht() Figure 3.n PredictorComponentBuilder predictorComponentID : Integer <<virtual>> buildPredictorComponent() 1 0.15: Logical model of HISSTO’s PredictionManager component: UML notation 47 . n 1 1 PredictorInput 1 Machi neLea rnerInput 1 Calibrator 0. n Node 1 ThermalSimulatorBuilder Ac oust icS im ul atorB ui ld er 1....
n ControlCommand <<virtual>> execute() 1 1 1 DeviceControl Command state : String execute() Ma nualCon trol getUse rInput() issu eCo mm and () PredictiveControl performPrediction() testControlOption() issueCommand() 1 1 1.The ControlManager subsystem in HISSTO (Figure 3.n Prediction predictControlOutcome() Tester testControlOp tions() 1 1 TestAlgorithm runTestAlgorithm() Figure 3.. the PredictiveControl object triggers a series of actions involving the Predictor.n ControlStrategy 1 <<virtual>> issueCommand() 1 0. and the Commander objects.n ScheduledControl identifySchedule() IssueCommand() 1.n 1 Co mm and er executeControl() 0.n 1.16) prepares and performs the actual control operations..16: Logical model of HISSTO’s ControlManager component: UML notation 48 .n Control control_ID : Integer selectControlMode() performControl() 1 createControlLog() 1 0.....n 1 Inv oker 1. Different courses of a control action are provided based on the Strategy design pattern which defines a family of the ControlStrategies and make them interchangeable (Gamma et al. 1995). For example. and support undoable operations (Gamma et al 1995). queue or log requests. the Tester. Each ControlStrategy is also a Facade so that the complex implementation details can be encapsulated. ControlManager <<static>> cotrolManager : ControlManager* communicationManager : CommunicationManager* monitoringManager : MonitoringManager* eventManager : EventManager* environmentDescriptor : EnvironmentDescriptor* predictionManager : PredictionManager* databaseManager : DatabaseManager* systemManager : SystemManager* userInterfaceManager : UserInterfaceManager* <<static>> Instance() registerToTheSystem() getSubSystem() getTransducer() getControlZoneView() getPredictor() getWeatherData() getUserPreference() createControl() logControlHistory() 1 Con trolLog 1 0. This Strategy pattern lets the control algorithm vary independently from clients that use it...n 1 Device Actuator 1 1 deviceID : integer actuateDevice() reportDeviceStateChange() 1. The way a ControlCommand is carried out in HISSTO is organized through the Command design pattern which encapsulates a control request as an object in order to parameterize clients with different requests..
n DatabaseTenant dataHandle : Long getData() setData() searchData() UserPrefere nce Data TransducerData Machi neLearner TrainingPattern WeatherData ControlSchedule En viron me ntDesc ri ption Dat a ControlHistoryLog Neural Network ei ght Da ta EventLog Figure 3. The Transaction object in this model is created by the DatabaseManager and has only a limited lifetime until it commits or rolls back after performing the requested data operations. Every Connection has one of three different stages at a give point of time. and CommunicationClosed.17: Logical model of HISSTO’s DatabaseManager component: UML notation The CommunicationManager subsystem deals with messaging and data exchange services among HISSTO’s components (Figure 3.The DatabaseManager subsystem in HISSTO (Figure 3. CommunicationListen. a Connection is handled either in a remote mode or in a local mode. namely.18)..17) provides a single interface for all data transactions invoked by the persistent data storage requests. This subsystem maintains concurrency and consistency throughout global or domain-specific data operations on multiple DatabaseTenants within a distributed environment.. A Connection object is created when a communication request is issued and the parties participating in the communication are identified. Depending on the actual distribution of all connection parties.n Transaction transactionID : Long state : String save() delete() update() query() 1 1. The Strategy design pattern used 49 . DatabaseManager <<static>> _instance : DatabaseManager* communicationManager : CommunicationManager* systemManager : SystemManager* userInterfaceManager : UserInterfaceManager* monitoringManager : MonitoringManager* eventManager : EventManager* environmentDescriptor : EnvironmentDescriptor* controlManager : ControlManager* predictionManager : PredictionManager* <<static>> instance() registerToThe System() getSubSystem() createTransaction() terminateTransaction() 1 0. CommunicationEstablished.
and SlideBars. This pattern lets the UserInterfaceManager treat individual Window objects and compositions of Windows uniformly (Gamma et al 1995).18: Logical model of HISSTO’s CommunicationManager component: UML notation The UserInterfaceManager subsystem in HISSTO (Figure 3.n Connection connection_ID : Integer connectionState : Integer connect() monitor() disconnect() 1 1 ConnectionImpl 1 1 CommunicationState <<virtual>> open() <<virtual>> close() <<virtual>> acknowledge() RemoteConnectionImpl LocalConnectionImpl 1 1 MessagingImpl Communication Established open() close() acknowledge() Communication Listen open() close() acknowledge() Communication l osed open() close() acknowledge() TCPConnectionImpl Figure 3.19) provides multiple ways of constructing interfaces mainly to monitor and control building systems states for different types of users. The key element in this subsystem is the Window object which is composed of multiple UserInterfaceComponents such as Buttons. EditBoxes.. The Composite design pattern is useful in organizing objects into a tree structure to represent part-whole hierarchy.here allows the Connection object to alter its behavior when its internal state changes (Gamma et al 1995). A 50 . This design of the CommunicationManager is motivated to keep the system flexible in dealing with both distributed and central system integration scheme. Co mm uni cationManage r <<static>> _instance : CommunicationManager* systemManager : SystemManager* controlManager : ControlManager* environmentDescriptor : EnvironmentDescriptor* predictionManager : PredictionManager* eventManager : EventManager* monitoringManager : MonitoringManager* databaseManager : DatabaseManger* userInterfaceManager : UserInterfaceManager* <<static>> instance() registerToT heSystem() getSubSystem() identifyCommunicationParties() createConnection() monitorConnection() terminateConnection() 1 0.
.. Depending on the characteristics of the detected event.n Bro wse rWi ndow getData() EditorWindow setData() Button ComboBox Kno b SlideBar Chart EditBox ListBox ScrollBar EventBrowserWindow Figure 3.19: Logical model of HISSTO’s UserInterfaceManager component: UML notation 51 . UserInterfaceManager activates/deactivates corresponding methods on the pertinent UserInterfaceComponents. Assembling Windows on-demand is a useful strategy when the system needs heterogeneous user interfaces varying dynamically based on the types of interactions in different contexts.Window object can be either a BrowserWindow or an EditWindow depending on the use case it is supposed to satisfy. Each UserInterfaceComponent is responsible for displaying or receiving information necessary for the system functionality. UserInterfaceManager <<static>> _instance : UserInterfaceManager* communicationManager : CommunicationManager* systemManager : SystemManager* databaseManager : DatabaseManager* monitoringManager : MonitoringManager* eventManager : EventManager* predictionManager : PredictionManager* environmentDescriptor : EnvironmentDescriptor* controlManager : ControlManager* <<static>> instance() registerToTheSystem() getSubSystem() createUserInterface() getUserInput() displaySensorVal() displayEvent() deleteUserInterface() 1 Coordinate x : Integer y : Integer width : Integr height : Integer 1 1. The UserInterfaceManager detects various user interface events derived from the system’s use case model. n 1. There can be nested Windows as is indicated by the Composite design pattern.n Window windowID : Integer zoom : Integer assembleWindow() displayWindow() updateWindow() deleteWindow() resizeWindow() 1 1 UserInterfaceComponent uiComponentID : Integer create() update() delete() display() 0..
If training is necessary to update a selected Predictor. So.4 Hardware/software mapping Eventually. and the UserInterfaceManager subsystem are always activated. The SystemManager. prediction. a separate thread should be given to this training process while maintaining control operation in a different mode. For this reason. In the PredictorManager. the EventManager. the CommunicationManager. test. HISSTO must collect sky condition data while performing event detection process allowing users to monitor the current systems states. there certainly are hardware/software dependencies. The entire MonitoringManager subsystem is designed to decouple the system from vendor-specific Sensors or Actuators so that HISSTO can have flexibility and extensibility for potential infrastructure changes. the Simulator and the Machine Learner are all portable as long as the operating system supports C/ 52 . The ControlManager subsystem should be able to handle multiple control requests concurrently each of which can be in one of those five states: training. only a single on-going global control instance is allowed to be activated and all other localized controls are multi-threaded from it. Considering potentially multiple computation-intensive processes at a given time.4. At any given time. at the instrumentation level. CPU and memory management are crucial for a successful system operation. multi-threading is absolutely required for this system to guarantee desired system operation. The tactics for this are: • • Reducing computational load whenever it is possible Reducing concurrency by activating computation intensive processes in a serial fashion rather than in parallel one unless it is absolutely necessary • Use of multi-threading when concurrency is inevitable To avoid conflicting concurrency which eventually should be globally regulated. The SystemManager.3. There should be careful software design to handle potential computational overload. there are at least five threads running. actuation or idling. Occupants and Facility Manager are expected to interact with HISSTO through the internet which provides a platform-independent remote access.3 Concurrency identification Concurrency is one of the most important design considerations for HISSTO’s system development. the CommunicationManager. and the UseInterfaceManagher must have separate threads so that they can run concurrently. Most data acquisition or control devices come with hardware plus software designed to be operated on either a PC or an UNIX machine. For example. 3. the MonitoringManager.4.
For HISSTO’s local interface. 53 . There can be another way of achieving this such as the combination of JAVATM and CORBATM technology enabling web-based monitoring and control as is demonstrated in the OWL project (Bruegge and Dutoit 1999). and compatibility. and simulations are all memory intensive tasks. HISSTO’s dependency on the ActiveX technology is imposed by the Labview programming environment responsible for performing most of the instrumentation tasks. machine learner training.21 shows the physical connectivity of HISSTO’s instrumentation within a real building.20 illustrates the topological structure of HISSTO’s hardware components and Figure 3. Both MonitorVI and ControlVI maintain connections to the client side ActiveX controls embedded in the HTML documents and loaded from an ActiveX container such as MicroSoft ExplorerTM web browser. The necessary communications between the remote ActiveX controls and the remaining HISSTO subsystems is done by the WebServer (for publishing HTML files) and DataSocketServer (for bi-directional remote information exchanges). simultaneously. the overall system performance is influenced by many factors. In terms of performance issue.C++ compilation. the dimmers. ActiveXTM technology combined with Data socketTM communication is used. Each VI accomplishes the connection to the core of HISSTO written in C/C++ through the library calls to the program compiled in DLL (Dynamic Link Library) form. reducing computational load for each job and minimizing concurrency for those tasks are absolutely necessary. Considering cost. performance. and the number of simulation sessions for prediction are all contributing factors to the system performance. Data acquisition. the size of machine learner training patterns and steps. the software/hardware integration is made by using LabviewTM programming environment. Figure 3. A VITM (Virtual Instrument) is an unit program component which can be a collection of other VIs. therefore. the PC is a desirable choice as the target platform for HISSTO. This development environment uses a graphical programming (G programmingTM) technique which allows a programmer to drag/drop and connect graphical symbols representing specific programming elements. Each VI has a front panel and a source code editor with which a programmer can build code and user interface. and the light sensors. allocating a high performance CPU for those tasks is also critical. The UseInterface for HISSTO is designed to monitor the sensor values and to actuate the target systems from an web browser. To make this possible. For the same reason. The MonitorVI and the ControlVI can receive/send signals through a PC board or a serial port on the computer from/to the transducers such as the louver. This is a "location-independent cross-platform" solution with a proper security provided. The sampling rate of data acquisition.
4.5 Data management Data management in HISSTO is important because the system is heavily dependent on the various forms of data. This includes sensor data which 54 . Persistent data is any data which maintains its state across system reboot. HISSTO will be likely to generate and handle a significant amount of persistent data.20: Deployment diagram describing HISSTO’s hardware topology: UML notation Daylight Station RS-232 Louver controller M Data Ballast PC W/S AT-MIO/AI E board Analog signal Photometric Sensors Figure 3. HISSTOClient HISSTOServer manual Listen To Events Issue Commands executive Poll Sensors Evaluate Control Options Generate Control Commands Publish Device States Listen To User Inputs A/D Converter nonpreemptive Sensor Actuator Convert A-D Convert D-A Process Signal Louver Luminaire Figure 3..21: Physical connectivity for HISSTO’s system instrumentation 3.
For example. SytemManager 1. The IP address of the machine is automatically detected.n logOn() operateSystem() logOut() 1 1 SecurityManagerInterface <<virtual>> authenticate() <<virtual>> authorize() <<virtual>> editAccessList() 1 1 SecurityManager Occupant FacilityManager etA ccessPrevil iege () heckGl obal Da ta() checkAccessList() authenti cate() 1 authorize() edit Acce ssList() AccessList add() delete() update() 1 Figure 3.. Depending on the sampling rate. Each sensing or actuation device also should be given a unique ID so that the ControlManager can access the right device whenever it is necessary. the SecurityManager checks the AccessList to authorize the User’s access request.is kept for months. 3. multiple sensor readings need to be buffered before being processed for obtaining useful sensor data. corresponding control services are made accessible by the SystemManager object (Figure 3. an User object is instantiated asking for the user ID and the password. A final category of data can be characterized as semi-persistent data.6 Global resources management Security issue in HISSTO is critical especially for protecting the system from non-authorized users. Other persistent data includes simulation inputs and outputs since this data can be used for machine learner training purpose. Then.4. The system builds a number of volatile and temporary data stores. predictions done by a simulator or a hybrid predictor for each control interval are temporal so is the evaluation result of each control option based on the performance indicators and preference functions. if not years. When a HISSTOWebInterface is called.n User name : String password : String ip_address : String accessLevel : Interger 1.22). This is a common concern for any web-based information system development..22: Access privilege management in HISSTO: UML notation 55 . This data is persistent in its nature but is not guaranteed as such. Depending on the User’s access privilege. An example of this is the weight value for each link in neural network which should be maintained for the initialization of a specific neural network until it has to be updated.
To realize such an event driven system. handling various events in an organized way is crucial for maintaining each subsystem truly independent. Once the thread is spawned. The design of each subsystem's internal control should in no way affect the design of the other subsystems.23). An event driven system has the advantage over the procedure-driven system in term of its simplicity. each procedure’s state needs to be preserved. the system must collect sky condition data from the sensors while allowing the user to raise the dimming level of the luminaires at the same time.3. This is because events could easily complicate on-going operations across different subsystems if they are not handled and notified properly. In the EventManager subsystem. 56 . the EventManager spawns a thread to handle that request. and better error handling capability (Rambaugh 1991). the need for concurrent system control comes from the fact that most subsystems in HISSTO should be able to run in parallel. This design guarantees automatic update of the critical information for control and maintenance of the building systems while reducing the complexity of distributing the impact of an event across different subsystems. For example. modularity. On the other hand. In an event driven system. This suggests that the system needs to be multi-threaded. For example. as an abstract class. The EventObserverInterface. Threads allow the system to do multiple processes. flexibility.7 Software control implementation External control among the subsystems of HISSTO is based on the combination of an event driven and a concurrent system. initialization of simulations or a machine learner training. Internal control is specific to each subsystem.4. the system is free to process other requests even before the first one is completed. the software control is always back to a “dispatcher” which regulates all external control flows. In HISSTTO. the response to a direct user control request etc. a predictor update process after detecting a change event. provides an interface among various Events and EventObserver objects so that any other subsystem can access the caring events through it. Each time an event occurs making a certain action to be performed. (Figure 3. all detected events are constantly observed by the EventObservers in terms of the states of the event sources and their implications. It activates or delegates specific sensing or actuation. the SystemManager component acts like “dispatcher” controlling all other subsystems. An event driven system is the proper choice because the entire system can be organized as the collection of separate modules and components and their dependencies are not pre-defined but based on ad-hoc events.
Thus. one control mode can be activated anytime but there will never be the case where more than one control mode is activated for a specific control thread. the ManualControl is activated. the desired control action is directly passed to the DeviceControlCommand object. When it is the ScheduledControl mode. The ManualControl needs not to invoke any control outcome prediction or evaluation process. This subsystem invokes a Control object in a combination of polling and event handling for process control. One thread which polls sensor data through an instance of the MonitoringManager. These different strategies of a control thread are mutually exclusive. The control mode transitions based on the user’s selection rely on the Strategy design pattern (Gamma et al.. the PredictiveControl takes initiative which is handed over to the ScheduledControl during off-work hours. A typical PredictiveControl process involves the outcome predictions and evaluations of the candidate control options followed by the execution of the decided systems state changes. one thread that actuates the ballasts or the louver motor. During office hours. the related building description Views and the Predictors should be updated via updated simulation outputs and the machine learner 57 . When a certain change occurs. For example. and one thread that deals with the system state change based on the outcome predictions and tests of various control options or the request a user has made. the control actions is guided by a pre-determined system operation schedule.23: HISSTO’s event notification process based on the Observer design pattern: UML notation The ControlManager subsystem’s internal control deals with various processes. This component is decomposed essentially into three layers..n 1 AlertE vent alertEventCode : String getEventSourceState() setEventSourceState() FailureEvent failureEventCode : String getEventSourceState() setEventSourceState() ChangeEvent changeEventCode : String 1 getEventSourceState() setEventSourceState() 1 Figure 3. When an user selects a manual control mode.n 1. 1995). The ControlStrategy object delegates control actions based on the three different control strategies.Event EventObserverInterface <<virtual>> update() 1 1 eventSourceState : Double attach() dettach() notify() handleEvent() EventObserver observerState : Double update() 1.
yet effective way to guarantee user satisfaction and energy conservation in building operation when a proper algorithm is deployed.retraining process.25 shows the possible transitions among three different building systems control modes.24: HISSTO’s alternative control strategies based on the Strategy design pattern: UML notation Figure 3. the control could be done by a pre-defined schedule which dictates minimum control actions necessary until it is overridden by any user for his/her extended work in the space. . 58 .24). A manual control request can interrupt any other control mode and remains so. The Strategy design pattern for each control strategy helps to encapsulate the control-mode-specific implementation detail (Figure 3. The default control mode is always the PredictiveControl mode considering the fact that it can be the most sophisticated. During off-work hours including weekend time. as long as the user maintains the control initiative. Another case in which the transition to the ScheduledControl mode needs to be made is when the Predictor requires update due to a building configuration change or other event which makes the current Predictor’s knowledge invalid. Control control_ID : Integer ControlStrategy selectControlMode() performControl() 1 createControlLog() 1 <<virtual>> issueCommand() ManualControl getUserInput() issueCommand() PredictiveControl performPrediction() testControlOption() issueCommand() ScheduledControl identifySchedule() IssueCommand() Figure 3.
Idle do/ Check Next Schedule [ Update f inished ] Prediction en tr y / Ch eck Curre nt Stat es do / Perf orm Pr edi ctions ex it / Sel ect Control Opt ion Actuation do/ Set State exit/ Update Current State [ Time ma tche d ] Select Automatic Control Manual Control User Ov erides Control Idle do/ Wait f or User Input Actuation do/ Set State exit/ Update Current State [ User input detected ] Figure 3. Various Devices are the potential receivers of the DeviceControlCommand.Start Sy stem entr y / Tur n O n Sy ste m do/ Dis pla y Control Edi tor exit/ Enga ge C ont ro l Building Operation Automatic Control Iterate Cy cle[ iterationPeriod =600 sec ] Predictiv e Control Scheduled Control [ Predictor needs update ] I dl e do/ Wait Until Next Cy cle Actuation do/ Set State exit/ Update Current S. it could be the sampling rate change Command. 59 .25: State transition diagram describing HISSTO’s control mode changes: UML notation The other design decision regarding the internal control of the ControlManager subsystem is the way it handles a control command suggested by the Commander object. The Execution object creates DeviceContorlCommand object which is stored and invoked by the Invoker object.. For a Sensor. The Command design pattern allows delayed or conditional execution of the given control commands (Figure 3..26).
Device Actuator 1 deviceID : inte ger actuat eDevic e() re port DeviceStat eChange () 1..n 1 Commander executeControl()
0..n ControlCommand <<virtual>> execute()
Device Con trol Command state : String execute()
Figure 3.26: HISSTO’s handling of control command based on the Command design pattern: UML notation
The user interface of HISSTO uses event driven process control. There is one thread for each of the controls on the screen at once so that the user interface can continue to operate even if one of the controls is waiting on a blocking call to another subsystem. There is also a thread of control corresponding to each input modality such as mouse/keyboard. Finally, there will also be a thread to handle the communications with other subsystems.
The first step in the start up of HISSTO is to initialize the SystemManager subsystem. Then the ComunicationManager subsystem must be initialized followed by the MonitoringManager and the EventManager subsystems. The UserInterfaceManager is the last subsystem to be initialized. Since this subsystem
handles user interface components on the web, the WebServer and the DataSocketServer should be started first before a user can actually access HISSTO’s remote user interfaces. Initialization of each subsystem involves starting its internal data structures and any additional resources which it may need. For example, before the ControlManager is performing control decision making, it must obtain sensor readings through the MonitoringManager subsystem. It is possible that a subsystem may attempt to invoke a service which is not yet registered. In this case, the registry will raise an exception to inform the subsystem that the particular service is not yet available. This exception must then be handled on an individual basis by the invoking subsystem.
HISSTO’s termination process will be initiated through the user interface. Once started, the system will terminate in the following order: 1) ControlManager 2) MonitoringManager 3) UserInterfaceManager 4) and finally, CommunicationManager. Even though this is a possible termination process, it could rarely happen except when the SystemManager needs to interrupt controlled operation for the maintenance purpose. Most of the time, at least the SystemManager, the ControlManager, the MonitoringManager, and the EventManager are running while other subsystems are activated or deactivated depending on the event captured by the EventManager. The user interface could be separately terminated by closing web browser, which doesn’t have impact on the operational status of the remaining subsystems. When any given subsystem terminates, it removes all public services from the registry. This may include, but is not limited to, freeing of memory, de-allocating network resources, and releasing any mutual exclusion locks that certain components of a subsystem may be holding.
HISSTO consists of many subsystems, any of which may fail at any time. No failure can be properly dealt with unless it is detected, but once a failure is detected, the system can take the appropriate measures to correct the problem. A failure in the subsystems other than the UserInterfaceManager will not bring down the entire system. In this case, an error message on the browser notifies that the user interface is not accessible, and the user will not be able to interact with the control system. If the ControlManager does not function properly to yield a control decision, the system automatically switches itself to either manual or scheduled control mode, until the problem is identified and fixed. This could be also the case when all or part of the sensors are not working. If data acquisition is interrupted for long time because of sensor failure, the machine learner’s retraining might be necessary with the new data set after sensing is restored. To restore service after a subsystem crash, the system should follow its initialization procedure. All failures should be as limited in scope as possible. When failures occur, those parts of the system that can remain functional will do so, and restoration of services will take place as smoothly and rapidly as possible.
Adaptation to the system changes
Any information system is expected to be under modifications primarily because multiple sources of change (Bruegge and Tutoit 1999). New vendor or new technology enforces system change by replacing system components; New implementation could be necessary to enhance system performance; New views might be needed for solving usability problems. Therefore, the creation of additional views on the same data is required. HISSTO has a subsystem which describes building configuration. This building configuration has multiple Views and is subject to changes based on the updates or the replacements of building components. In HISSTO, a View is an object describing geometrical and physical properties of the target Zone or the ZoneComponents. This View object can be very complex. By applying the Builder design pattern, construction of a View can be separated from its representation. A client such as the EnvironmentDescriptor instantiates a ViewBuilder as well as a ViewBuilderDirector. The ViewBuilderDirector does a View construction through the ViewBuilder object and the EnvironmentDescriptor gets this View object using getView() method (Figure 3.27) .
Enviro nmentDescriptor 1
ViewBuilderInterface <<virtual>> buildView() 1 1
1 ViewBuildDirector co nstructV iew()
Vi ew Vi ewBuilde r
currentView : View* buildView() getView()
viewType : String viewList : Integer* attachViewElement() dettachViewElement() updateViewElement()
ViewBuildDirector new ViewBuilder
new ViewBuildDirector constructView() buildView()
Figure 3.27: Construction of a View based on the Builder design pattern: UML notation
3. For example.10 Design priorities The system design priorities in HISSTO are summarized in Table 3.4. should the need arise. performance and accuracy are two potentially conflicting criteria. the level of detail in describing building for the simulator influences both the accuracy of prediction and the speed and easiness for preparing the predictor. Subsystem-level design trade-offs are specific to that subsystem's functionality. and larger maximum training steps for a Machinelearner can potentially increase accuracy while sacri- 63 . but should also be portable to multiple platforms. Portability The client portion of the system should be usable on multiple platforms. Location transparency In order to facilitate portability and ease of use.4. efficiency: The control process should be handled completely.4: Design priorities of a typical building control system Criteria Correct functionality Efficiency Description The control system should be as bug-free as possible.4. • Specificity vs. and should be easily ported to a new architecture. It should be efficient in order to maximize the system’s effective life-span. System-level design trade-offs are identified as follows: • Functionality vs. increased patch numbers in lighting simulation. the actual location of the server on which the "core" of the system is running should be transparent to the clients. a situation where one design goal must be sacrificed in order to achieve another. The system should responds to user input and facilitate efficient solutions to problems. Extensibility The control system should be extensible enough that it will eventually be able to handle even multiple control domains in a building. Reuse of components The individual subsystem should be easily extractable and applicable to other systems if similar functionality is required. generality: The system should run efficiently according to the user's preferences.: Table 3. but it should also be handled quickly. More frequent sampling rate for the sensors. The system should be relatively self-explanatory so that the amount of time required to become familiar with it is minimized. 3. Fault tolerance Ease of use The system should be resilient to individual server and client failure.11 Design trade-off Established design goals can sometimes lead to an impasse. For the MonitoringManager and the ControlManager subsystem.
5. The advantage of a simulation-based prediction is that the simulator can calculate those performance indicators (i. Computational and time intensiveness is on its disadvantage list. and the Commander. the choice should be made depending on the individual case considering possible tolerance for each criterion.ficing performance. 3. HISSTO’s Predictor has two sub-components: the Simulator and the MachineLearner. glare and solar gain) that can not be measured in most of actual building settings.1.5 Object design Object design of HISSTO focuses on the detail design of the Predictor. Those sky condition variables are obtained from either measurement or the local weather file.1 Object design of the Predictor The Predictor is a critical part of HISSTO which provides intelligence for control decisions. The first group includes indoor illuminance values on the specified reference points. Louver angle and luminaire dimming levels in the target space are also among simulation input variables. The outputs of the simulation can be classified into three categories. three sensor values (global and diffuse horizontal irradiance and global horizontal illuminance) are included as inputs. Since both of these two criteria are important. The second group consists of the reference parameters mostly for glare calculation. LUMINA is used to predict the values of the daylight performance indicators for candidate control options. the Tester. 3. Figure 3.28 explains the viewer specific simulation input variables for calculating glare indices chosen in HISSTO. Third group represents seven visual performance indicators intended to be used for visual environment control options evaluation. To inform LUMINA of the current sky condition. LUMINA itself can be used as a predictor for daylight responsive lighting control.e.1 Object design of the Simulator In HISSTO.5. Developing a test algorithm to evaluate the control options generated by the Predictor is the main task of the Tester design. 64 . With properly filtered data set and sufficiently calibrated simulation parameters. structuring the ways in which a building control operation is performed using different hybrid predictors The Predictor design task includes the detail design of the hybridization processes between a simulator and machine learners and the training process of the resulting hybrid predictors. 3.
Both internal and external surfaces are defined by the x. environmental variables (E0-E5) and the system variable (S1) are the inputs. Site information should be given in terms of its latitude.28: Viewer information necessary for LUMINA’s glare calculation A simulator needs a more sophisticated set of inputs than a machine learner due to the descriptive nature of its modeling process. For glare calculation. Table 3.5 shows the variables selected for the inputs and the outputs of LUMINA as a predictor in HISSTO. specularity. Among all listed variables. and diffuseness. 65 . Windows are separately defined with a similar set of properties. reference parameters (R1-R17) and performance indicators (P1-P7) are the outputs.y.z) view direction x-axis azimuth y-axis Figure 3. y. LUMINA. and z coordinates of the vertices along with their physical properties such as transmittance. reflectance. Reference points on which the illuminance values are to be calculated should be given before running the simulation. as a combined daylight and electrical lighting simulation program. longitude. both the position of the viewer and the view angle need to be specified. Luminaire schedule along with the IES (Illuminance Engineering Society) formatted luminaire information file is also a necessary part of inputs. etc. height from the sea level.altitude viewer(x. needs to have the description of a building and its systems as a part of inputs.
developing an efficient way of calculating electrical light is required.Table 3. In HISSTO.5: Input and output parameters for LUMINA ID E0 E1 E2 E3 E4 E5 S1 R1-12 R13 R14 R15 R16 R17 P1 P2 P3 P4 P5 P6 P7 Variable Unit Month Day Hour Description m d h I hd I hg Ehg LRL En L b1 L b2 Ed Ei L max Em Uf DGI CGI G CRT Q P W⁄m W⁄m lx degrees 2 2 Horizontal diffuse irradiance Horizontal global irradiance Horizontal global illuminance Light redirection louver angle Indoor illuminance value for the nth reference point Background luminance for DGI calculation Background luminance for GCRT calculation Illuminance on the eye ball due to the direct light component Illuminance on the eye ball due to the indirect light component Value of the maximum luminance patch on the CRT screen Average illuminance across all indoor reference points Uniformity factor in the target space Daylight Glare Index CIE Glare Index due to the electrical light Glare on the computer screen Total solar radiation through the glazing area Total power consumption by the electrical light lx lm lm lx lx lm lx - W W One of the most important issues in the simulator design is how to maximize its search in the system state space within a given amount of time. the Luminaire Matrix is devised to avoid otherwise too heavy computation for electrical light simulations. Since the potential combinations of the possible luminaire dimming states can grow exponentially (assuming there are multiple luminaires in the control zone). 66 . The solution for that is to separately deal with electrical lighting calculation and daylight simulation.
Measurement-based neural networks in HISSTO are classified into this category. Since obtaining actual sensor readings for those performance indicators is difficult. On top of those two categorization of neural network usages.5. To train this neu- 67 . • Copying Existing Model (CEM) is another way of designing a neural network which can be trained from the data set generated by an existing model or a human expert.2 Object design of the MachineLearner The MachineLearner objects for HISSTO are all supervised learners implemented in neural networks using back-propagation as the weight-tuning rule and sigmoid function for treshold function. Since simulation can generate all output parameters necessary for HISSTO’s prediction capability by having almost the same inputs most of the neural networks require.3. For this reason. Different designs of the neural networks provide exploratory benefits in the hybridization process. At least two major ways of the hybridization are covered: direct knowledge transfer from the Simulator to the neural networks through a copying mechanism and the neural network assisted simulator calibration.e. a neural network has to rely on the simulator for predicting the values of such indicators. Since a Simulator can provide prior knowledge for a MachineLearner. especially for minimizing computational load and time otherwise necessary for simulations. it is necessary to have the neural networks trained and operable with only limited number of indoor illuminance sensors. Estimated system parameters by this method can be used for system behavior prediction or system control purpose.1. • Number of indoor illuminance sensors is an issue because it is not feasible to use a large number of sensors in a target space beyond certain period of data collection phase. glare indices and solar gain through glazing) creates a challenge. a couple of other issues are identified in selecting proper neural network designs for HISSTO. supervised learning is the best choice compared to others. This neural network obtains average illuminance and uniformity factor calculated from the direct measurement or the predicted indoor illuminance values by other neural networks. • Neural network’s lack of access to a certain group of visual performance indicators (i. • System Identification (SI) is the process in which a neural network identifies system behavior by observing system’s inputs and outputs. copying simulation knowledge into a neural network becomes an attractive option. then predicts the rest of outputs (both reference parameters and remaining performance indicators) based on the training done with weather file-based simulations.
a simulation based neural network in CEM category. Even though it can only predict a single reference point illuminance without being able to deal with average illuminance and uniformity factor. Generation of the training patterns for the SMNN needs a dedicated period of time during which the measurement of sky condition parameters (E3-E5) is performed. This neural network predicts all reference parameters and performance indicators necessary for HISSTO. 6) SMCNN in CEM category is the calibration neural network designed to learn the simulator’s indoor illuminance prediction errors.ral network. simulation generated training patterns must be used. seven different neural networks are created. Considering those MachineLearner design issues in HISSTO. 68 . because only the simulator can generate the values of those unmeasured performance indicators. The expected advantage of this neural network is the agile preparation of training patterns as long as the pertinent local weather file is acquired. this neural network is trained with the training set generated by the local weather file-based simulations. Each of the neural networks designed in HISSTO varies mostly by its inputs and outputs as well as the source of the training patterns. Similar to WFNN(1). 3) SMNN. it provides a fairly scalable solution in the real lighting control setting because placing a large number of sensors in any control zone is not always a practical solution. uses the training patterns generated by the simulator with the recent sky condition measurement. It predicts only a subset of performance indicators such as average illuminance and uniformity factor which can be calibrated by the indoor illuminance sensors during its operation time. This neural network is designed to replace the simulator on condition that its predictive capability matches that of the simulator within an acceptable deviation range. 1) MSNN(1) in SI category uses measurement data for training and predicts indoor illuminance distribution across multiple reference points. 2) MSNN(2) in SI category uses only one indoor illuminance sensor for its training. Once trained. The accuracy of prediction can be the major issue with this neural network since it does not have any chance of being calibrated based on the real measurement. 4) WFNN(1) in CEM category uses local weather file to generate a sufficient number of training patterns using the simulator. it can calibrate simulation output by predicting the possible deviation of the simulator’s output from the measurement. 5) WFNN(2) in CEM category is another type of neural network addressing calibration issue.
7) BRIDGENN (Bridge Neural Network) in CEM category is to relate other neural network’s indoor illuminance prediction to the predictor outputs such as reference parameters (R13-17) and the performance indicators (P3. The neural networks take solar azimuth and solar altitude as a part of input parameters instead of the time stamp. Table 3. and P6) other than the average illuminance and uniformity. P5.6: Input and output parameters for the neural networks designed in HISSTO ID E1 E2 E3 E4 E5 S1 R1-12 R13 R14 R15 R16 R17 P1 P2 P3 P4 P5 P6 P7 Variable Unit degrees degrees Solar azimuth Solar altitude 2 2 Horizontal diffuse irradiance Horizontal global irradiance Horizontal global illuminance Light redirection louver angle Indoor illuminance value for the nth reference point Background luminance for DGI calculation Background luminance for GCRT calculation Illuminance on the eye ball due to the direct light component Illuminance on the eye ball due to the indirect light component Value of the maximum luminance patch on the CRT screen Average illuminance across all indoor reference points Uniformity factor in target indoor space Daylight Glare Index CIE Glare Index due to the electrical light Glare on the computer screen Total solar radiation through the glazing area Total power consumption by the electrical light Description φ θ I hd I hg Ehg LRL En L b1 L b2 Ed Ei L max Em Uf DGI CGI G CRT Q P W⁄m W⁄m lx degrees lx lm lm lx lx lm lx - W W 69 . the neural network is trained with the training patterns that the local weather file-based simulation generates. This is for reducing the number of input variables while maintaining the minimum meaningful set of data to train neural networks.7 show the input and output parameters for each of those neural networks designed in HISSTO.6 and Table 3. To do this. Table 3. then predicts the rest of other outputs needed in HISSTO. This neural network gets average illuminance and uniformity value (measured or predicted by another neural network) as a part of its inputs.
But. The character C after that denotes the calibration routine with the number of illuminance sensors used for the calibration. a subset of neural networks designed in HISSTO have the outputs predictable only through being trained with the data set generated by the simulator. Separate from the naming convention. Table 3. SM: SiMulation. a special naming convention is established. WF: Weather File).In principle. Figure 70 . Therefore. The last two letters NN indicates that the final form of the hybrid predictor is a neural network. all input and output parameters of a neural network are supposed to be measured. (MS: MeaSurement.7: Selected input(I) and output(O) parameters for each neural network designed for HISSTO MSNN(1) E1 E2 E3 E4 E5 S1 R1 R2 R3 R4 R5 R6 R7 R8 R9 R10 R11 R12 R13 R14 R15 R16 R17 P1 P2 P3 P4 P5 P6 P7 I I I I I I O O O O O O O O O O O O SMNN I I I I I I O O O O O O O O O O O O O O O O O O O O O O WFNN(1) I I I I I I O O O O O O O O O O O O O O O O O O O O O O SMCNN I I I I I I O O O O O O O O O O O O MSNN(2) I I I I I I O WFNN(2) I I I I I I O O O O O O O O O O O O BRIDGENN I I I I I I O O O O O I I O O O - To represent each of the different neural network designs. The First two capital letters indicate the source of the neural network’s training patterns. the validity of those neural networks can only be judged relative to the simulator’s predictions instead of being directly compared with the measurement.
or to re-calculate all outputs after being calibrated by other system component such as the calibration illuminance sensor value (i.e. MSNN(1)).29 shows the iconic representation of the neural networks designed for HISSTO. CEM M/S/W Number of Sensors for training Single/Multiple ?1 2 2 B 1 1. Source of Knowledge M: Measurement S: Simulator W: Weather file Neural Network Type 1. WFNN(2)). and WFNN(1) .can perform necessary predictions without being combined with any other system component or neural network. the combinations of the component neural networks.e. MSNN(2). or the other neural network predicting the simulation error (i.8 shows the design characteristics of HISSTO’s seven component neural networks serving as the major building blocks of the hybrid predictors.e.e. Others need BRIDGENN to be functional as the predictor due to the incompetence of reaching those output variables only accessible by the simulator (i.i. and the illuminance sensors provide intended variations in the hybrid predictors design and implementation. SI 2. Necessity of BRIDGENN Sensors in operation P S Simulation Absence of an element Prediction Figure 3. Some neural networks in HISSTO . Designing multiple component neural networks and the different ways of hybridization to come up with various hybrid predictors are considered to be appropriate especially when each hybridization strategy has potential benefits and disadvantages depending on the context and available resources. the simulator.3.29: Iconic representation scheme of the neural networks designed for HISSTO Table 3. SMCNN). BRIDGENN 2. 71 . In HISSTO. SMNN.
Table 3.8: Design characteristics of the neural networks to construct HISSTO’s predictors ID MSNN(1) Icon
Inputs E1-5, S1
-NN trained with measurement data having multiple indoor illuminance values -This NN needs the BRIDGENN to predict unmeasured performance indicators. -NN trained with measurement data having a single indoor illuminance value -This NN Predicts only a point illuminance value and doesn’t need the BRIDGENN. -NN trained with the measured sky condition-based simulations -This NN predicts all performance indicators including the unmeasured ones. -NN trained with the data generated by the local weather file-based simulations -This NN predicts all performance indicators including the unmeasured ones. -NN trained with the data generated by the local weather file-based simulations -This NN only predicts only indoor illuminance profile and needs the BRIDGENN to complete the prediction. -NN trained with the data generated by the local weather file-based simulations having average indoor illuminance and uniformity as a part of its inputs. -This NN predicts the unmeasured performance indicators. -NN trained with the simulator error patterns -This NN needs simulations before calibration and the BRIDGENN to complete the prediction. -This NN needs indoor illuminance measurement to maintain its calibration capacity.
R1 or,... R12
R1-17, P1-P3, P5-P6 R1-17, P1-P3, P5-P6 R1-12
E1-5, S1, P1-2
R13-17, P3, P5 P6
Figure 3.30 shows the generic structure of the neural network design in HISSTO which has only one hidden layer in addition to input and output layer. Depending on the type of neural network designed in HISSTO, each neural network keeps only a subset of input and output nodes illustrated in the diagram. For example, BRIDGENN takes average illuminance (Em) and uniformity (Uf )as the additional inputs instead of having them as outputs, which requires two more nodes in the hidden layer.
E1 E2 E3 E4 E5 E6 E7
E8 E9 E10 E11 E12 Lb1 Lb2 Ed Ei Lmax Em Uf DGI GCRT Q
Figure 3.30: Generic architecture of the neural networks in HISSTO
Ihd Ihg Ehg LRL Em Uf
Neural network training parameter set-up
Each neural network can have an unique set of training parameters to get the best result. Normally, learning rate and momentum are the most sensitive training parameters to be optimized. With multiple trials, it has been identified that learning rate 0.0001 and momentum 0.95 work for the most of neural networks designed for HISSTO. The termination point of a neural network training process can be determined by specifying either the maximum allowable prediction error or the maximum number of training steps the neural network adjusts the weights of its links. Maximum number of training steps also has implication for the speed of the training: around 2,000 steps yields fairly stable result across most of the neural networks tested.
Neural network training
Training of a neural network in HISSTO is done with the training patterns either from the simulations or the measurement. Partitioning of the data set for training and testing is important especially for the proper validation of the neural network’s prediction capability. In HISSTO, basically three different data partitioning approaches have been tested (Figure 3.31). Generally speaking, the train and test data set of a neural network need to be adequately apart in time dimension to check how far the trained neural network’s prediction capability can be generalized. One way of doing this is the “shuffle-and-select” method in which a train or a test data set is randomly chosen among multiple segments of a data pool which are shuffled every time a new selection is to be made. This is a good way of testing neural network’s prediction performance when the training patterns are homogeneously distributed throughout the entire data pool. Nevertheless, this data set partitioning scheme is less meaningful in HISSTO’s actual system control operation context not to mention the embedded technical complexity of implementing this method.
The second approach is the “cumulative” data set partitioning method. The entire cumulated data set up to the current control time interval is used for training a neural network to predict control outputs for the next control time interval. As time passes, the data set for training is growing in its size, which is a critical shortcoming of this approach. Without giving certain incentive for the most recent part of the data set, the trained neural network’s up-to-date prediction capability possibly fades by being enforced to fit weights across the entire training patterns distributed throughout the cumulated data sets. Even though it is not a scalable solution considering the magnitude of data the control system eventually gets during its continuous operation, this scheme can be useful when the data collection is necessary only for a certain period of time.
8 for the best training result. In this method.e. this scheme can keep the neural networks continuously being updated. One of the important issues with this training scheme is to determine the size of the time window to maximize the neural network’s learning capability while minimizing the computational load without losing a grip on the environmental changes. First.2 and 0. Since daylight responsive lighting control requires at least quasi-real-time states of the environmental variables (i. preferably between 0. This "sliding time window" approach is scalable because it always uses a constant volume of data to train neural networks on a regular basis. This requirement is a challenge for the implementation of neural networks designed in HISSTO.31: Different neural network training schemes based on the progression of time Neural network training pattern encoding & output decoding A neural network requires all the values in a training pattern between 0 and 1. all the inputs of the neural network should first be encoded and its outputs must be decoded. The challenge comes from two different sources. Shuffle and select Training Prediction Cumulative Sliding time window Figure 3. sky condition). the input and output values for these neural networks have all different boundary of variations some of which are too widely dispersed to make them fit the 75 . Experiments show that a 4-5 days of time window guarantees a reasonable performance of the neural networks designed for HISSTO. This is because the sigmoid function that a neural network uses to modify node input value for better fitting non-linearity performs well within this value range. training is done with only the most recent data set to predict the control outcome of the next control time interval.The third approach is “sliding time window” method. For this reason.
On the other hand. Logarithmic scale can compress a large number into a small one. Both scale and offset are normally used for encoding training patterns and decoding prediction outputs. a = ( α – β ) ⋅ ( max ( t i ) – min ( t i ) ) and b = ( 2α – β ) – a ( 2max ( t i ) – min ( t i ) ) Where ti ′ is the encoded value for the ith variable in T.3) Where pi is the decoded value of p′i .8 and t i ∈ T . Especially for scaling. any element variable of T or P can be encoded/decoded by applying both scale ( a ) and offset ( b ) based on the following process: a ⋅ max ( t i ) + b = α for .2 < β < α .4) The same transformation could be done with logarithmic scale. a special numeric transformation process has been devised.required value range. pi ′ – b p i = -------------a (3. either arithmetic or logarithmic scale can be applied depending on the characteristics of a specific variable. the ith variable in the output P. or t i = -------------. certain variables are changing too dynamically (i. where t ( sm ) i is the simulated value and t ( ms ) i is the corresponding measurement of the ith variable in T. ti ′ = a ⋅ t i + b (3.2 ≤ α ≤ 0. Preliminary test shows that even combination of these two different scaling methods would work properly. a ⋅ log min ( t i ) + b = β for 0. tightly bounded variables such as uniformity and glare indices can be processed with arithmetic scale. and P is a prediction output of a neural network.8 and t i ∈ T a ⋅ min ( t i ) + b = β for 0. environment variables and reference parameters which depend on the sky condition at a particular moment).2 ≤ α ≤ 0. making the neural network comparatively sensitive to the value changes in the lower magnitude with any widely ranged variable. To ensure every encoded variable in a training pattern is within the desired value range. Secondly. t ( sm ) i a ⋅ log max ( ti ) + b = α for 0. Sky condition variables such as outdoor illuminance values are suitable for this type of manipulation. Suppose T is a training pattern. which yields. which yields.especially for the simulation calit ( ms ) i bration network (SMCNN). max ( t i ) a = ( α – β ) ⋅ log ------------------ min ( t i ) 76 .2 < β < α .e.
Thrun 1991). the ith variable in the output P. The predictor’s role in HISSTO is to provide means for directed exploration throughout the building systems control state space. ti ′ = a log ( ti ) + b (3.either simulation or measurement (Chang and Mahdavi 2001). There are three groups of hybrid predictors designed in HISSTO.2 and 0. 3. simply setting a fixed boundary for each variable in a training pattern based on its probable range of variations can be effective in preventing this type of problem.b = 2α – β – a ( 2 log max ( t i ) – log min ( t i ) ) Where ti ′ is the encoded value for the ith variable in T. Three hybrid predictors use measurement based neu- 77 . One of the problems for applying this method comes when the range of a value in all training patterns can not properly accommodate the current or future variations of its value the neural network uses in operation. every variable in a training pattern T can be converted into the number between 0.1. They vary mainly by their inputs and outputs as well as the source of knowledge for training .7) By using this encoding/decoding scheme. The internal process and the characteristics of a hybrid predictor varies depending on the component neural networks the hybrid predictor uses. a (3. decoded output element Pi is p ( sm ) i p ( ms ) i = --------------------p′–b 10 i -------------.3 Object design of the hybrid predictors Six different hybrid predictors have been designed and implemented for HISSTO. a p i = 10 (3. pi ′ – b -------------. the directed exploration utilizes certain exploration-specific knowledge for guiding its process (cp. Unlike undirected exploration where actions are generated randomly with uniform probability distribution. Sometimes.5.6) When the neural network is SMCNN.5) Where pi is the decoded value of p′i .8 and the neural network output is decoded back properly after the prediction is made. One way to solve this problem is to use squeezed α and β values so that the encoded inputs for prediction could be kept within the desired ranges.
it has problems in predicting those reference parameters necessary for calculating glare indices as well as the other performance indicators which cannot be measured. The other two hybrid predictors are based on the model copy neural networks. For this reason. therefore. crucial for this predictor. 4) Predictor using weather file based simulation copy neural network: (WFC0NN) This hybrid predictor uses WFNN(1) trained with the patterns generated by the simulator based on the 78 . both average illuminance and uniformity are calculated and fed to the BRIDGENN as a part of its input. The six hybrid predictors designed in HISSTO are: 1) Predictor using multi-sensor measurement-based neural network: (MSC0NN): This hybrid predictor is using MSNN(1) trained with measurement data only. With just one indoor illuminance reference point. using multiple sensors for the component neural network training can only be allowed during a limited period of time. 3) Predictor using simulation copy neural network: (SMC0NN) This hybrid predictor uses SMNN trained solely with the training patterns derived from the simulation inputs and outputs. Since MSNN(1) solely relies on the measurement. this hybrid predictor’s performance depends heavily on the similarity of the weather conditions prior to the time performing actual prediction. The decay pattern of its prediction capability as time passes then becomes interesting issue for observation. Selecting an adequate reference point is. Another issue for this predictor is the number of indoor illuminance sensors. The BRIDGENN then predicts all other reference parameters as well as the other performance indicators. Therefore. How closely the component neural network can copy the simulator is another issue. some of the performance indicators selected for evaluating control options can be meaningless. feeding only one sensor value to this predictor for training is a sustainable operation scenario in reality even though its coverage of control performance criteria is significantly limited. Having ever-present multiple sensors within a control zone is not realistic. Since the simulations for training pattern generation is based on the measured sky conditions of the past. uniformity factor can not be calculated with a single illuminance value. Unlike MSC0NN’s case. For example. The remaining one hybrid predictors uses a neural network built for the simulator calibration. This predictor doesn’t need the BRIDGENN because the source of knowledge for this predictor is the simulator and each training pattern includes all output variables to be predicted. once the MSNN(1) predicts the indoor illuminance distribution. 2) Predictor using single sensor measurement-based neural network: (MS1NN) This hybrid predictor uses MSNN(2) neural network trained with only one indoor illuminance sensor value.ral networks.
The neural network component of this predictor doesn’t need any measurement data as long as the local weather file is available.local weather file instead of relying on the simulations with the measured sky conditions during the past 4-5 days of time period. 1: getZoneDescription() LightingZone: View LightingSimulator: Simulator 2: getSkyCondition() LightSensor: Sensor 3: getSimulatorOutput PredictionErrorIdentifier: rainingPatt ernBuilder : NNTrainingPattern 4: getMeasurement 5: generateTrainingPattern 6: getTrainingPattern 7: performTraining NeuralNetworkTrainer: MachineLearnerTrainer SMC12NN: MachineLearner Figure 3. SMCNN. Then the simulation output (i. daylight simulation must be done first. should be retrained for a successful adaptation to the changed condition. 5) Predictor using neural network calibrated simulation: (SMC12NN) The simulation result can be calibrated on-line to capture the system behavior more precisely.e. indoor illuminance profile) is calibrated by SMCNN.32 and Figure 3. In the beginning of each month. the simulator takes current month’s weather file data and generates hourly-based predictions over the reference parameters and the defined performance indicators for each louver control option for the entire month. its prediction capability should be almost as good as or even better than the one trained by the simulator with measured past sky conditions (SMC0NN).33 describe SM12CNN’s training and prediction process. To operate this hybrid predictor. Figure 3. Whenever a change occurs. this predictor’s decay pattern over time is important to consider. the component neural network for simulation calibration. Since the component neural network also relies on the simulator for training. SMCNN should be retrained on-line.32: Collaboration diagram describing SM12CNN’s training process: UML notation 79 . Whenever a change occurs. Since SMCNN uses multiple indoor illuminance sensor values for training. average indoor illuminance and uniformity are calculated and the BRIDGENN completes the rest of the prediction task. This approach assumes that the simulation prediction error has a certain pattern so that a machine learner can capture and compensate for it as long as the building configuration and systems configurations are remain unchanged. This hybrid predictor uses SMCNN. a neural network to calibrate simulation output by learning the simulator’s prediction error.
it uses BRIDGENN to generate the final prediction output (Figure 3. In this calibration process.LightingZoneView: View 1: getZoneDescription() 2: getSkyCondition LightingSimulator: i mulat or 3: generateOutput() 6: getSkyCondition() imulat orPredic tion: Prediction LightSensor: Sensor Database 4: getTrainedNNWeights() NeuralNetworkBuilder: MachineLearnerBuilder 5: createNeuralNetwork() SMC12NN: Machi neLearner 7: generateOutput() 9: getDeviat ion() 8: getPrediction() PredictionCalibrator: Calibrator 10: calibratePrediction() FinalPrediction: Prediction SimulatorErrorPrediction: Prediction Figure 3. Figure 3. feeding only one indoor illuminance sensor value to this network for real-time calibration makes this predictor practical in actual building operation. Identifying a proper reference point is.34 shows an exemplary collaborative process between WFNN(2) and BRIDGENN to perform necessary prediction with WFC1NN hybrid predictor in HISSTO. the difference between the measured illuminance value and the predicted one is applied uniformly to the other reference points. After the calibration. 80 . It then calibrates the prediction result using only one real-time indoor illuminance sensor value. crucial for this hybridization scheme. Similar to MS1NN’s case. a neural network trained with weather file-based hourly simulations to predict the indoor illuminance profile. therefore.36).33: Collaboration diagram describing SM12CNN’s prediction process: UML notation 6) Predictor using calibrated weather file based simulation copy neural network: (WFC1NN) This hybrid predictor uses WFNN(2).
R1 (E1) R2 (E2) ....R17 DGI GCRT Q Em Uf BRIDGENN Predict Default Inputs Em Uf Figure 3..9 and Figure 3....35 illustrate how each designed hybrid predictor is trained and performs prediction in relation to the other system elements......34: Exemplary collaborative process between WFNN(2) and BRIDGENN in WFC1NN Table 3.. 81 . Calibrate Default Inputs Predict WFNN(2) Sensor R12(E12) R13 ...
Predict indoor illuminance profile using the Simulator 2.9: Diagrammatic description of each hybrid predictor’s prediction process Predictor MSC0NN M ? Prediction process Description 1. Calibrate the output with one illuminance sensor value 3. Calibrate simulation output with SMCNN 3. Predict remaining preference indicators using BRIDGENN WFC1NN W W B P 82 . Predict all performance indicators using SMNN WFC0NN W P 1.Table 3. Predict indoor illuminance profile using MSNN(1) 2. Predict remaining performance indicators using BRIDGENN 1. Predict remaining performance indicators using BRIDGENN 1. Predict all performance indicators using WFNN(1) SMC12NN S SM W B P 1. Predict a single illuminance value on the chosen reference point using MSNN(2) W B P MS1NN M ? SMC0NN S P P 1. Predict indoor illuminance profile using WFNN(2) 2.
indoor illuminance solar azimuth solar altitude sky condition louver d1 d6 input c1 training calibration output training pattern t1 deviation neural network d4 s2 d2 d3 s1 s3 n2 n1 b1 prediction output simulator d5 bridge neural network time building geometry local weather file MSC0NN: d1 t1 WFC0NN: d2 WFC1NN: d2 s1 s1 d4 n2 d5 b1 MS 1 NN: d1 t1 d4 n1 n2 SMC12NN: d3 s2-d1 d6 c1 d5 b1 n1 SMC0NN: d3 s1 s3 d4 n2 t1 d5 d4 b1 n1 t1 d4 t1 d4 t1 d3 Figure 3. After this. Furthermore. The Tester in HISSTO realizes this process with an evaluation algorithm which filters all candidate target systems states through a predefined set of visual environment assessment criteria. an additional evaluation process to select desirable louver angle needs to be performed. Because the hybrid predictors are only for the daylight prediction task. the electrical lighting calculation is done separately by using a pre-calculated Luminaire Matrix.2 Object design of the Tester Once the daylight prediction is completed by any predictor in HISSTO.5. these performance indica- 83 . selection of the most desirable luminaire dimming level combination with the chosen louver angle should be performed. 3.2.35: Training and operation processes of the different hybrid predictors in HISSTO 3.5.1 Visual performance evaluation criteria An attractive feature of a model-based control strategy is the diversity of the performance indicators that can be incorporated into the control option evaluation process.
glare due to electrical light (CGI.8 (3.5 L b + 0. This can be formularized in the following expression: n 1 E m = -. glare due to daylight (DGI. Hopkinson 1971). This can be expressed in the following formula.e. the most widely accepted one is the Daylight Glare Index (DGI.tors need not to be strictly limited to strictly visual criteria (i. illuminance level).⋅ n ∑ Ei i=1 (3. the simulator predicts the values of the seven performance indicators: average illuminance calculated from the illuminance values on the multiple reference points in a space. solar gain (Q).07 ⋅ ω ⋅ L s 0.478 ∑ ---------------------------------------------0. cp Einhorn 1979). glare on the computer screen (GCRT). uniformity of illuminance distribution on the specified plane in a space. Ls ⋅ Ω DGI = 10 log 0.4) Where Ω Ls Lb ω solid angle subtended by the glare source (window) modified by the position factor luminance of the glare source in the field of view luminance of the background total solid angle subtended by the window at the eye 84 .3) where ESD is the standard deviation and Em is the average value of all illuminance values. This can be expressed in the following formula which yields the uniformity value between 0 and 1 (UE. Among many daylight glare models. heat gain and energy use). Hopkinson 1971). cp. DGI: Daylight glare can be a critical source of visual discomfort since it is usually in the direct field of view having high luminance. Glare index assessing discomfort glare due to the daylight specifically addresses this condition. Mahdavi and Pal 1999): Em U E = ---------------------E m + ESD (3. Uniformity: The uniformity factor indicates the evenness of the illuminance distribution within a space.e. In HISSTO. Average illuminance: Mean of the indoor illuminance values for all reference points to represent the light level in a target space. and electrical lighting power consumption (P).2) where n is the number of all reference points. cp. but also can include cross-domain performance criteria (i.
The formal expression of this glare is: Ed 2 1 + -------500 -----------------.5d + 4.CGI: The glare caused by an electrical lighting source.12 ( 1 – e ) 2 d + 1. Both daylight and electrical lighting source should be considered.⋅ L s ω . This glare can be expressed in the following formula: n 1 G CRT = ------------------.011 d 3 horizontal distance to the glare source parallel to the line of sight horizontal distance to the glare source perpendicular to the line of sight height of the glare source above the eye position GCRT: Glare on the computer screen depends on the locations and the relative angle of the viewer and the computer monitor.6) where n is the number of interior surface patches. Einhorn 1979 and accepted by the CIE Committee of Discomfort Glare as the standard glare index to be used for discomfort glare due to the electric lighting.18 d 2 2 x 2 where x = – --------.6 0. Ai is the area of the ith surface patch. β i is the angle between the normal of eye patch and the line connecting the centers of surface patch and eye patch.⋅ π ⋅ L max i=1 ∑ ------------------------------------2 r i L i ⋅ cos α i cos β i (3.s D S d = --. This CIE Glare Index was developed by H. L i is the luminance of the ith surface patch.5) where Ed Ei Ls ω direct illuminance on the vertical surface of the eye indirect illuminance on the vertical surface of the eye luminance of the glare source in the field of view total solid angle subtended by the window at the eye –2 Here.. s = --H H D S H 0. r i is the distance between the 85 . It is given by: p –2 d e x = ----------------------------------. α i is the angle between the normal of the ith surface patch and the line connecting the centers of the ith surface patch and eye patch.s + -----------. the solid angle ω is modified by p which is a position factor to take into account the occupant’s line of sight relative to the glare source.∑ --------CGI = 8 log 2 2 Ed + Ei p (3. D.+ 0.
This can be roughly estimated based on the following formula: n 1 Q sol = -. b: Uniformity. Power consumption: The amount of energy consumed by the luminaires in a target space.8) where n is the number of luminaires in the target space and P i is the power consumption of the ith luminaire. E i is the illuminance on the ith window patch. and η is average luminous efficacy. f: solar gain in winter g: solar gain in summer.ith surface patch and eye patch. User's preference for the desired values of those selected visual performance indicators may be expressed by graphic means.7) where n is the number of window patches.e. e: GCRT. h: power consumption). A i is the area of the ith window patch. L max is the maximum luminance value identified among CRT screen patches.⋅ η ∑ Ei ⋅ Ai ⋅ τi i=1 (3. τ i is the solar transmittance of the ith window patch. Illustrative examples of preference functions for the performance indicators are given in Figure 3. 86 . Operation of daylight systems (i. c: DGI. d: CGI.36 (a: Average illuminance. The power consumption by electrical lighting can be represented with the following formula: n P= ∑ Pi i=1 (3. Solar gain: The amount of solar radiation introduced to a space through the glazing area. louver or blind) has also energy implication since it can increase/decrease heating/ cooling load. This concern is left out in calculating power consumption because it requires a complex tracking of all related HVAC components and is already taken into account by the solar gain performance indicator.
4 2.8 2.8 g.4 0. 87 .6 0.1 2. e c.8 c.2 1.3 0. An example of such a utility function U is given in the equation 3.0 Figure 3.7 0. 0 2 4 6 8 10 12 14 16 18 20 0 0.1 0.0 0. d.2 0 a 0 lx 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 b 0 0.9 below.0 c. 0 3 6 9 12 15 18 21 24 27 30 d.5 1. g.36: Exemplary preference functions for HISSTO’s performance indicators Those preference functions provide the basis for the derivation of objective function toward the evaluation of control options (Mahdavi 1993).4 0. h c. h 0. d.9 1. e Preference index 0.0 1.1.6 0.7 3. An objective function may be based on a single performance indicator.2 0.9 1.6 0.8 0. Preference index 0.4 0.5 0.2 0 f f. or on a weight aggregate of two or more performance indicators.6 d a a b b 0.3 0.
For HISSTO.3 Object design of the Commander After the best control option is identified.9) ∑W = 1. 3. WE: Weight for average illuminance WUE: Weight for uniformity factor WDGI: Weight for Daylight Glare Index WCGI: Weight for CIE Glare Index WGCRT: Weight for Glare on CRT WQ: Weight for solar gain WP: Weight for luminaire power consumption PE: Performance indicator for average illuminance PUE: Performance indicator for uniformity factor PDGI: Performance indicator for Daylight Glare Index PCGI: Performance indicator for CIE Glare Index PGCRT: Performance indicator for Glare on CRT PQ: Performance indicator for solar gain PP: Performance indicator for luminaire power consumption Needless to say. Another part of the Commander object’s services is to ensure a seamless execution of the control commands through the software/hardware interfaces connecting the actuators to HISSTO. such weightings involve subjective and contextual considerations and may not be standardized. preference functions and the weighting mechanism are intended to provide the user of the system with an explorative environment for the study of the relative implications of the impact of various performance indicators in view of preferable control strategies (Mahdivi et al. where U=WE x PE + WUE x PUE + WDGI x PDGI +WCGI x PCGI + WGCRT x PGCRT + WQ x PQ+ WP x PP 0 ≤ W. Rather. The luminaire control is 88 . The louver control is done through a serial connection to the louver motor control circuit board. the Commander object checks other on-going control actions to avoid any potential global conflict before the execution of the control decision. 1999).5. P ≤ 1 and (3. both light redirection louver and luminaire actuation interfaces use digital signals generated by the DeviceControlCommand. This is done by predefined heuristic rules implemented as “if-then” clauses.Maximize U.
10 shows the list of primary sensor types and their characteristics used in HISSTO. buffer size. the captured data may need to be further processed to yield desirable data format. the sensors used in HISSTO should have generic representation to maintain vendor-independency and interoperability.implemented through an A/D converter embedded on a PC control board passing a converted analog signal to each luminaire ballast for dimming. This is important especially for the predictors to provide the required services. This is absolutely necessary to avoid the situation where an instantaneously retrieved sen- 89 . Both illuminance and irradiance sensors are the key elements in visual environment control in addition to the systems state sensors such as the luminaire power level sensor and the louver angle sensor. This type of elaborated control command execution process is specified in one of the member functions of the DeviceControlCommand object. For detail information.1. and reading rate for an optimized data acquisition.10: Various sensors used in HISSTO Type Photometric sensor Pyranometer Luminaire power level sensor Louver position sensor Description measures illuminance measures irradiance reports luminaire power level reports louver angle Sampling rate 1/min 1/min 1/state change 1/state change On the server side. For example. and logging of the sensor values. the ControlCommand will be executed by gradually increasing (or decreasing) the luminaire power level up (or down) to the desired state instead of changing it at once. the calculated average of all scanned data instances can represent the sensor value for that specific time interval. After all active sensor channels are properly scanned and read. if one sensor channel is scanned during several minutes. Considering possible future extension.1.1 and A. see Appendix sections A. For example. Based on the Command design pattern. when there is significant difference between the current luminaire dimming level and the newly suggested one.4 Object design of the Sensor The sensors connected to HISSTO capture necessary physical properties of the environmental or the system objects based on the proper sampling rates. Table 3.2. The data acquisition period can be customized and the data collection cycle can be activated every 10 minutes between 9 AM and 6 PM. 3.5. the system maintains connection to all sensors for monitoring purpose. The data acquisition module performs sampling. Table 3. The facility manager can specify scan rate. a ControlCommand can be further elaborated for better control quality. processing.
blind controller.6 Object design of the daylight responsive lighting control Instead of evaluating all possible instances of combined daylight and electrical lighting control options. Luminaires can also be either turned on or off during a data acquisition cycle. In HISSTO. Numeric display is added for each sensor to provide more precise reference. Different color can be used to identify each sensor in the graph. or the other internal or external daylight modification devices. some are derived only from electrical light (i. Multiple indoor illuminance sensor values along with global horizontal illuminance. An Actuator object in HISSTO can provide interfaces to dimmable luminaire ballasts. Depending on what type of systems the target control zone has. Modeling of an Actuator should allow vendor-independent and inter-operable interfacing so that it can accommodate a variety of manufacturers and products without significant code change. corresponding actuators should be identified and modeled properly. For this purpose. Certain performance indicators address only daylight (i. global horizontal irradiance. This is basically for the facility manager rather than ordinary users.e. which is 90 . computations of glare indices require the knowledge of background luminance. numeric or graphic indicators confirm the outcome of this control action. HISSTO assesses each of those separately and combines the results later to avoid an explosion of the control state search space. Training of a neural network requires a systematic data acquisition plan.1. HISSTO’s monitoring component has the server side interface for displaying the captured sensor values on the screen in both numeric or graphic format. power consumption).5 Object design of the Actuator The Actuator represents a mechanical/electrical device responsible for controlling a system. A.5. Once the collected data set is properly processed. For example. 3. and others need to consider both. heat gain).5.2 section in the Appendix A describes the data acquisition module interface. scheduled or predictive control of the louver and the luminaires.e.4. it can be either displayed or archived with an attached time stamp. the increment of the louver angle changes within a data acquisition cycle as well as the duration of the data collection after the louver is positioned at a certain angle before it moves to the next position. and diffuse horizontal irradiance value are among the ones to be displayed. Once a control action is completed. light redirection louver controller. The actuation in HISSTO is finalizing either manual.sor value does not represent the overall trend due to the fluctuation of its value over time. the facility manager can specify the louver status (either on or off). 3. the data acquisition interval is typically set to 10 minutes.
5) Final values of glare indices are generated by synthetically combining the glare components . 105 degrees).5.37 shows how the daylight prediction and evaluation for the louver angle selection can be integrated with the outcome of electrical light calculation based on the Luminaire Matrix look-up process. The following steps show a typical daylight responsive lighting control process designed for HISSTO: 1) Outdoor daylight sensor values.both daylight components and electrical light components . A pre-calculated Luminaire Matrix for electrical lighting assessment is proven to be effective. The louver position and the preferable luminaire power levels are identified by selecting the one which maximizes the utility value from all tested louver-luminaires control options. The variables used in the diagram are depicted in Table 3. direct and indirect illuminance on the eye surface for CGI calculation and the luminance on the computer screen for GCRT calculation. and current time are identified. Figure 3. 91 . current louver position. Examples of those glare components are background luminance.affected by both daylight and electrical light. luminance of each window patch for DGI calculation.calculated in step 3 and 4 for each louver-luminaires set. each of all nL such combinations are searched out to identify the corresponding illuminance contribution and power consumption along with the glare components (electrical light components) for CGI and GCRT calculations. current luminaire power levels. 4) For each luminaire. Then.9) is selected (i. 3) Another round of simulation for the chosen louver position is performed to generate glare components enabling glare indices calculations when the selected louver position is combined with various sets of luminaire power level combinations. Calculated performance indices for each louver position are further processed to generate the utility value U based on the defined preference indices and corresponding weights. from the Luminaire Matrix. 2) Either a simulation or a hybrid predictor-based prediction is done for each of the eight louver positions based on the input data. Finally. analog signals are sent to the louver controller and the luminaire ballasts to update the control states of those devices. This is possible since the pre-calculated glare components are additive in generating the final glare indices. then the louver position that maximizes the utility (see equation 3. n steps of candidate power levels (current power level plus (n-1)/2 steps below and (n-1)/2 steps above the current power level) are identified.e.
37: Activity diagram of HISSTO’s daylight responsive lighting control process: UML notation 92 .Simulate Daylight wit h a L ouver An gl e [else] [ a ll l ou ve r a ng le s are cov ere d ] Select Best Louver Angle Simulate Daylight with Chosen Louver Angle :DaylightComponent Se lect a Dimm ing Scheme Select Dimming Level for a Luminaire Lookup Luminaire Ma tri x Identify Ill uminan ce Identify LbDL Identify LbCRT Identify Ed Identify Ei Identify Lm ax Add Ill uminan ce Add LbDL Add LbCRT Add Ed Add Ei Add Lm ax [ el se] [ all luminaire dimming levels are covered ] Get Daylight Components Add-up Illuminance Ca lcu la te DGI Add-up LbCRT Add-up Ed Add-up Ei Add-up Lmax Evalu ate Co nt rol Op tio ns [else] :Contro lOpt ionEvaluati onTa ble [ all luminaire dimming schemes are covered ] Se lect Be st Control Option Figure 3.
At every control interval. The ManualControl mode is maintained until the user explicitly terminates the mode unless a special event enforces it to be finished.1 Object design of the ManualControl The ManualControl in HISSTO overrides other control modes and doesn’t require any prediction or evaluation process to come up with a desired control option. Negotiations in the ManualControl can be implemented through two negotiation algorithms. The other negotiation scheme is to take mean value of all requested control state values for each target system. To resolve potential conflicts among different users in the same control zone. Voting is one way where the control action is selected based on the votes from all users.3. the louver angle could be specified explicitly based on the outdoor temperature and the time of day as is illustrated in Table 3.6.11. the system needs to continuously monitor and log the changing values of the performance indicators for future reference. This is also important for the training of a machine learner designed to capture user control preference. user prioritizing. This ScheduledControl mode can be overridden by any user interruption whenever it is requested.5. and negotiation. Simple chronological order method takes control command from each user and simply executes them in a chronological order. 93 . User prioritizing method implements control action based on the priority given to each user. In this control mode. the system calculates and archives the resulting values of the performance indicators mainly for the validation of the specified control schedule. Those two negotiation schemes are only for intra-zone control.5. so the inter-zone control conflict resolution schemes need to be further studied. Considering the case an user chooses to evaluate every manual control action he/she specifies. HISSTO simply checks current conditions and looks up this table to identify and send control command to the Commander object. the implementation of the ManualControl provides three alternative solutions: simple chronological order.2 Object design of the ScheduledControl Implementation of the ScheduledControl involves the specification of the “conditions-action” rules formatted into a table. For example. 3. Providing an user with useful information for an informed control decision is therefore critical.6.
at each time step. The electrical lighting settings are expressed in terms of the dimming level of the luminaires. This strategy requires.6. a reduction of the possible number of system states. The objective of this control strategy is to arrive at a configuration of daylight settings and electrical lighting settings that would accommodate the desired states for both daylight and electrical light systems. This is achieved.4) 94 . The daylight settings are expressed in terms of the position of the external light redirection louvers.Table 3. For example. due to the potentially unmanageable size of the resulting control alternatives. then.5. All the necessary output variables for electrical light part are.11: Louver operation schedule example Outdoor Temperature >= 76 F >= 76 F >= 76 F <= 70 F <= 70 F <= 70 F Time 8:00 AM-10:00 AM 10:00 AM-12:00 AM 12:00 AM-3:00 AM 8:00 AM-10:00 AM 10:00 AM-12:00 AM 12:00 AM-3:00 AM Louver angle (degrees) 45 30 0 105 105 90 3. Let L be the number of luminaires (or luminaire groups) and D the number of dimming states considered for each luminaire. each of the multiple luminaires in the target control zone can be at one of possible power level states. The total number of possible combinations n is then given by the following equation. separately calculated using a luminaire matrix before final integration. Moreover. by dimming the luminaires in groups. only a subset of possible dimming states are considered for each pair.3 Object design of the PredictiveControl Daylight-based dimming of the electrical lights via a predictive control strategy can be implemented in three different ways. n = DL (3. Predictive Control with the Simulator The first predictive control approach involves the concurrent assessment of various combinations of both daylight and electrical lighting control variables (Mahdavi et al. for example. only daylight prediction is done by the chosen predictor. 1999). When any of the hybrid predictors is used for control.
a larger number of dimming options may be considered and evaluated towards the selection of a preferred combined daylight and electrical lighting settings. As a result. change of the louver position and/or change of the dimming level of a luminaire) is generally sufficient to allow for the simulator for dealing with a larger number of louver angle options. as well as flexibility in the definition of the dynamic element (such as the position of observer. Each hybrid predictor goes through training or retraining process in case of system or building configuration changes before being used in the continuous control operation. The results are then combined with a preprocessed matrix of various luminaire dimming levels. Predictive control with hybrid Predictors The control approach using any of hybrid predictors is implemented through almost the same process as the “simulation plus luminaire matrix” control scenario. Predictive control with simulation plus Luminaire Matrix The second predictive control algorithm involves a sequential approach (Mahdavi et al. 95 . However. the limitation of possible dimming options at each time step to the immediate adjacent positions may result in the inability of the search process to transcend local minima and/or maxima. The advantage of this approach is the potential of extending the system control state space. the preferable louver position is derived by the daylight simulations. The concurrent assessment of daylight and electrical light options allow for the real-time incorporation of changes in room and aperture configuration. First. The typical time interval between two actuation events (i. or position of the luminaires. Instead of using simulator. if L = 2 and D = 3. 1999). This luminaire matrix needs only be re-generated if there is a change either in the configuration of interior space or in the number. This Luminaire Matrix can be computed ahead of the real-time control operation based on the assumption that the incident (electrically generated) light at any point in the space can be calculated by the addition of individual contributions of each luminaire. Combining the results of the selected louver setting with the matrix of electrical lighting states does not require real-time simulation and is thus highly efficient from the view point of computational time.e. The results are then combined with the electrical light evaluation through the Luminaire Matrix look-ups. etc. one of the hybrid predictors is used to predict the values of the daylight performance indicators for each louver control option.For example. type.). 9 possible electrical lighting control states will result.
1). The facade system includes a set of three parallel external moveable louvers which can be used for shading purpose and – to a certain degree – for light redirection purposes (Table A.Table A.1: Intelligent Workplace at Carnegie Mellon University as the system test bed 96 . 1997).4 System test & evaluation 4. Illuminance measurements have been performed intermittently in this test space during the entire year of 1998.4 . The western section of a south bay in IW is dedicated to lighting studies. For these measurements. the test space is illuminated using four indirect luminaires (Figure A. and visual quality.5). The glazing has low emission coating for heat loss control and 0.3 . Figure 4. The test bay features are illustrated through Figure 4.8 m above the floor (Table A. In terms of electrical lighting. About 60% of the external wall of the space consists of glazing. a recently established laboratory at the Carnegie Mellon University campus for demonstration and hands-on study of advanced building systems/technologies and their integration (Hartkopf et al.3 shading coefficient with high visible transmission. An array of 12 illuminance sensors is located along the central axis of this space at the height of about 0.1 Test bay & control target systems HISSTO’s system test was performed in the Intelligent Workplace (IW).3) the locations of which are depicted in Figure 4.3.4. Outdoor daylight conditions are monitored using a total of 11 illuminance and irradiance sensors that are installed on the daylight monitoring station on the roof of the IW. eight standard louver positions have been considered as per Figure 4.1 to Figure 4. This multi-layered enclosure is to reduce thermal loads to enhance air quality. This test area is isolated from the rest of IW using white-colored partitions.
Figure 4.3: Illustrative photos showing exterior and interior of the test bay 97 .2: HISSTO system test bay at the Intelligent Workplace Figure 4.
25m 12 11 10 9 8 7 6 5 4 3 2 1 L1 L2 N 1.1m P8: 105 ° P7: 90 ° P6: 75° P5: 60 ° P4: 45 ° P3: 30 ° P2: 15 ° P1: 0 ° Figure 4.8m L4 L3 4.6m L4 L3 . and indoor sensors 98 . luminaires.4: Plan and elevation view of the test bay with louver.3.
Figure 4.2).4. The angle of the light redirection louver installed exterior to the test bay was also retrieved through a serial port. The daylight station also measures horizontal and vertical irradiance/illuminance values of the sky for each orientation (Figure A. Figure 4. All 23 outdoor daylight sensors as well as 12 indoor illuminance sensors are connected to a couple of PC boards (Table A. The daylight station includes a “ diffuse ring” which measures diffuse horizontal irradiance by using an adjustable ring designed to block the direct solar insolation during daytime (Figure A.5).1). Throughout the entire data collection period.4 and Figure A.6) installed on the host computer so that the data acquisition process could be controlled through the software interfaces implemented as the VITMs (Virtual Instruments for both systems control and monitoring) in LabviewTM environment (Figure A.5: Daylight station on the top of the IW measuring sky condition parameters 99 . the diffuse ring needed to be adjusted every 3-4 days to be tuned for solar path change.2 Data acquisition for system test A special instrument having a collective daylight sensing devices named "daylight station" was constructed and installed on the roof top of IW to measure sky conditions.5 shows the daylight station installed on the roof of IW.
1 9 9 8 4500 Figure 4. 1 9 9 8 . Figure 4.7 demonstrate the daily and seasonal variations of the local sky condition. .6: Sky illuminance and irradiance measurements on a typical clear day (3/13/98) x 10 14 4 12 Global horiz ontal illum inanc e[lx] 10 8 6 4 2 0 0 500 1000 1500 2000 2500 3000 3500 4000 D a t a i n s t a n c e w i t h t i m e p r o g r e s s i o n : J u l .2.8).7: Sky illuminance measurement showing daily and seasonal variations 100 .4. Figure 4.6 in Appendix A) and the accuracy of the daylight simulations compared to the measurement (Figure A.7 in Appendix A) are also analyzed. x 10 10 9 8 7 6 global east west south north 4 1000 900 800 global 700 east west 600 W/ m2 500 north south diffuse x + + 5 4 3 2 1 0 6:00 AM 400 300 200 100 9:00 AM 12:00 PM Time 3:00 PM 6:00 PM 0 6:00 AM 9:00 AM 12:00 PM Time 3:00 PM 6:00 PM Figure 4.1 Analysis of the measurement Based on the collected measurement data (Table A.6. various analysis were performed to identify the characteristics of both the environmental and the system factors affecting the experiments. . Daylight availability in the test bay (Figure A.N o v.
8 shows an exemplary daily indoor illuminance profile for all reference points with cyclic louver angle changes. Considering the fact that the windows and the louvers are all facing west side of the test bay. During the initial louver angle increase operations. Both sensor-dependent (above) and louver-dependent (below) anaysis show that the impact of the louver angle changes is greater in the afternoon than in the morning.9 also illustrates the impact of the lover angle changes on the indoor illuminance distribution in the test bay measured at certain point of time. The indoor illuminance level actually decreases up to 30 . The impact of the cyclic louver angle changes is emerged as the wave-like concave curves across all sensor positions. each of these openings are gradually covered by the upper louver panel until the angle reaches at the certain degree where the gap between louver panels start continuously growing and accepting more daylight.40 degrees of the louver angle and continue to increase after that until the maximum louver angle is reached. this result can be easily explained.8: Indoor illuminance profile on all reference points influenced by the louver angle changes Figure 4. Each line represents different illuminance sensor position in the test bay. 101 .Figure 4. This phenomenon can be partially explained by the geometrical configuration of the three louver panel sets which has slight openings between the middle louver panel and the upper or lower louver panel. The other distinctive trend is the non-linear pattern of the indoor illuminance profile change when the louver angle is gradually increased. 7000 6000 5000 Illum inance[lx] 4000 3000 2000 1000 0 0 50 100 150 200 D a ta in s t a n c e w ith t im e p ro g re s s io n : 7 :0 0 A M -4 :0 0 P M 250 Figure 4. The overall daylight pattern during the chosen day shows gradual increase/decrease of the indoor illuminance level with the progress of time.
5000 4500 4000 3500 3000 2500 2000 1500 1000 500 0 S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 Sensor position P0 P30 P60 P90 5000 4500 4000 3500 3000 2500 2000 1500 1000 500 0 S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 S12 Sensor position P0 P30 P60 P90 Illuminance[lx] Indoor illuminance profile at 7:00AM (sensor dependent) 5000 4500 4000 3500 3000 Illuminance[lx] Illuminance[lx] 2500 2000 1500 1000 500 0 P0 P15 P30 P45 P60 P75 P90 P105 Louver position S1 S3 S5 S7 S9 S11 Illuminance[lx] Indoor illuminance profile at 1:00 PM (sensor dependent) 5000 4500 4000 3500 3000 2500 2000 1500 1000 500 0 S1 S3 S5 S7 S9 S11 P0 P15 P30 P45 P60 P75 P90 P105 Louver position Indoor illuminance profile at 7:00 AM (louver dependent) Indoor illuminance profile at 1:00 PM (louver dependent) Figure 4. The first group of data sets (D1-D3 in Table 4.2.1) came from the original measurement after being processed by applying a series of data filters for consis- 102 .2 Data preparation for testing the predictors Collected measurement data was filtered and further processed to be used for testing the developed system. There are two different data sets used for the tests.9: Indoor illuminance profiles depending on the sensor positions and louver angles (3/13/98) 4.
1) Data group Date Data set ID Instances 3/33/4 D1-1 453 3/53/13 D1-2 501 3/143/15 D1-3 445 3/163/21 D1-4 391 D 1 (3/3-3/31) 3/223/23 D1-5 477 3/233/25 D1-6 452 3/263/31 D1-7 334 4/14/4 D2-1 513 4/44/5 D2-2 435 D 2(4/1-4/18) 4/64/8 D2-3 468 4/94/13 D2-4 560 4/164/18 D2-5 610 D3(4/19-4/22) 4/194/20 D3-1 259 4/214/22 D3-2 382 Table 4. Table 4.1: Test data sets filtered by the validation criteria (see A.2) for the system test is obtained by filtering out the data segments of which the measured solar irradiance and illuminance values deviate more than ± 20 % from the predictions made by the Perez sky model (Perez et al. Table 4.3: Neural network training data sets generated by the weather file based simulations Data group Instances WD3 (3/1-3/31) 2232 WD4-1(4/1-4/30) 2088 WD4-2(4/19-4/30) 861 103 . These other data sets were obtained in a different building configuration where all windows except two tinted glass windows adjacent to the central door were blocked to reduce incoming daylight for the test purpose. These data sets are prepared mainly for the performance tests of the hybrid predictors such as WFC0NN and WFC1NN trained with weather file-based simulations.7 (Ulah 1996). The second data set group (D4-D6 in Table 4.2. 1993) used in the simulator. LUMINA (Pal and Mahdavi 1999). and WD4-2 were collected while all windows of the test bay were not blocked.1.9).2: Test data sets within ± 20 % deviation range from the Perez model sky prediction Data group Date Data set ID Instances 3/33/12 D4-1 215 D 4 (3/3-3/31 3/133/19 D4-2 221 3/213/31 D4-3 157 D 5(4/ 1-4/18) 4/14/4 D5-1 115 4/84/18 D5-2 173 D6(4/19) 4/19 D6-1 41 Table 4.3 shows another type of data sets generated by the simulator (LUMINA) based on the Pittsburgh weather file (Table A.tency checks described in Table A. D6. All data sets except D3.
Example of the actual inputs to LUMINA are described in Table B. Table 4.2: for outdoor irradiance measurement). global and diffuse horizontal irradiance as well as the vertical illuminance and irradiance for each orientation were measured. 1997. Empirical validation checks the simulation’s prediction capability based on the measurement.Table B.1: for outdoor illuminance measurement) and 6 pyranometers (Table A.Dec. a simulation program can be evaluated in three different ways . Table 4. 12 indoor photometric sensors are placed inside of the test bay at the Intelligent Workplace (IW).3. 1998) using 17 photometric sensors (Table A. After implementing the test. Comparative studies The test case for this approach should be kept simple to ensure intercode input equivalence. a continuous measurement was performed (Jan. and comparative study.analytical verification. Empirical validation Establish data acquisition instrumentation is for a test cell. any subsequent analytical or empirical study and can also be used as a final check against other models already validated. empirical validation. Outline necessary data capture to allow careful specification of model input and study model output.3 4. Finally.4 shows the characteristics of those three methods in validating a building energy simulation program (Judkoff et al.4: Example of typical simulation validation methods Method Analytical verification Validation process Define test cases where the analytical solution is given. For the purpose of validating LUMINA. Analytical verification aims at perfecting the simulation’s internal algorithm.4. . The validation process itself was performed by comparing the simulation result with the corresponding 104 .5. Perform predictions for each test case and compared the results with the corresponding analytical solutions. While the angle of the light redirection louver (LRL) was changed based on the incremental cyclic operation schedule.1 . This method can precede the other methods to prove the need for. the comparative studies use other simulators in the same domain to benchmark the performance of a simulation program by comparing the outputs based on the same given inputs. a comparison between measurements and predictions is performed. 1983).1 Performance test of the predictors Prediction performance of the LUMINA In general. outdoor global horizontal illuminance.
a series of simulations were performed with filtered measurement data sets. The prediction error was generally higher on the clear days. 1800 measurement simulation 1600 1400 1200 1000 lx 800 600 400 200 0 0 100 200 300 400 Data index(hourly integrated) 500 600 Figure 4.12). RE m = n Ep. especially when the direct sunlight fell onto a subset of the reference points. For testing LUMINA’s performance. i measured illuminance on the ith reference point E p. Figure 4. To represent the sky condition in LUMINA. geometrical and physical properties of the test bay along with time stamp. global horizontal irradiance and diffuse horizontal irradiance. Once the sky condition and the louver angle were given for a specific time segment. The overall performance of LUMINA was evaluated by calculating it’s relative prediction error (Figure 4.11) and charted in histogram (Figure 4.1) where Em. measured global and diffuse outdoor irradiance values were provided to LUMINA’s Perez sky model. LUMINA performed better on cloudy or partly cloudy days.measurement data. i predicted illuminance on the ith reference point This mean relative error value gives the same weight to the prediction errors that are n times higher or 1/ n times lower than the measurement. To facilitate the comparison.10 shows a comparison between LUMINA’s prediction and the actual measurement on a partly cloudy day. i 1 ⋅ exp -.∑ ln ---------- – 1 ⋅ 100 [%] n E m. Output of the simulation included indoor illuminance values on the pre-defined 12 reference points. The input variables included outdoor global horizontal illuminance. The mean relative error used for evaluation was based on the following definition (Mahdavi 1997a).10: LUMINA’s indoor illuminance prediction performance in the test bay 105 . the illuminance values at 12 reference points in the test bay were both measured and simulated for various times/dates and different light redirection louver angles. LUMINA predicted the indoor illuminance distribution profile. i i=1 (4.
11: Relative error of LUMINA’s indoor illuminance prediction 1 000 900 800 700 Frequency 600 500 400 300 200 100 0 20 60 40 -6 0 80 0 0 0 0 0 -1 0 -8 -4 -2 10 0 Re la tive e rror(% ) Figure 4.3. S.13).80 60 40 20 0 % -20 -40 -60 -80 0 100 200 300 400 Data index(hourly integrated) 500 600 Figure 4.3.2) 106 .1 Test of the neural networks’ learning convergence Various neural networks designed for constructing the hybrid predictors in HISSTO are implemented based on the neural network code developed by Thrun. Total 453 training patterns (D1-1: 3/3/98 -3/4/98) are used for this test.12: Histogram showing LUMINA’s relative indoor illuminance prediction error distribution 4. and tested to characterize their learning convergence trends (Figure 4. The error of each neural network is calculated by the following formula: ∑ ( ot – o p ) 2 ⁄N (4.2.2 Prediction performances of the hybrid predictors 4.
2 0.4 0.13: Learning curves of the neural networks in HISSTO 107 .1 0 0 EError rror 500 1000 1500 Training s tep Training step 2000 Error E rror 0 500 1000 1500 Training step 2000 WFC0NN 0.5 0.where o t is target output.1 0 Error Error 0 500 1000 1500 Training s tep Training step 2000 0 500 1000 1500 2000 Training step Training s tep WFC1NN 0.1 0 SMC0NN 0.4 0. MSC0NN 0.4 0.6 0.1 0 Error E rror 0 Error E rror 500 1000 Training step 1500 2000 0 500 1000 1500 2000 Training step Figure 4.6 0.2 0.4 0.3 0.6 0.4 0.6 0.3 0.3 0.3 0.2 0.5 0.4 0.5 0.2 0.2 0. and N is the number of training patterns.2 0.5 0.5 0.5 0.1 0 SMC12NN Error E rror 0.1 0 BRIDGENN 0.6 0.6 0.3 0.3 0. o p is the predicted output.
Some hybrid predictors designed and implemented in HISSTO relied on all 12 indoor illuminance sensors for both training and operation.4.Apr. the data sets for the test case 1 were not filtered to assure a close match between the Perez model’s prediction of the sky luminance distribution and the actual measurement. a hybrid predictor such as WFC1NN used only one indoor illuminance sensor in operation phase. Optimality is usually an asymptotic result. 108 .2). The fact that those hybrid predictors trained by the simulator are less accurate than the ones trained with the measurement data can be partially explained based on the test data sets being used in each test case. This measure defines how near to optimality is sufficient. This also explains why WFC1NN shows significant performance increase simply by being compensated through the calibration using one indoor illuminance sensor. and so convergence speed is an ill-defined measure. which is a more practical approach in the actual control setting.2 Test of prediction accuracy Six hybrid predictors were implemented and tested to evaluate their prediction performance. On the other hand. Although. 3 . For each time window and each reference point in the test bay. Figure 4. the same test was done with different data sets further processed by the second filter (D4 and D5 in Table 4. the selected test data sets are divided into 12 partitions to implement “sliding time window” neural network training and operation scheme for the test.15 show the accuracy of HISSTO’s hybrid predictors in predicting indoor illuminance profile using a part of the filtered measurement data sets (Mar.1) measured for “all windows unblocked” test bay configuration were used. mean relative errors were calculated.14 and Figure 4.14 shows the evaluation result focused on the test case 1.2. 1998). their RMSs (Root Mean Square) of the indoor illuminance predictions were calculated across all 11 data partitions on which the predictions are made. MSC0NN is capable of predicting indoor illuminance distribution quite accurately since it learns directly from the measurement. only a limited segment of the total data was used for the test purpose. With a proper training scheme and the fine-filtered data sets. In the test case 1. More practical is the speed of convergence to near-optimality. With test case 2. Figure 4. 18. Both SMC0NN and WFC0NN. A related measure is the level of performance after a given time. having been trained with simulation outputs based on either the measured sky condition or the weather file. then. As opposed to the data sets for the test case 2. In this test. MSC0NN and SMC12NN perform very well in their predictions (within 5-15% range). For the test case 1. The speed of a neural network’s convergence to optimality was an issue during the test. which similarly requires that someone defines the given time (Kaelbling 1995). two filtered data sets (D1and D2 in Table 4. the data acquisition period covered almost the entire year of 1998.3. show a similar trend in terms of their prediction capabilities.
especially the ones trained by the simulator.14: Different hybrid predictors’ indoor illuminance prediction accuracy expressed in RMS of relative errors (Test case 1 with D1 and D2 data sets in Table 4. Among the hybrid predictors tested.e. 30 SM MSC0NN 25 SMC0NN WFC0NN SMC12NN 20 WFC1NN RMS of relative errors [%] 15 10 5 0 1 2 3 4 5 6 7 8 9 10 11 12 Reference point Figure 4. WFC1NN which is trained by the local weather file-based simulations and calibrated by one indoor illuminance sensor shows an overall performance improvement across all reference points. MSC0NN and SMC12NN perform very well (within 3-7% RMS error range) in their prediction of the indoor illuminance distribution.1) 109 . It is also implied that using actual sky scanner data (instead of using synthetic sky models) can potentially increase the simulator’s predictive performance as well as those of the other simulationbased hybrid predictors. WFC0NN turns out to be clearly capable of copying the simulator with sufficient number of training patterns. SMC0NN and WFC0NN) show a similar prediction trend even though they are slightly less accurate than the ones trained with the measurement data. All hybrid predictors demonstrate reasonably good prediction performance (within 3-13% RMS error range) across all reference points except the ones close to the window. Selection of the sensor position for calibration is an important factor for an enhanced performance of this predictor. The simulator and those hybrid predictors trained by the simulator (i. as expected. most of the hybrid predictors show improved prediction performance.In the test case 2.
30 SM MSC0NN 25 SMC0NN WFC0NN SMC12NN 20 WFC1NN RMS of relative errors [%] 15 10 5 0 1 2 3 4 5 6 7 8 9 10 11 12 Reference point Figure 4.2) Figure 4. Simulator’s output is used for the bench mark because other hybrid predictors are dependent on the simulator’s prediction for most of the output variables except average illuminance and uniformity. especially for those prediction output variables whose values can be predicted only by the simulator. WFC1NN’s predictive pattern is also similar to those of other simulation-based hybrid predictors. Certain predicted target variables (i. validation of the simulator becomes important. Even though it is calibrated by a single sensor. MSC0NN and SMC12NN show higher deviations for certain target variables. For a meaningful comparison among multiple hybrid predictors. Because these two hybrid predictors are heavily sensor dependent (either for training or for calibration). Predictions are made for the entire output variables (both reference parameters and performance indicators in Table 3.16 shows the relative prediction performances of all hybrid predictors in HISSTO.7) and their mean relative deviations from the simulator’s prediction are calculated. 110 . which needs to be further investigated.e. solar gain) deviate more from the simulator’s prediction.15: Different hybrid predictors’ indoor illuminance prediction accuracy expressed in RMS of relative errors (Test case 2 with D4 and D5 data sets in Table 4. Test result reveals that the predictive patterns of both SMC0NN and WFC0NN are similar to what the simulator predicts. they are affected by the difference between the measured indoor illuminance profile and the simulated one which is propagated to the other target variables by the BRIDGENN.
16-1). Therefore. Test case 2 50 SM MSC0NN SMC0NN WFC0NN SMC12NN WFC1NN SM MSC0NN SMC0NN WFC0NN SMC12NN WFC1NN Test case 1 50 Average of relative deviations from simulation [%] 40 Average of relative deviations from simulation [%] 40 30 30 20 20 10 10 0 0 -10 -10 -20 R13 R14 R15 R16 R17 P1 P2 -20 P3 P4 P5 P6 P7 R13 R14 R15 R16 R17 P1 P2 P3 P4 P5 P6 P7 Reference parameters and performance indicator Reference parameters and performance indicator Figure 4. the pure simulator performs reasonably well. it is difficult to say that the hybrid predictors in the test case 2 better perform than in the test case 1. The graphs in Figure 4. the difference between the "measurement-based" and the "simulation-based" hybrid predictors’ predictions is significantly reduced. Even without a measurement-based calibration process. SMC12NN also requires multiple simulations every time it is to be deployed. even a measurement-based hybrid predictor such as MSC0NN using this “BRIDGENN” is influenced by the improved simulation performance due to the further filtered sky condition inputs.When the hybrid predictors’ relative prediction performances of the test case 2 (Figure 4. This is also true with SMC12NN having the neural network trained to predict simulation errors.16-2) is compared to those of the test case 1 (Figure 4.16: Different hybrid predictors’ performance in predicting both reference parameters and performance indicators represented as the averaged relative distances from the simulator’s prediction Despite its high predictive accuracy. the need for ever-present multiple indoor illuminance sensors not to mention the difficulty of systematic measurement for the neural networks retraining in operation time makes MSC0NN less practical.16 are all showing the relative distances of the hybrid predictors’ predictions from the simulator’s outputs rather than the relative errors compared to the measurement. Since the “BRIDGENN” is trained by the simulator. The simulator’s memory-intensive and time consuming process is greatly improved in other simulation- 111 .
They are almost sensor-independent (WFC1NN uses only one indoor illuminance sensor for calibration) while maintaining enhanced prediction capability. SMC0NN and WFC0NN were both trained with the data sets generated by the simulator based on the different sources of sky condition data.based hybrid predictors. glare indices). or WFC1NN) to generate the final prediction outputs. SMC1NN. SMC1NN. which normally takes a day to complete the process. and WFC1NN were trained with either the measurement data or the data generated by the simulator and predict the indoor illuminance profile.e. when tested with the measurement data sets. it takes the average indoor illuminance (Em) and the uniformity factor (Uf) calculated from the indoor illuminance profile predicted by the other neural network (i. For example. both WFC0NN and WFC1NN emerge as the promising ones. This measurement-free architecture also makes them immediately available for operation once they are trained with the training patterns generated by the simulator based on the local weather file. Various tested hybrid predictors tested show different trends in this analysis. The necessary training patterns can be acquired by performing daylight simulations on weekends at the beginning of each month. Obviously. These predictors are clustered together in terms of their prediction trends and show less deviations from the pure simulator’s prediction. all simulation-based hybrid-predictors need a sufficiently accurate simulator to maintain their desirable predictive performance level. This clarifies why. BRIDGENN itself and any other hybrid predictor using BRIDGENN show differences in their prediction profiles compared with the other hybrid predictors such as SMC0NN and WFC0NN. This explains why the prediction pattern of each of those hybrid predictors is similar to that of BRIDGENN. then used BRIDGENN to predict the reference parameters and the performance indicators. 112 .e. MSC0NN. MSC0NN.17 coupled with their standard deviations. Simple averages of the relative distances from the simulator’s predictions over the reference parameters and the performance indicators are shown in Figure 4. They. SMC0NN approximately inherits the simulator’s prediction capability. Although BRIDGENN was initially trained with the simulation outputs based on the local weather file. This result implies that those hybrid predictors using BRIDGENN share a similar prediction trend although it is not easy way to identify the better performing group of predictors by having no way to measure the values of those performance indicators (i. Among the simulation-based hybrid predictors.
1) 113 .17: Average relative distance and standard deviation of hybrid predictor’s reference parameters and preference indicators prediction from the simulations (with D1 and D2 data sets in Table 4.100 100 Average & standard deviation of relative distances from simulation[%] Average & standard deviation of relative distances from simulation[%] 50 50 0 0 -50 R13 R14 R15 R16 R17 P1 P2 P3 P4 P5 P6 P7 -50 R13 R14 R15 R16 R17 P1 P2 P3 P4 P5 P6 P7 Reference parameters & preference indicators Reference parameters & preference indicators MSC0NN SMC0NN 100 100 Average & standard deviation of relative distances from simulation[%] Average & standard deviation of relative distances from simulation[%] 50 50 0 0 -50 R13 R14 R15 R16 R17 P1 P2 P3 P4 P5 P6 P7 -50 R13 R14 R15 R16 R17 P1 P2 P3 P4 P5 P6 P7 Reference parameters & preference indicators Reference parameters & preference indicators WFC0NN SMC12NN 100 100 Average & standard deviation of relative distances from simulation[%] 50 Average & standard deviation of relative distances from simulation[%] P3 50 0 0 -50 R13 R14 R15 R16 R17 P1 P2 P4 P5 P6 P7 -50 R13 R14 R15 R16 R17 P1 P2 P3 P4 P5 P6 P7 Reference parameters & preference indicators Reference parameters & preference indicators WFC1NN BRIDGENN Figure 4.
it is far more sustainable than MSC0NN except the fact that it can not use such performance indicators as the average illuminance and the uniformity factor. A heuristic rule suggests that the selected reference point better be a meaningful one to the occupant in the target space (i. As is shown in Figure 4.e. 20 15 RMS of relative errors[%] 10 5 0 1 2 3 4 5 7 8 9 10 11 12 Indoor illuminance reference point Figure 4.18). fewer number of output nodes in a neural network means fewer number of the weights the backpropagator should adjust. The test result is slightly better than MSC0NN. the selection of a calibration sensor position is an important issue for this predictor’s implementation. Since this predictor needs only one indoor illuminance sensor. Fortunately. those better-performed illuminance sensor positions are also the ones on which the simulator generates better predictions. to identify the best control option. there is no significant difference in terms of this predictor’s prediction capability across all tested indoor illuminance reference points except the first and the second one close to the window. Normally. possibly due to its significantly reduced neural network output number. not to mention the others.19.MS1NN trained with only one indoor illuminance sensor value was also tested to evaluate its prediction performance (Figure 4. It should be noted that WFC1NN uses the simulator to generate indoor illuminance profile 114 . the center of the desk). As MS1NN. One of the critical issues with this predictor is to decide which indoor illuminance reference point should be selected.18: MS1NN’s indoor illuminance prediction accuracy with the selected training sensor position WFC1NN also uses only one indoor illuminance sensor value to calibrate the predictions performed by its component neural network copying the weather file-based simulator.15. Referring to the pure simulator’s performance in Figure 4. taking the sensor position either close to the window or near by the opposite wall is better than choosing one of those in the middle.
3 Test of decay in prediction performance One of the critical issues in operating a hybrid predictor is to identify the decay rate of its prediction capability after the training is done. The expectation was that a hybrid predictor’s prediction performance decay rate change due to the gradual alterations of weather pattern and solar position are less than the decay rate change invoked by a sudden building configuration modification (i. The prediction decay profile of each hybrid predictor has been tested based on the data sets for test case 1 (D1. having calibration mechanisms based on the measurement.20 shows HISSTO’s five hybrid predictors’ decay trends in the prediction of the indoor illuminance profile.before being calibrated by any of those twelve indoor illuminance reference point values. The entire test data sets cover March 4 -April 22. performs well all the way up to the end of the test period. Figure 4. This is necessary especially to determine the hybrid predictor’s retraining schedule during operation time because the obtained knowledge needs to be updated as time passes. SMC12NN and WFC12NN.e.2. The test was done by training each of the hybrid predictors only with the first data partition in D1data group followed by a series predictions over the rest of the test data partitions. 115 .19: WFC1NN’s indoor illuminance prediction performance (a) and RMS of the predicted performance indicators’ relative deviations from the simulations (b) with different calibration sensor points 4. 50 E1 E3 150 RMS of relative deviations from simulation[%] Em UI DGI GCRT 40 E5 E7 E9 E11 RMS of relative errors[%] 100 Q 30 20 50 10 0 1 2 3 4 5 6 7 8 9 10 11 12 0 1 2 3 4 5 6 7 8 9 10 11 12 Illuminance calibration sensor point a) Illuminance calibration sensor point b) Figure 4. WFC1NN shows the least decay because it is always backed up by one indoor illuminance sensor value for calibrating its initial prediction.3. It was on April 19 when a major building configuration change occurred in the test bay where most of the windows were blocked except two tinted glass windows. glazing area change in the test bay). D2 and D3). For the similar reason.
21.SMC12NN performs second best because it compensates the simulator’s predictions based on the learned error function. it obviously undercompensates simulator’s prediction since the actual indoor illuminance suddenly drops and the predictor is not updated to deal with that event properly. Because MSC0NN was trained directly with the measured data set. all tested predictors’ decay curves remain relatively stable even throughout almost a monthly weather pattern and solar position changes up until the building configuration change event. a dramatic increase of prediction errors with D3-1 and D3-2 test data partitions can be easily explained.. When this predictor encounters the window change event. its prediction decay curve shows a fairly stabilized performance across the entire test period except with the last test time window. and WFC1NN show a similar decay trend to BRIDGENN which is not informed of the 116 .20: Indoor illuminance prediction decay curves of the hybrid predictors in HISSTO Hybrid predictors’ performance decay trends in the prediction of the reference parameters and the performance indicators are shown in Figure 4. MSC0NN. Again. As is expected.. all tested hybrid predictors perform fairly well until they encounter the building configuration change event at which they show significantly increased deviations. Considering the weather and solar position changes during this test period. SMC12NN. Their prediction decay curves show slightly increased overall deviations from the actual measurement compared to those of the other predictors. Since those predictors were initially trained based on the “all windows unblocked” building configuration.. SMC0NN and WFC0NN behave similarly in terms of their prediction decay profiles. 150 Avaraged abstarct value of relative errors[%] MSC0NN SMC0NN WFC0NN SMC12NN 100 WFC1NN 50 0 D1-3 D1-4 D1-5 D1-6 D1-7 D2-1 D2-2 D2-3 D2-4 D2-5 D3-1 D3-2 Tested data partition with time progression Figure 4. Being trained by the simulator. SMC0NN was trained with the simulation outputs based on the first data partition in the first test data set 1 (D1-1) whereas WFC0NN utilized the Pittsburgh weather file-based simulation outputs for its training.
especially for Ed and DGI prediction. This is because SMC0NN only relies on a limited number of training patterns (D1-1) whereas WFC0NN uses much larger number of training patterns covering the entire month (March).building configuration change event. . SMC0NN’s prediction decay pattern is different from that of WFC0NN.21: Decay curves of the hybrid predictors’ predictions for the reference parameters and performance indicators expressed in terms of their relative distances from the simulation outputs 117 . 1000 1000 Average of relative deviations from simulation [%] 900 800 700 600 500 400 300 200 100 MSC0NN Em Average of relative deviations from simulation [%] 900 800 700 600 500 400 300 200 100 SMC0NN Em UI DGI GCRT Q UI DGI GCRT Q 0 -100 D1-3 0 -100 D1-4 D1-5 D1-6 D1-7 D2-1 D2-2 D2-3 D2-4 D2-5 D3-1 D3-2 D1-3 D1-4 D1-5 D1-6 D1-7 D2-1 D2-2 D2-3 D2-4 D2-5 D3-1 D3-2 Tested data partition with time progression 1000 Tested data partition with time progression 1000 Average of relative deviations from simulation [%] 900 800 700 600 500 400 300 200 100 WFC0NN Em UI DGI GCRT Q SMC12NN Em UI DGI GCRT Q Average of relative deviations from simulation [%] D3-1 D3-2 900 800 700 600 500 400 300 200 100 0 -100 D1-3 D1-4 D1-5 D1-6 D1-7 D2-1 D2-2 D2-3 D2-4 D2-5 0 -100 D1-3 D1-4 D1-5 D1-6 D1-7 D2-1 D2-2 D2-3 D2-4 D2-5 D3-1 D3-2 Tested data partition with time progression 1000 1000 Tested data partition with time progression WFC1NN Average of relative deviations from simulation [%] Average of relative deviations from simulation [%] 900 800 700 600 500 400 300 200 100 Em UI DGI GCRT Q BRIDGENN Em UI DGI GCRT Q 900 800 700 600 500 400 300 200 100 0 -100 D1-3 D1-4 D1-5 D1-6 D1-7 D2-1 D2-2 D2-3 D2-4 D2-5 D3-1 D3-2 0 -100 D1-3 D1-4 D1-5 D1-6 D1-7 D2-1 D2-2 D2-3 D2-4 D2-5 D3-1 D3-2 Tested data partition with time progression Tested data partition with time progression Figure 4.
those prediction deviations are significantly reduced after updating the simulator and retraining the neural networks with the new simulation result. WFC0NN in the test case 1 shows a significant prediction deviation because of no predictor retraining even after the actual window configuration change event in D3-1. Actually. D2-5.4 Test of change adaptation performance The change adaptation capability of each hybrid predictor has been also tested using three continuous test data partitions. and D3-2 (Figure 4. Ed.23 illustrates how HISSTO’s hybrid predictors behave during the window configuration change event in terms of their predictions over the reference parameters and the performance indicators. On the other hand. Test case 1: weather file is not updated 200 MSC0NN SMC0NN WFC0NN Test case 2: weather file is updated 200 MSC0NN SMC0NN WFC0NN Average of relative errors[%] SMC12NN WFC1NN Average of relative errors[%] SMC12NN WFC1NN 100 100 0 D2-5 D3-1 D3-2 0 D2-5 D3-1 D3-2 Tested data partition with time progression Tested data partition with time progression Figure 4. As is shown in the test case 2. those variables showing more deviation are the ones sensitive to the direct solar component which happens to be cut down dramatically by the window 118 . Test result reveals that the three hybrid predictors such as MSC0NN.4.22: Change adaptation performance of the hybrid predictors in their indoor illuminance prediction Figure 4. Those test data partitions cover only the transition period from the original test-bay window configuration to the one changed. D3-1. This is because D3-1 and D3-2 include already updated measurement as well as the corresponding simulation result. WFC1NN is also influenced by the window configuration change event. now show the increased relative distances from the simulator’s prediction over a subset of the output variables (i.e. Similar to the other tests.3. SMC0NN.22). DGI and Q). Those two predictors such as MSC0NN and SMC12NN which once maintain fairly stable indoor illuminance predictions throughout the same change event. and SMC12NN recover their indoor illuminance prediction capabilities after a single catastrophic deviation shown with the test data partition D3-1. the “sliding time window” neural network training scheme was used for this test. This is mainly because of not updated BRIDGENN.2.
Test case 1: weather file is not updated Average of relative deviations from simulation [%] 800 700 600 500 400 300 200 100 0 -100 D2-5 D3-1 D3-2 Test case 2: weather file is updated Average of relative deviations from simulation [%] 900 800 700 600 500 400 300 200 100 0 -100 D2-5 D3-1 D3-2 Em UI DGI GCRT Q 900 SMC0NN Em UI DGI GCRT Q Tested data partition with time progression 900 800 700 600 500 400 300 200 100 0 -100 D2-5 D3-1 D3-2 900 800 700 600 500 400 300 200 100 0 -100 D2-5 Tested data partition with time progression Average of relative deviations from simulation [%] WFC0NN Em UI DGI GCRT Q Average of relative deviations from simulation [%] Em UI DGI GCRT Q D3-1 D3-2 Tested data partition with time progression 900 800 700 600 500 400 300 200 100 0 -100 D2-5 D3-1 D3-2 900 800 700 600 500 400 300 200 100 0 -100 D2-5 Tested data partition with time progression Average of relative deviations from simulation [%] WFC1NN Em UI DGI GCRT Q Average of relative deviations from simulation [%] Em UI DGI GCRT Q D3-1 D3-2 Tested data partition with time progression Tested data partition with time progression Figure 4. WFC1NN is better in dealing with this type of change event than WFC0NN because of its calibration capability using one indoor illuminance sensor. Again.blocking event. these deviations are mostly corrected after updating neural networks by the retraining process. The selected calibration sensor informs WFC1NN of the change event although it still uses the previous version of BRIDGENN to generate the final prediction outputs. SMC0NN shows no difference in both test case 1 and test case 2.23: Change adaptation profile calculated as the closeness of each hybrid predictor’s reference parameters and performance indicators prediction to the simulation output 119 . WFC0NN and WFC1NN are the ones mostly affected by this event.
The assignment of numbers are mostly based on the performed in this research.24 show the evaluation result of HISSTO’s hybrid predictors based on the various prioritized criteria. Table 4. yet it shows how each hybrid predictor performs relative to the others.5: Overall evaluation of the hybrid predictors designed in HISSTO Evaluation criteria BN MSC0NN MS1NN SMC0NN WFC0NN SMC12NN WFC1NN Evaluation criteria Accuracy Agility in prediction Sensor independency in operation Change adaptation Decay resistance Sensor independency in training Total credit 10 9 8 7 6 5 4 7x10 6x9 6x8 2x7 5x6 2x5 2x4 234 1x10 5x9 6x8 5x7 5x6 2x5 4x4 194 7x10 1x9 6x8 6x7 2x6 3x5 5x4 225 7x10 2x9 6x8 6x7 3x6 4x5 6x4 240 7x10 4x9 6x8 2x7 6x6 5x5 2x4 231 7x10 5x9 6x8 5x7 4x6 6x5 6x4 276 300 250 200 Credit 150 100 50 0 MS12NN MSC0NN MS1NN MS1NN SM12NN SMC0NN WF12NN WFC0NN SMC12NN SMC12NN WFC1NN WFC1NN Predictor Predictor Figure 4.24: Overall evaluation of the HISSTO’s hybrid predictors in a graphical form 120 . The number in each cell. Adding up all those number calculated for each cell in a specific predictor column yields the total credit it gets.Table 4. Each number in a cell represents a relative strength of the corresponding predictor in the specified evaluation category. There are needs for some assumptions and possible subjectivity in this evaluation.5 and Figure 4. then is multiplied by Base Number (BN) given in the second column to reflect the priority.
(Em+Esd)-1 (4.3) Here Em and Esd are the mean and standard deviation of the illuminance levels measured at various locations in the test space.1 Test of predictive control scenarios Test of simulation-based control Three simulation-based control scenarios have been tested. The second test is a combined daylight and electrical lighting control scheme with combined daylight and electrical light simulations. where U = Em.4. The first function aims at minimizing the deviation of the average (daylight-based) illuminance level Em in the test space from a user-defined target illuminance level Et (say 500 lx).2) The second objective function aims at maximizing the uniformity of the illuminance distribution in the test space as per the following definition (cp. as well as the outdoor measurements at time interval ti are used as model input) for the four louver positions selected.4. Based on the predefined objective functions. Two illustrative objective functions were considered for the test. the simulation tool predicts the expected illuminance levels in the space for the time interval ti+1 (test space geometry and photometric properties.4. automatic determination of the "optimal" louver position (among four standard louver positions) toward fulfillment of daylight related objectives was tested (Mahdavi et al.4 4. Mahdavi and Pal 1999): Maximize U. 4. The first scenario uses the simulator to select the best louver angle considering only daylight. The simulation-based louver control scenario in this case followed the GAT approach. Minimize(|Et–Em|) (4.1 Test of simulation-based daylight control As an initial feasibility test of the simulation-based control approach. 2000). At time interval ti. The third test case is a simulation plus Luminaire Matrix based daylight and electrical lighting control scheme which dramatically reduces computational time and memory demand. 121 .1.
including a total of 28 time instances. 100 67 Control Quality 33 0 1 2 3 4 Ranked Louver Positions Figure 4.25: Mapping between ranked louver positions and control quality index These results (Table 4. is less affected by the absolute errors in the predictions of the simulator.6: Overall evaluation of the daylight control test Control Quality Index Illuminance 73.the simulator identifies the louver position which is likely to maximize the light distribution uniformity or to minimize the deviation of average illuminance from the target value. Table 4. To numerically evaluate the performance of this simulation-based control approach via a “control quality index”.25. To evaluate the performance of this simulation-based control reasoning. 100 points was assigned to intervals in which the simulation-based recommendation matched the position which was empirically found to be the best. Intermediate cases were evaluated as per the Figure 4. the measured average illuminance and the uniformity were ranked according to the degree to which they fulfilled the objective functions. which. in contrast to the illuminance.7) demonstrate an encouraging potential for the feasibility of the proposed approach.8 Uniformity 98. The better performance in the case of the uniformity indicator is due to the "relative" nature of this indicator. 0 was assigned to the interval. In those cases where the recommendation was furthest from the optimal position.9 122 . The resulting illuminance levels were measured for the four louver positions and for all selected time intervals during the test period.6 and Table 4.
870633 0.61774 123 .833 319.29732 2.641637 0. For electrical lighting part of the simulations.75 340.786171 0.9: 10 best control actions suggested by the pure simulator θn+1 (lvr)[degree] L1[%] L2[%] L3[% L4[%] Em[lx] ] 0 15 0 0 15 15 0 0 0 15 20 20 20 30 30 20 40 30 20 20 20 20 30 20 20 30 20 30 40 40 20 20 30 20 20 30 20 30 40 40 20 20 20 30 30 20 40 30 20 20 339 317.333 340.417 318.10193 2. three luminaire power levels.07142 5.08702 5.1. a total of 18 simulation runs are necessary to come up with the 10 best control actions presented in Table 4. 9 Iglob[W/m2] 81 Idiff[W/m2] 78 Eglob[lx] 9.07142 5.641822 0.865349 0.29732 2.790394 DGI 5.870633 Q[W] 2.379 θn (lvr)[degree] L1[%] 30 L2[%] 30 L3[%] 30 L4[%] 30 0 Table 4.657004 0.785386 0.865349 0.10193 2. the daylight responsive lighting control scenario is tested.Table 4.29732 2.10193 2.29732 2.870633 0. only three louver angle (the current angle plus 15 degrees higher and lower than the current angle) are considered for daylight part simulations.788617 0.865349 0.4.29732 2.667 340 339.583 UE 0.29732 2.665948 0.08702 5.786909 0.789567 0.9.08702 5.07142 5.785798 0.10193 P[W] U 96 96 120 120 120 120 144 144 144 144 0. To reduce potentially a large number of simulation sessions.08702 5.865349 0.656926 0. Table 4.8: Initial state for the inputs to the predictor Year 1998 Month . The initial systems states with the time stamp serve as the simulation inputs (Table 4.681061 0. 3 Day 14 Hour.07142 CGI GCRT 0 0 0 0 0 0 0 0 0 0 0.865349 0.833 318.784183 0.632771 0.790177 0. the immediate higher state.632759 0.667 340.8). namely the status quo. and the immediate lower state are considered for each of two groups of luminaires (L1+L4 and L2+L3).7: Percentages of instances with a specific control quality index Control Quality Index Percentage of total instances Uniformity 100 67 33 0 96 4 Illuminance 61 11 17 11 4.632856 0.78515 0.865349 0.870633 0. As a result.2 Test of simulation-based daylight responsive lighting control Control with the pure simulator Using only the simulator.08702 5.08702 5.
70231 6.794675 0.667 274 259 253 240.91505 8.883333 0.803117 0.99742 1.76282 6.10193 P[W] 0 0 0 0 0 0 0 0 U 0.94012 8.81757 8.917 438.400497 0.5 UE 0.88125 0.13 illustrates 10 best daylight responsive lighting control options 124 .786916 DGI 8.828507 0.Control with the simulator plus Light Matrix Another test is performed using the simulator and a pre-calculated Luminaire Matrix (Table B.8 34.9175 0.2. all 625 different luminaire power level combinations (five discrete power levels for each luminaire) are evaluated to obtain the 10 best dimming options (see Table 4.862567 0.26749 2.920625 0.45.421751 0.8 29 29 29 U 0.845033 0.75 296.534375 0.5 438.94012 9.705528 DGI 7.395913 (weight factors: wEm: 0.00905 9. Compared to the simulation only approach.6 40.79449 0.4. Table 4. this simulation plus Luminaire Matrix approach is much more efficient toward extending the search coverage in the control option state space. wCGI: 0.868681 Q[W] 0 0 0 0 0 0 0 0 0 0 0 P[W] 46.833438 0.833 313.801891 0.707153 0.583 452. wUE: 0.431753 0.954213 0.925105 0.5 452.5 467 453.88125 0.8).083 UE 0.02388 1.10569 2.333 452.4 40.881458 0.11: 10 best control actions suggested by the simulator + Luminaire Matrix predictor θn+1 (lvr)[degree] L1[%] L2[%] L3[%] L4[%] 105 105 105 105 105 105 105 105 105 105 105 20 20 20 10 10 10 20 20 20 10 20 20 10 20 10 10 20 10 20 20 20 20 20 20 20 20 20 20 10 10 20 20 10 20 10 10 20 10 10 10 10 20 20 10 Em[lx] 481.862567 0.786228 0.917 468.845208 4.8 34.04048 9.975549 0.1.691845 0.698465 0.786916 0.98526 5. After the most promising louver position is identified by the system (Table 4.12 shows MSC0NN-based controller’s evaluation of each louver position based on the seven performance indicators.29442 6.61243 5. wGCRT: 0.90942 8.838888 0.04048 CGI 0 0 0 0 0 0 0 0 0 0 0 GCRT 0.2 Test of hybrid predictor-based control With the same initial systems state. the alternative control schemes with various hybrid predictors are tested. Table 4.847292 0.793781 0.952298 0.974216 0.559583 0.6 40.05.694442 0.12.494544 0.72461 8.8 34.456758 0.10: Order of the best louver positions suggested by the simulator θn+1 (lvr)[degree] 105 90 75 60 0 45 30 15 Em[lx] 323.679896 0.083 438. wDGI: 0.79449 0.29732 1.6) based on the same initial systems states (Table 4.86377 6.827625 0. Table 4. Daylight simulations for eight different louver positions (from 0 to 105 degrees in 15 degrees interval) are performed first. wQ: 0.25 467.05) Table 4.687529 0.822971 0.81446 8.668466 0.91875 0.992165 0.6 34.25 237. and wP: 0.98669 2.9907 Q[W] 2.81997 8.89557 CGI 0 0 0 0 0 0 0 0 GCRT 0.845208 0.10).03.850581 0.954792 0.803996 0.868681 0.00177 5.99867 2.941588 0.81087 0.11).
805 UE 0.08477 P[W] 0 0 0 0 0 0 0 0 U 0.702954 0.8 34.326 353.6815 6.8).581875 0.995556 1.074 170.03569 CGI 0 0 0 0 0 0 0 0 0 0 GCRT 0.5 289.46231 6.524105 0.844186 0.58683 6.suggested by the controller using MSC0NN and the Luminaire Matrix look-ups.69931 0.6 40.815359 0.75 320.788894 0.914115 0.510435 0. Test results of the other hybrid predictor-based controllers are presented in the Appendix B.74633 0.583958 0.72866 4. Table 4.848127 Q[W] 0 0 0 0 0 0 0 0 0 0 P[W] 46.28021 2.551042 0.985962 0.5862 7.79376 CGI 0 0 0 0 0 0 0 0 GCRT 0.75 UE 0.4 40.6205 4.844884 0.62125 0.844139 0.359 343.854437 0.8 Table B. Table 4.37001 6.71267 7.51375 0.74466 0.833871 0.2185 5.917224 0.137 146.426448 0.62746 7.71607 7.811451 DGI 7.920692 0.42856 0.35295 4.783019 0.585208 0.587083 0.369602 Table 4.697505 0.27666 2.25548 2.583 332.702124 0.057 182.521608 0.15 show the result of WFC1NN-based controller test under the same conditions (Table 4.167 305.45921 7.911358 0.85102 8.6 40.86159 1.411213 0.912815 0.389 329.5895 7.88111 1.00118 1.72047 7.694596 0.505 177.68033 4.52683 0.19675 2.01058 1.103 350.991702 0.2 U 0.530025 0.58194 7.738368 DGI 6.083 333.421119 0.547917 0.8045 1.76631 CGI 0 0 0 0 0 0 0 0 GCRT 0.746484 0.424007 0.700442 0.298 161.63399 6.917 319.54948 4.452 100.6 34.743364 0.24498 2.83685 1.417149 0.45613 4.549792 0.18524 3.26605 2.74962 1.04345 Q[W] 1.483013 Table 4.746651 0.682252 DGI 4.15.25 UE 0.810467 0.915258 0.8 29 23.835154 0.835906 0.2276 2.517443 0.417 319.90946 1.794022 0.808058 0.74544 0.98368 0.988443 0.487803 125 .949 187.14: Order of the best louver positions suggested by the WFC1NN predictor θn+1 (lvr)[degree] Em[lx] 105 90 75 60 45 30 15 0 355.824164 0.5459 P[W] 0 0 0 0 0 0 0 0 U 0.055 295.13: 10 best control actions suggested by the MSC0NN predictor θn+1 (lvr)[degree] 105 105 105 105 105 105 105 105 105 105 L1[%] L2[%] L3[%] 20 20 20 10 20 20 10 10 10 10 20 20 10 20 20 10 20 10 10 10 20 20 20 20 10 20 20 20 20 10 L4[%] 20 10 20 20 20 10 10 20 10 10 Em[lx] 348.8 34.12: Order of the best louver positions suggested by the MSC0NN predictor θn+1 (lvr)[degree] Em[lx] 105 90 75 60 45 30 15 0 190.745619 0.939 338.52436 6.912167 0.528404 0.802422 0.14 and Table 4.6 40.401688 0. Table B.5 334.787762 0.701299 0.89522 1.933347 Q[W] 2.809252 0.833 334.846073 0.941 347.
841503 0.25 UE 0.25 second to perform the same task.16: Final systems control states (with utility values) suggested by the tested predictors Predictor type SM SM + LM MSC0NN SMC0NN WFC0NN SMC12NN WFC1NN 0 105 105 105 105 105 105 θn+1[degree] L1[%] 20 20 20 20 20 20 20 L1[%] 20 20 20 20 10 10 10 L1[%] 20 20 20 20 20 20 20 L1[%] 20 20 20 10 10 10 10 U 0. the time necessary for finishing one session of simulation was found to be approximately 38 seconds while a neural network took only 0.83521 0.804443 0.803636 0.8 40. The exact numbers could vary depending on various parameters.6 34.33 second.Table 4. Eight louver positions and five dis- 126 . most controllers that use any of the hybrid predictors suggest similar systems control states for both the louver and the luminaires.809281 0.583 483.83329 DGI 5.68 0.840976 0.789809 0. As is shown in Figure 4.823649 0.6 34.333 498.825736 0.5 498 482.834702 0.8 40.62 0.83 4.4.833744 0.78 0.95437 5.26.824078 0.834286 0.824455 0.833 499.819778 0.8 U 0.842869 0. The controller using a hybrid predictor performs best in this test.794113 0.824168 0.810098 Q[W] 0 0 0 0 0 0 0 0 0 0 P[W] 34.825299 0.86764 5.824922 0.98031 5.89529 5.5 483.15: 10 best control actions suggested by the WFC1NN predictor θn+1 (lvr)[degree] L1[%] L2[%] L3[%] L4[%] 105 105 105 105 105 105 105 105 105 105 20 20 10 20 10 20 10 20 20 10 10 20 20 10 10 20 20 10 20 20 20 20 20 20 20 10 20 10 10 10 10 10 10 20 20 10 20 20 20 20 Em[lx] 484.8 40. The controller that uses only the simulator comes up with different control recommendation due to its limited system control state search space (necessary to reduce computation time).805352 0.77 0.824601 0.75 497.823319 As is shown in Table 4.79 0.3 Test of the convergence time to reach a control action decision Another test shows the differences of the time necessary for a controller to finish all predictions and evaluations to come up with a control decision.75 0. Table 4.98552 5.95211 5. The average time for an instance of the Luminaire Matrix lookup and control option evaluation was approximately 0.16.86985 5.8 40.8 34. yet this clearly shows that prediction speed could be enhanced by using neural network-based hybrid predictors trained with the simulator.95729 5.790587 0.167 482.84194 0.25 484.824972 0. Table 4.17 shows the time needed to identify the desirable system control states for three controllers using different predictor types.6 34.86478 5.98325 CGI 0 0 0 0 0 0 0 0 0 0 GCRT 0.6 34.799874 0.835689 0.
000 combined louver-luminaire states separately (equivalent to 190. The total time measured reveals that less than a half of the time is necessary for a hybrid predictor method compared to the simulator plus Luminaire Matrix method.000 seconds) to finish the task. The system still has to evaluate each of the control options to identify the best. The total time needed for this method covers luminaire matrix lookups and the evaluation of all 625 luminaire power level combinations. When only the simulator is used. When the simulator and the Luminaire Matrix are used jointly. The hybrid predictor. a total of nine simulations are necessary to finish the same task including eight for the identification of the best louver position and one more for calculating daylight glare components based on the determined louver position before being integrated with the ones induced by the electrical light.based controller performs also disjointed daylight-electrical light predictions.33 second. Time for Control Decision [sec] 200000 150000 Prediction Evaluation 100000 50000 0 SM SM+LM WFC0NN+LM Type of Predictor Time for Control Decision [sec] 1000 800 600 400 200 0 SM+LM WFC0NN+LM Type of Predictor Prediction Evaluation Figure 4.crete power levels for each luminaire are considered in the test. each).26: Time requirements comparison needed for a single control time window 127 . which takes approximately 1. it has to simulate each of 5. Except two seconds dedicated to finish the daylight predictions for eight different louver positions. it follows the same process as the simulator plus Luminaire Matrix method.650 seconds (0.
• In terms of fulfilling non-functional requirements. any hybrid predictor being trained or calibrated by the measurement data shows an improvement in precision compared to the pure Simulator (TH2).000 304 2 1. Change adaptation is made possible by updating simulation inputs or having at least one indoor illuminance sensor for constant on-line calibration as is in the case of WFC1NN (TH4).650 244 244 Total [sec] 191.17: Comparison of the required time for HISSTO’s different controllers Controller SM SM+LM NN+LM Prediction [sec] Evaluation [sec] 190.650 548 246 4. fulfillment of the use cases. Web-based remote access to HISSTO is demonstrated by 128 . • All use cases involving direct user interactions are covered or at least tested on the internet environment (UC1-UC7). a hybrid predictor such as WFC1NN can perform well with only one indoor illuminance sensor for real-time on-line calibration (NFR3).5 Overall evaluation of HISSTO The evaluation of HISSTO mainly addresses the three different categorized issues . In terms of prediction accuracy. HISSTO’s entire design is structured for enhanced modularity and flexibility through a wide use of the software design patterns and component technology so that it can facilitate potential system extensions (TH5).Table 4. and those non-functional requirements specified in the requirement elicitation section. those hybrid predictors in HISSTO can predict within ± 20 % relative error range by having approximately 400 training patterns along with the encoding/decoding technique devised in this thesis (NFR1). All hybrid predictors except MS1NN cover seven visual performance indicators for the evaluation of the candidate control options by inheriting the simulator’s modeling capability (TH3).the verification of the thesis hypotheses. Most hybrid predictors finish the desirable control option selection process under 5 minutes after evaluating eight louver positions and 625 luminaire power level combinations (NFR2). • The speed to reach the control decision is significantly increased by using neural networks across all hybrid predictor-based controllers (TH1). Trained with weather file-based simulations.
Designed and implemented.17). Compared to pure simulator. hybrid predictors are much faster in their predictions (Table 4.18 shows the summary of HISSTO’s evaluation results: Table 4. Enhanced adaptation to A pure neural network is expected to perform on-line change adaptation. Designed and implemented. WFC1NN uses only one sensor to calibrate initial predictions.15). System flexibility and adaptability Monitoring Manual control Scheduled control Predictive control Predictor update Preference change Predictor parameters change System flexibility is maximized by using design patterns and componentbased object-oriented programing methodology. and energy use into control criteria.18: Assessment of the research outcome including HISSTO implementation ID TH1 TH2 TH3 TH4 Description Prediction speed Prediction accuracy Richness of control criteria Evaluation Compared to pure simulator. Future cross-domain interfacing is facilitated by decoupling implementations from their corresponding representations based on the various software design patterns (NFR5). heat gain. Designed and implemented Designed and implemented Designed and implemented Designed and implemented. Figure 4. hybrid predictor such as WFC1NN can generate more precise predictions (Figure 4. TH5 UC1 UC2 UC3 UC4 UC5 UC6 UC7 NFR1 Prediction error NFR2 10-15 minute time limit All tested hybrid controllers finish the process within 10 minutes. Web-based monitoring and control are implemented and tested. NFR3 Minimum sensors NFR4 Remote access NFR5 Cross-domain interface API for cross-domain operations needs to be built. the changes which is not possible due to the sensor number limit in operation. HISSTO incorporates glares. Manual update is necessary when a change occurs.14. Beyond a mere average indoor illuminance based control.using LabviewTM and ComponentWorksTM (based on ActiveX technology) off-shelf development environment (NFR4). Table 4. 129 . Designed and implemented Most of measurement-based hybrid predictors and calibrated predictors satisfy ± 20 % relative error limitation.
2) Prototypical realization of a new perspective on building simulation as a key contributor for enrich- 130 . setting up an interface for the simulator and the machine learners.1 Contributions The hybrid approach explored in this thesis is a new framework capable of efficiently handling building systems operation while satisfying multiple control criteria. This process included control instrumentation.5 Conclusion 5. Beyond mere reactive operations based on environmental sensing. allowing effective adaptation to the potential system or environment changes. This thesis illustrated how computational modeling could be applied to enrich the informational repertoire of systems control operation for lighting. the implementation and the test of such hybridizations. model-based building control allowed for proactive evaluation of a richer set of control options. A highly attractive feature of this model-based strategy came from its potential for a transparent and high-level integration of multiple control agenda. This new approach is trying to augment simulation’s role as a building design evaluation tool to a building operation support tool. particularly the following results were achieved: 1) Development of a framework enabling an effective model-based building control. and establishing a control option evaluation process (using GAT & BDI method along with the user preference function definition). The anticipated synergistic effects between simulation and machine learning have been validated through the design. On the other hand. Reducing computational load in prediction also enables extended search in the system state space within a limited time window. By performing this research. A complex control strategy has been formulated to simultaneously address economical and ecological considerations in providing appropriate levels of building visual performance. Maximization of flexibility in control software development and user-system interaction have been also demonstrated by fully exploiting software design patterns as well as the distributed component technology. The hybrid system turned out to be a promising option especially to avoid heavy computational load and to reduce sensor dependency. simulation’s capability in building control is proven to be enhanced by being combined with machine learning technique. This has been demonstrated in daylight responsive lighting control domain.
2 Future Research Future research must address some of the remaining issues. vendor-independent sensing and actuation as well as the location-independent and cross-platform user interactions. 5. This has been done by using various software design patterns in addition to the web-based component technology such as Active XTM. the proposed system’s implementation needs to be extended to accommodate multi-zone controls. The design and the training scheme for each machine learner in HISSTO needs to be further improved.ing control criteria even across multiple domains. possibly based on reinforcement learning technique. Considering the scope of the present work. This also requires 131 . Controlling lighting systems across multiple zones in a building requires extended work. reducing its heavy sensor-dependency. A machine learner. Since a typical building has multiple rooms and control zones. further research must address: 1) Extension of the control scope from one control zone to multiple zones in a building. whereby glare indices and solar gain were considered as a part of control option evaluation criteria beyond conventional simple illuminance assessment. 6) Development of object-oriented and modular control software architecture. and enabling it to capture the behavioral characteristics of complex control parameters. The hybridization approach is shown to be promising in resolving this issue as well. One of these hybridization schemes which uses weather file-based simulations as the source of neural network training knowledge turns out to be especially promising. 5) Exploration of the ways to deal with gradual or radical changes in building systems or the environment. 3) Extending simulation’s role in building control by providing a machine learner with prior knowledge to be trained with. 4) Development and validation of the various ways for on-line hybridization between the simulator and the machine learner to enhance control system’s prediction capability in both speed and accuracy. This has been demonstrated through the pilot tests in the daylight responsive lighting control domain. HISSTO’s applicability to the different building control domains also needs to be explored. could learn how to customize personal preference functions for a specific occupant. The control system developed has been tested only with one space of a building although it’s software design allows to control an entire building. The developed system architecture integrates the software components for on-line predictor hybridization.
6) Extension of the developed hybrid system’s applicability to multiple control domains in a building. 4) Development of a machine learner that captures individual user’s system control preference by observing each user’s response pattern to the control actions made by the system. If there is a sensor to measure a certain control parameter. Currently the control system is using a file system to organize persistent data storage. However. The same hybrid control system architecture can be used for the other building control domains such as thermal or air flow control. 2) Identification of possible sensing mechanisms for currently non-measurable control parameters (i.e. Capturing individual user preference is desirable because then the control actions can be customized for each individual user. which can greatly influences the prediction performance of the neural network. By recording and interpreting accumulated user interactions with his/her physical environment. based on reinforcement learning technique) can learn how to customize default preference functions and weight factors of all performance indicators for that specific individual. The implementation of a DataBase Management System (DBMS) can increase both efficiency and effectiveness of all data operations. New sensing technologies may ease this problem in the future. glare. The implemented system for this research focused on the visual environment in a building.the development of a global control algorithm which can handle potential conflicts among different control zones. How extensively the system architecture must be restructured and/or extended to cover such other building control domains remains to be further investigated 132 .if not impossible . a machine learner (possibly. certain parameters such as glare and solar gain are difficult . 5) Database implementation to allow sophisticated and efficient data operations especially for better online machine learner training and trend analysis. This has impact on the effectiveness of the neural network training pattern encoding and predicted output decoding. in practice.to be measured with a reasonable cost. solar gain) to validate simulator’s prediction. 3) Improvement of the design and the training scheme for each machine learner in the system to enhance its prediction capability. One example is the way to estimate the variation range of each control parameter within a certain period of time. it can dramatically enhance hybrid predictor’s prediction capability. especially for the system diagnostics and the trend analysis.
. Bruegge. (1996).4. Curtis. (2001). D. ISBN 85 . Cambridge. J. S. Adam Hilger Ltd... Maxion. NY. P. Requirement Analysis Document: OWL Project. Connectionist Learning for Control. T. Object-Oriented Software Engineering: Conquering Complex and Changing Systems. S. The MIT Press. Neural Networks for Control. Pittsburgh. Accord. New Jersey. S. Rio de Janeiro. G... A. (1999). OWL project team members (1996). R. PA. Wiley. II. PA. GA. Vol.) (1995). MA. P. Frameworks. B. Reading.. Georgia Institute of Technology. ASHRAE Transactions. Simulation of the PID Algorithm for Direct Digital Control Applications. Doing Hard time: Development of Real-time systems with UML. Carnegie Mellon University. Objects. 849 . pp. CMU-CS-99-102.3 . Carnegie Mellon University. School of Computer Science. (1985). Scenario-Based Design: Envisioning Work and Technology in System Development. Brandt. Chang. MA. 133 . New York. Longman Inc. Comparing Methods for Multivariate Non-participant Regression. B. Shavit.901939 . B. A. Mahdavi. Douglass. A. Olszewski. G. (1999).856. Energy simulation in Building Design. Barto. Pittsburgh. Brazil. School of Computer Science. L. (1999). MA. A. R. Experimental Results for a Network-Assisted PID controller. Caroll.6 References Banks. Bruegge. 96-21-3. Prentice Hall. Proceedings of the Workshop on HVAC Controls Modeling and Simulation. G. Addison-Wesley... J. (1990). (1984). Dutoit. A Hybrid System for Daylight-Responsive Lighting Control. A. M. Proceedings of the Seventh International Building Simulation (IBPSA) Conference. (Ed. and Patterns. Clark.
(1992). Vol. pp. A. ASHRAE Transactions: Research.. School of Computer Science. B. E. A.. E. Addison-Wesley Publishing Company Inc. GA. (1971). 1. pp. Curtiss. Vol 143. A Computational Environment for Active Lighting Design. L.. New York. A Methodology for Validating Building Energy Analysis Simulations. (1997).. P.. (1998). M. Georgia Institute of Technology.. (1996). (1984). D. SERI (Sustainable Europe Research Institute) Report. 90. Discomfort Glare: A Formula to Bridge Differences. Hartkopf. School of Architecture. Reinforcement Learning: A survey. Neural and Genetic Algorithm Control Applications. Park. N.. 2. No. D. Vlissides. F. Gamma. A Dynamic Building /HVAC/ Control Systems simulation Program. Construction Research and Development Journal. H... J. PA. 134 ..D. P.. Vol. Carnegie Mellon University.. Z. Experimental Results of a Predictive Neural Network HVAC Controller. G. Vol. A. 401-415. Nyongesa. J. Judkoff. Assawamartbunlue. 3. (1994). D. (1997).V. 11. P. Shankavaram. S. Lighting Research & Technology. J. R.. Carnegie Mellon University. F. Van Nostrand Reynold. D. Moor. Glare from Windows. Jeannette.. (1983). An integrated approach to design and engineering of intelligent buildings . L. Li. Johnson. Kaelbling. May W. Burch.. S. O’Doherty. published by Elsevier Science B. Architectural Interior Systems: Lighting. R. Pittsburgh. No. IEEE Proc. A Comparison of Sky Luminance Models with Measured Data from Garston.. Kelly. PA.. p. D. R. Kremers. 6.. Steffy G.. A. (1979).J. D. Wortman. Segil. Clark. HVACSIM+. Ph. W. Learning Systems in Intelligent Control: An Appraisal of Fuzzy. Littlefair. R. No 4. Acoustics. H.. W. Hopkinson. Vol. R. No. V. Flynn. E. Helm. R. 2.The Intelligent Workplace at Carnegie Mellon University. A.. Loftness. Air Conditioning.-Control Theory Appl. Pittsburgh. Littman. Design Patterns: Elements of Reusable ObjectOriented Software. J. Mahdavi. 4. 23-28. pp. (1995). R... 169-175. Lee. K. Proceedings of the Workshop on HVAC Controls Modeling and Simulation. Thesis. C. J. O. (1995). V. Linkens. G. Kreider. in Automation in Construction.Einhorn. J. J..
(1995). and Control Strategies on Energy Use in Commercial Buildings. A.. Volume 28. 53. Exploring Model-Based Reasoning in Lighting Systems Control. A Prototypical Realization of an ’Open’ Daylighting Simulation Environment. V. Pal. McGraw Hill Inc. Vol.230. 1994: pp. Nonlinear HVAC Computations Using Neural Networks. Toward and Entropy-based Light Distribution Uniformity Indicator. 29... S.. Mahdavi. Berberidou. Summer 1994. Number 1.35. (1999).214. Vol. Machine Learning. K. pp.) pp. pp.. Proceedings of Building Simulation '99.. S. Sixth International IBPSA Conference. Narendra. V. Mahdavi. 291 .. L. A. A. A. 25 . 24 . USA. Junge. (Ed. T. Proceedings of the Fifth International IBPSA Conference. 34 . Proceedings of the CAAD Futures '97 Conference (the 7th International Conference on Computer Aided Architectural Design Futures). Kyoto. A. Winter 2000. Munich. A. Journal of the Illuminating Engineering Society. 62 -71. P. V. Nair. I. Germany. Modeling-Assisted Building Control. (1994). No 4. Winter 1999. Simulation-based Integration of Contextual Forces into Building Systems Control. 25. ASHRAE Transactions 99 (1). "Open" Simulation Environments: A "Preference-Based" Approach. 219 . Mahdavi.40. GESTALT. (1990). S.. Kumar. Chang.United Kingdom. Loftness. Pennsylvania. Zoning. A. Winter 1995.. A. Carnegie Mellon University. S. (1997). Effects of Lighting. Mathew. Solar Energy Vol. pp. Journal of the Illuminating Engineering Society of North America. (2000).. pp.29. Mahdavi. Pittsburgh. pp. Pal. Number 1. Hartkopf. (1993). S. V. 135 . Journal of the Illuminating Engineering Society. 315-322. Mahdavi. Mahdavi. R. An Adaptive Control Using Neural Networks. Japan. Mahdavi. Chang. Neural Networks for Control. Mistry. (1997a). Vol. Number 1. S. (1997b). V. Toward a Simulation-assisted Dynamic Building Control Strategy. I. S. M. (1999). 195 . Proceedings of CAAD Futures '93. Journal of the Illuminating Engineering Society. Mitchell. Pal. No 2. Volume 24. Mahdavi. (1993).294. pp.
Seals. Proceedings of Building Simulation '99. Tse. Overview of Design and Capabilities. Cambridge. Integrated Computational Analysis of the Visual Environment in Buildings. Thrun. L. Vol 28. Ph.. Unified Modeling Language (UML) Specification. W. A comprehensive approach to modeling and evaluating the visual environment in buildings. Mechanical and Electrical Equipment for Buildings. W. M. Neural Networks for Control. New York.D. A.Preliminary Configuration and Validation. V. Rumbaugh. J. L. All-Weather Model for Sky Luminance Distribution . MA. 136 . No 2. Thesis.586. A. (1991). W. OMG: Object Management Group (1997). Ulah. The Role of Exploration in Learning Control. and Adaptive Approaches. Lorenson. B. pp. 579 . Blaha. F. Object-Oriented Modeling and Design. W. Pal. (1993).The MIT Press.(1995). 1. J. R. B.0. P. Japan. Prentice Hall. (1991).. 50. 235-245. Springer-Verlag. Chan. Springer-Verlag... Webos. S.. Tellier. (1991). Fuzzy. Premerlani. Solar Energy Vol. Reynolds. Eddy. R. Van Nostrand Reinhold. Building Automation on the Information Superhighway. II. The MIT Press. The State of the Art in Physically-based Rendering and its Impact on Future Applications. J. No. Photorealistic Rendering in Computer Graphics. So. V. 1993: pp. MA. Sillion. John Wiley & Sons Inc.. Cambridge. Berlin. T. 1991. PA. Pittsburgh. Carnegie Mellon University. Handbook of Intelligent Control: Neural. F. P.. (1996). M. Ver. Mahdavi.. (1999). (1992). Pal.. Bouatouch. Physics Based Lighting Models: Implementation Issues. Lighting Research and Technology. Sixth International IBPSA Conference. NY.. (1992). International Daylighting Measurement Program-Singapore Data I: Quality of data gathered over a long period. Berlin. ASHRAE Transactions: Research. S. Michalsky. Photorealistic Rendering in Computer Graphics. Perez. (1999). (1998). Vol. K. 3. Kyoto. ISBN 4-931416-02-0.
Appendix Appendix A: HISSTO’s test instrumentation Appendix B: HISSTO’s control performance evaluation 1 .
................A10 A.1.......................2.........4.................1..................2.1.....2...........................................................2...................A2 A............4......1............................3......................................1................... Luminaire.............................1............. Pyranometer sensor.................. Louver and luminaire control interface ................................1................................................1 Instrumentation ........A1 A......................1.......2.....2...............................2...........................A3 A... Data quality control and measurement samples................. Data Acquisition Hardware.......................................1...............A12 A..................2.............................................2........................................................1.................................1....................................................................... Actuators ................................................1..........................A6 A..............A7 A..........................................1.........................A8 A.......................... Data acquisition PC board ........................................ Data analysis .........................A11 A... Pyranometer with a shadow ring ....................... Data collection in the Intelligent Workplace .2.......2.................................1..................2.....A1 A....4.................. Sky measurement data filtering criteria......A4 A................. Daylight measurement sample..............................A14 i .......... Transducer monitoring interface..........A4 A..............................1...A13 A...1.........A10 A......... Daylight availability in the IW ....A10 A..............A7 A..........1..1......A1 A.................1..............A13 A............ Examples of simulation vs..............A HISSTO’s test instrumentation Table of Contents A..A5 A1......................3...2..................4....1.........................1................. Daylight station.......... Sensors .....3..................1.............................................2.................2............. measurement ............................... Light Redirection Louver controller .....................................................................1.....................A6 A....................... Photometric sensor .....3..2....... Control and monitoring hardware interfaces ..................A1 A.......................1.............1...........................2..1..............2.. Pittsburgh weather file sample........2.........1.........................................
1. The relative spectral response of the silicon photo-diode is not uniform across the entire solar radiation range.1.1 Photometric Sensor For the system tests.1.15 % per ° C maximum Cosine correction Azimuth Tilt Detector Cosine corrected up to 80 ° angle of incidence ± 1 % error over 360 ° at 45 ° elevation No error included from orientation High stability silicon photo-voltaic detector A. This sensor uses silicon photo-diode to sense solar radiation. A1 . The specification of this sensor is summarized in Table A.4 µ m and increases nearly linear to a maximum at about 0.95 µ m and then decreases almost linear to a cutoff close to 1. The response is very low at 0.1 HISSTO’s test instrumentation Instrumentation Sensors A. Table A.1: Specification of the photometric sensor used for the test Sensor attribute Abstract calibration Sensitivity Linearity Stability Response time Description ± 5 % tractable to NBS 20 µ A per 100 klux 1% maximum deviation up to 100 klux less than 2% change per 1 year 10 µ s Temperature dependence ± 0.2 Pyranometer sensor This sensor is used for measuring solar radiation received from the whole hemisphere.1.A A.1 A.1 shows this sensor’s specification .2. LI-210SATM Photometric Sensor is used for measuring illuminance levels (lux). Table A.2 µ m.1. The LI-210SA Photometric Sensor has been calibrated against a standard lamp and its uncertainty of the calibration is ± 5 %.
the ring axis maintains the angle same as the latitude of the observation site with the horizontal plane.15 % per ° C maximum Cosine correction Azimuth Tilt Detector Cosine corrected up to 80 ° angle of incidence ± 1 % error over 360 ° at 45 ° elevation No error included from orientation High stability silicon photo-voltaic detector A.1.1. ± 5 % maximum.1).Table A. For this purpose. Figure A. Shadow ring is the mechanism to block the direct solar insolation during daytime based on its adjustment. The shadow ring is supposed to be shifted along the ring axis considering the sun’s declination. A2 . On top of that.1 shows the principle of how this device works as intended. the axis of the shadow ring is set to be parallel with the polar-axis. typically ± 3 % 80 µ A per 1000 W m –2 –2 1% maximum deviation up to 3000 W m less than 2% change per 1 year 10 µ s Temperature dependence ± 0. As a result.2: Specification of the pyranometer sensor used for the test Sensor attribute Abstract calibration Sensitivity Linearity Stability Response time Description Calibrated against PSP. the actual sensor should be positioned along the ring axis.3 Pyranometer with a shadow ring This instrument is intended to measure the diffused component of solar irradiance from the hemispherical sky received by a horizontal or an inclined plane (Figure A. The device is constructed based on the KIPP & ZONEN’s CM-121 shadow ring guideline.
total 10 sensors (5 photometric and 5 pyranometer sensors) are placed onto the total five surfaces of a black-painted wooden box surrounded by a hallow rectangular wooden cover for screening ground reflection out.1.1. this simplified version of a sky station has been creating critical data for overall local sky condition throughout the entire data acquisition period (Figure A.1: Principle of the shadow ring blocking the direct solar insolation component The ratio between the ring width and the ring radius is set approximately to 0. or North. Along with the shadow ring. East. To adjust the ring position depending on the solar declination. West.North pole shadow ring solar insolation in June sensor solar insolation in December horizontal Figure A. South.2) A3 .185. A. For both outdoor illuminance and irradiance measurements (horizontal and vertical). Those 5 surfaces are oriented in such a way that a pair of both photometric and pyranometer sensor on each surface faces either upward.4 Daylight station This device is designed and built for monitoring sky condition all the year round. the ring itself has been moving downward or upward at every 3-4 days interval.
Wall bracket is for 2-point fitting and compensation for wall angles of ± 30 . 180 35 H = 100 270 65 90 L = 637 (mm) 0 (degree) Figure A. This luminaire unit has a Vshaped extruded aluminum section with ribbed texture along the sides.2 Actuators A. The luminaire’s dimensions are 275 x 638 x 100 mm.1. The flux of the lamp is 4800 lm having color rendering index 1 B.2..1 Luminaire Lighting fixture in IW (Figure A.2: Daylight station for measuring horizontal and vertical irradiance and illuminance in every orientation A. photometric sensor pyranometer Elevation view Plan view Figure A.3) is a wall-mounted luminaire with high frequency ballast enabling indirect/direct light distribution and has the fitting of two TC-L 55W lamps. This unit is equipped with EVG Zumtobel PC-C ballast having 120W of the connected load and 10 W of the power loss.1. It has also a matt anodized aluminum reflector for asymmetrical lighting pattern.3: Specification of the luminaire installed in the test bay A4 .
1=ascii Meaning If bits 4 and 6 are both set. Argument bytes (such as those following the ‘l’ command) are not echoed. unless a user is attempting to test the device using a dumb terminal. Carnegie Mellon University and has the following characteristics (Table A. Mode 1: Human interface .3): Table A. Communication mode (bit 7) should be 0. Table A. then the louvre is initialized. but not initialized controller serial communication mode.the status byte is returned as a binary string such as "00110010".A. where each bit is represented by the ASCII character ‘1’ or ‘0’. Mode 2: Computer interface .the status byte is returned as the actual byte representing the status value Table A. A5 . If the louvre is not initialized.3: Design characteristics of the louver controller in the IW test bay Attribute Interface Protection Initialization Communication Format Description 9600bps RS-232 serial interface Watchdog timer protection Hardware-controlled initialization All received command bytes are echoed back to verify that the message has been received.2 Light Redirection Louver Controller The light redirection louver controller in IW test bay was constructed by Jamieson Shulte. which can be selected by transmitting the ASCII characters ‘1’ or ‘2’ to the sensor. 0=hex. Communication Modes There are two modes of communication. School of Computer Science.4: Status bit assignment for checking the louver state in the IW Bit 0 1 2 3 4 5 6 7 louvre is fully open louvre is fully closed louvre is moving open louvre is moving closed louvre is running a program (moving to a preset position) unused controller has been reset.2. it cannot execute an absolute movement (using the 'd' command) since the current position will be completely unknown.4 shows the louver’s status bit assignment of the light redirection louver in the IW.1.
5 shows the control command byte bit assignment of the light redirection louver in the IW. 24-bit Digital A6 . and 8 digital I/O lines. This board allows the data acquisition capability up to 100 kS/s and shows 12-bit performance on 16 single-ended analog inputs. 1 byte argument.3. A. 1 byte argument. two 24-bit 20 MHz counter/ timers.6.05 to ± 10 V 2 12 bits 100 kS/s up to ± 10 V 8 2. read destination. MIO-16E-10TM PC board from National InstrumentTM is chosen for a ISA slot on the control host computer. This board features digital triggering capability. Table A.1. read time-out. two 12-bit analog outputs.1. 1 byte returned. Table A. read error threshold.Table A. initialize louver hardware.1 Data acquisition PC board Data acquisition PC boards is to process sensing and actuation signals.6: Specification of the DAQ PC board used for the system test Attributes Analog inputs Resolution Sampling rate Input range Analog outputs Resolution Output rate Output range Digital I/O Counter/Timer Triggers Specification 16 SE/8 DI 12 bits 100 kS/s ± 0. 1 byte returned. load destination.5: Control byte bit assignment for controlling the louver in the IW Signal ’1’ ’2’ ’a’ ’s’ ’x’ ’o’ ’c’ ’p’ Action change to text mode change to byte mode report louver position report interface status clear angle to 0x00 open louver (no program) close louver (no program) stop louver Signal ’l’ ’r’ ’t’ ’e’ ’d’ ’g’ ’i’ Action load time-out. load error threshold. A brief summary of the device’s specification is described in Table A. 1 byte returned. 1 byte argument.3 Data acquisition hardware A.
the louver or each luminaire gets converted analog control signal which dictates its system state.A.4. is passed to the louver motor controller through a RS-232 serial cable. For this. When data acquisition is necessary.1 Louver and luminaire control interface Data acquisition and control command execution operations are implemented in VIsTM (Virtual Instrument) based on the LabviewTM’s G programmingTM language considering its comparability to the various hardware components selected for the system development. accordingly. A7 . the louver controller does status checks during the louver re-positioning process and initializes louver angle after one cycle of data acquisition is completed. The louver control signal.e. Each luminaire is controlled based on a discrete integer value between 0 and 10 for its dimming. 15 degrees) until it hits 105 degrees.4 illustrates the program structure for the louver and the luminaire control in the test bay. The louver state can be set to any of the angles between 0 and 113 degrees except in the automated data acquisition mode where the louver angle is continuously increased in a fixed increment (i.4 Control and monitoring hardware interface A. on the other hand.1. converted into an analog voltage so that the ballast can set the power level. the actual data sampling should be done only after the louver reaches the specified target angle. Figure A. The luminaire control signal is passed to the PC board. Depending on the system control mode.1.
the entire sensor values vector is archived in a file with time stamp attached to it. and displayed depending on the request.2 Transducer monitoring interface Figure A.4: Program structure of the luminaires and louver control VI (Virtual InstrumentTM) A.1. Since both the luminaire and the louver don’t have sensors attached to them.4. On a daily basis. A8 .. a user can specify start point and end point as well as the sampling interval of the data acquisition operation using this interface. logged. When the monitored sensor values are to be logged.5 shows the program structure of the data acquisition process. The program constantly checks the current time in seconds and triggers the defined action when it matches a specific point of the time schedule. Integration of the sampled sensor values over time is done by calculating average of all those values. Figure A. Data sampling rate and buffer size could be customized for each sensor channel. only photometric and pyranometer sensor values are sampled.
Figure A.5: Program structure of the data acquisition VI (Virtual InstrumentTM) A9 .
7: Sky condition data filtering criteria Test on irradiance data I g < 1.05E g E g < 1. recommended as the standard by the US Daylight Availability Working Group and US Bureau of Standards.7).A.2 Data collection in the Intelligent Workplace A.2.2I eeo cos ih I D < I eeo I g > 40Wm I d < 1. all sensors have been continuously monitored and calibrated.1 Sky measurement data filtering criteria To isolate potential errors of in sensor readings. On top of that.15 ( I D cos i h + Id ) I v > ID cos iv I g > I D cos i h –2 Test on illuminance data Eg < 1. .05I g I g < 1.15E D cos ih + E d Ev > ED cos i v Eg > ED cos ih where I g : Global horizontal irradiance ( Wm –2 ) –2 I eeo : Extra-terrestrial solar irradiance ( Wm I D : Direct normal solar irradiance ( Wm I d : Diffuse horizontal irradiance ( Wm –2 ) ) –2 ) –2 I v : Solar irradiance on a vertical surface ( Wm g: ) i h : Solar angle of incidence on the horizontal ( ° ) Global horizontal illuminance ( lx ) Extra-terrestrial solar illuminance ( lx ) Direct normal solar illuminance ( lx ) Diffuse horizontal illuminance ( lx ) Solar illuminance on a vertical surface ( lx ) eeo : D: d: v: A10 . was used to filter the original sky measurement data to avoid inconsistencies (Table A.1 Data quality control and measurement samples A.2.2E eeo cos ih ED < Eeeo Eg > 150lx Ed < 1.1. Class B category tests proposed by Perez. Table A.
p-b1: indoor illuminance values A11 . c: day. m: vertical illuminance (west). l: vertical illuminance(east). h: louver angle.2.2 Daylight Measurement Sample Table A.8: Measurement data sample g 14024 14147 13882 13810 13734 13668 13668 13612 13568 13532 13506 13482 13467 13462 13459 13461 13467 13222 13488 13488 13505 13523 13541 13571 105 41 50 47 44 99 41 49 47 44 5284 5272 95 41 49 47 44 5296 90 41 49 47 44 5307 5900 5916 5929 5952 85 42 49 47 44 5320 5885 84 42 49 47 44 5320 5885 5643 5643 5639 5637 5638 5640 80 41 48 46 44 5206 5803 5525 75 42 49 47 44 5349 5849 5654 5432 5383 5432 5432 5430 5429 5429 5434 70 42 49 47 44 5365 5834 5662 5436 65 42 49 47 44 5382 5820 5670 5438 731 745 764 784 822 822 842 875 914 964 60 42 48 47 44 5402 5812 5681 5442 725 55 42 48 47 44 5426 5798 5694 5445 711 530 541 551 567 592 622 651 651 677 713 762 818 50 43 48 47 44 5456 5788 5710 5452 702 516 45 43 48 48 44 5482 5780 5727 5454 699 509 419 427 439 452 466 487 517 537 562 562 600 637 682 731 40 43 48 48 44 5507 5771 5743 5458 689 505 412 35 43 48 48 44 5533 5765 5759 5464 685 500 407 362 365 374 383 395 410 426 443 473 488 518 518 554 589 629 664 30 43 48 48 44 5554 5757 5776 5471 683 498 403 358 310 313 317 324 334 343 358 368 384 408 413 442 442 469 499 528 553 25 44 48 48 44 5576 5747 5799 5479 685 498 402 354 304 24 44 48 48 44 5576 5747 5799 5479 685 498 402 354 304 270 270 274 278 282 290 299 308 321 331 348 364 370 398 398 422 447 471 491 20 44 48 48 44 5594 5737 5821 5491 694 505 404 355 302 268 15 44 48 49 45 5598 5728 5842 5503 707 513 410 357 300 267 239 240 243 243 248 252 257 263 268 282 296 306 321 332 341 364 364 387 404 424 438 9 44 48 49 45 5601 5722 5864 5516 723 528 420 365 303 268 240 5 45 48 49 45 5725 5763 5950 5609 749 545 435 376 308 271 247 239 234 229 231 235 235 241 244 248 256 264 276 291 300 314 324 332 353 353 374 390 405 420 0 44 48 49 45 5584 5718 5909 5541 783 573 458 396 318 278 252 246 h i j k l m n o p q r s t u v w x y 220 215 210 209 211 213 213 220 223 227 232 242 251 265 275 287 298 308 329 329 350 367 383 394 z 203 199 194 193 196 199 199 205 209 211 220 229 238 250 259 270 279 286 309 309 327 343 361 372 a1 201 196 194 192 194 198 198 203 208 212 221 230 239 252 262 271 279 287 308 308 328 344 360 371 b1 214 208 204 204 205 211 211 219 221 225 236 244 255 267 279 288 300 308 328 328 350 366 383 396 a b c d e f 98 3 3 11 118 109 98 3 3 11 119 110 98 3 3 11 117 108 98 3 3 11 116 107 98 3 3 11 116 106 98 3 3 11 115 106 98 3 3 11 115 106 98 3 3 11 114 105 98 3 3 11 114 105 98 3 3 11 114 105 98 3 3 11 114 105 98 3 3 11 113 105 98 3 3 11 113 105 98 3 3 11 113 105 98 3 3 11 113 105 98 3 3 11 113 105 98 3 3 11 113 105 98 3 3 11 111 102 98 3 3 11 113 105 98 3 3 11 113 105 98 3 3 11 114 105 98 3 3 11 114 105 98 3 3 11 114 105 98 3 3 11 114 105 where a: year. n: vertical illuminance (north). o: vertical illuminance (south). j: vertical irradiance (north). b: month.A.1. e: global horizontal irradiance. d: hour. j: vertical irradiance (west). f: diffuse horizontal irradiance. i: vertical irradiance (east). k: vertical irradiance (south). g: global horizontal illuminance.
4 2.1 107.3 Pittsburgh weather file sample Table A.9 45.4 71.7 82.7 57 63.1 104.7 46.2 180.2 118.2 83.7 102.2 0 0 0 0 0 BTU/hr-ft2 (E-Tot) 0 0 0 0 0 0 18.3 29 2.9: Local weather file sample (Pittsburgh) Month 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 Day 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 Hour 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 BTU/hr-ft2 (S0-Diff.8 0 0 0 0 0 BTU/hr-ft2 (W-Tot) 0 0 0 0 0 0 2.7 97 97.9 48.4 31.9 33. WFC0NN and WFC1NN).7 23.6 62. Table A.7 84.2 6 0 0 0 0 0 A12 .3 27.A.6 36.5 98.5 14.5 82.) 0 0 0 0 0 0 3 23.1 31.8 0 0 0 0 0 BTU/hr-ft2 (S0-Tot) 0 0 0 0 0 0 3.4 103.6 93 160 154 196 126 134 102 52.2 89. new simulation sessions are performed based on the proper portion of this weather file to generate new training patterns for those hybrid predictors requiring update.3 29.9 is a partial sample of Pittsburgh weather file used by the simulator to generate training patterns for the hybrid predictors’ (i.1 24.7 121. When a new month begins or when some change occurs in the building configuration.4 113.4 2.6 135 127.1.6 62.e.9 38.4 124.9 108.8 60.2 211.7 39.2 53.1 121.5 70.3 1.1 14.2.9 0 0 0 0 0 BTU/hr-ft2 (S-Tot) 0 0 0 0 0 0 4.
2. 1998. this result would imply that (particularly due to a large percentage of the glazing area) the test bay has almost no need for ambient electrical lighting with only occasional task lighting.6 shows the percentage of time during which the illuminance levels at the various sensor positions in the test bay were above 250.1 Daylight Availability in the IW The western section of a south bay in the IW has been dedicated to the indoor illuminance monitoring task.6: Percentage of time during which the average indoor illuminance level at each reference point is more than the specified value A13 .2. Continuous illuminance measurement have been performed between Jan. Figure A.81 m above the floor. recent occupancy experiences with the issues pertaining to the demand for solar control and glare control emphasize the need for further studies in this domain. 100 90 80 70 60 % 50 40 30 20 10 0 2 4 6 Senso r 8 10 12 above 250 lx above 500 lx above 1000 lx Figure A. eight standard louver positions have been considered. Considering only the illuminance criterion. 500. An array of 12 illuminance sensors is located along the central axis of this space at the height of 0. if necessary.2. During this measurement period.2 Data anaysis A. or 1000 lux (the graphs are based on the time interval from 9 AM to 5 PM and the data collected between March 3 and March 30). 1997 and Dec. However. This area is separated from the rest of IW with white-colored carton boxes.A.
2 Example of simulation vs. measurement on a cloudy da Figure A.7 shows the examples of the measured and the simulated single reference point indoor illuminance predictions during different times of a clear day and a cloudy day. measurement Figure A.2. measurement on a clear day b) Simulation vs.A. 5000 4500 4000 3500 5000 simulation measurement 4500 4000 3500 simulation measurement Illuminance [lx] 2500 2000 1500 1000 500 0 6 AM Illuminance [lx] 3000 3000 2500 2000 1500 1000 500 0 6 AM 9AM 12PM 3PM 6PM 9AM 12PM 3PM 6PM Time of day Time of day a) Simulation vs.7: Comparison between the measurement and the simulation for the time dependent change of a single indoor illuminance value on a clear day (a) and a cloudy day (b) A14 .2.
HISSTO’s control performance evaluation
Table of Contents
B.1 Preparation of the Predictors.............................................................................B1
B.1.1. Instantiation of the knowledge source ..............................................................................B1 B.1.1.1. Simulation input data..........................................................................................B1 B.1.2. Luminaire Matrix sample..................................................................................................B3
Test of HISSTO as a lighting control system ....................................................B5
B.2.1. Control system performance test with different predictors...............................................B5
HISSTO’s control performance evaluation
Preparation of the predictors
B.1.1 Instantiation of the knowledge source
B.1.1.1 Simulation input data Table B.1-Table B.5 show a subset of the input values to the LUMINA simulation program used for a series of experiments performed in the test bay at the Intelligent Workplace. Describing both internal and external surfaces and windows along with their physical properties was the main part of inputs to LUMINA. Considering the geometrical sophistication of the test bay configuration, certain level of simplification has been applied to those non-critical building components to reduce the complexity and the time needed to finish necessary simulations. Reference points, site information, viewer information (for glare calculations), and luminaire information were added to the inputs. All dimensions describing the locations and sizes of the building components and systems are based on the metric scale. Light redirection louvers were described as the dynamic external surfaces which change their states in operation time. Part of the inputs to LUMINA came from the measurement done with the daylight station on the roof-top of the IW. .
Table B.1: Properties of the internal surfaces in the test bay
a 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 b 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 c 1.8 3.8 3.8 0 0 0.4 0 0.4 0 0.4 0.4 3.8 2.8 0 0 0 0.5 0.5 0.5 d .8 4.3 4.3 -0.3 2.7 0 -0.3 4.3 4.6 0 4.3 0 1.5 -0.3 0 1.6 -0.3 4.6 -0.3 e 0.9 0 3.1 0 0 0 0 0 0 3.1 3.1 3.1 4.7 3.1 0 0.8 2.4 2.4 2.4 f 1.4 3.8 3.8 0 0 0.4 0 3.8 0.4 0.4 3.8 0.4 3.8 0 3.8 3.8 0.4 0.4 0.5 g 0.8 0 0 1.6 4.5 0 -0.3 4.3 4.6 4.3 4.3 0 1.5 4.6 0 1.6 -0.3 4.6 4.6 h 0.9 0 3.1 0 0 3.1 3.1 0 0 3.1 3.1 3.1 4.7 3.1 0 0.8 2.4 2.4 2.4 i 1.4 3.8 3.8 0 0 3.8 0.4 3.8 0.4 2.8 3.8 2.8 3.8 0.4 3.8 3.8 0.4 0.4 0.5 j 0.8 0 0.9 1.6 4.5 0 -0.3 4.3 4.6 1.5 1.5 0.9 0.9 4.6 4.3 2.7 4.6 -0.3 4.6 k 1.2 3.1 4.7 0.9 0.9 3.1 3.1 3.1 3.1 4.7 4.7 4.7 4.7 3.1 0 0.8 2.4 2.4 2.6 l 1.8 3.8 3.8 0 0 3.8 0.4 0.4 0 2.8 2.8 3.8 2.8 0.4 0 0 0.5 0.5 0.5 m 0.8 4.3 0.9 -0.3 2.7 0 -0.3 4.3 4.6 0.9 1.5 0.9 0.9 -0.3 4.3 2.7 4.6 -0.3 -0.3 n 1.2 3.1 4.7 0.9 0.9 0 0 3.1 3.1 4.7 4.7 4.7 4.7 3.1 0 0.8 2.4 2.4 2.7 o 0.8 0.7 0.7 0.7 0.7 0.7 0.7 0.7 0.7 0.7 0.7 0.7 0.7 0.4 0.3 0.3 .7 0.7 0.7 p 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 q 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 r 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 s 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 t 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 u 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 v 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Table B.2: Properties of the external surfaces outside of the test bay
a 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 b 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 -0.88 -1.30 -0.88 -1.30 -0.88 -1.30 0 -0.2 -0.3 -0.3 c d 4.6 0.3 1.4 2.9 4.0 4.6 1.6 4.6 -2 -2 -2 -2 -2 -2 -2 -0.3 -2 -2 e 2.2 0.9 0.9 0 0.9 0.9 0 0 2.69 2.69 2.25 2.25 1.80 1.80 3.1 2.9 3.1 3.1 0 0 0 0 0 0 0 0 -0.88 -1.30 -0.88 -1.30 -0.88 -1.30 -0.3 -0.2 -0.3 -1.1 f g -0.3 -0.3 0.3 1.4 2.9 4.0 -0.3 2.7 6 6 6 6 6 6 -2 4.6 6 -2 h 2.2 0.9 0.9 0 0.9 0.9 0 0 2.69 2.69 2.25 2.25 1.80 1.80 3.1 2.9 3.1 3.1 0 0 0 0 0 0 0 0 -1.30 -0.88 -1.30 -0.88 -1.30 -0.88 -0.3 -0.2 -1.1 -1.1 i j -0.3 -0.3 0.3 1.4 2.9 4.0 -0.3 2.7 6 6 6 6 6 6 6 4.6 6 6 k 3.1 2.2 2.2 2.2 2.2 2.2 0.9 0.9 2.69 2.69 2.25 2.25 1.80 1.80 3.1 3.1 3.1 3.1 0 0 0 0 0 0 0 0 -1.30 -0.88 -1.30 -0.88 -1.30 -0.88 0 -0.2 -1.1 -0.3 l m 4.6 0.3 1.4 2.9 4.0 4.6 1.6 4.6 -2 -2 -2 -2 -2 -2 6 -0.3 -2 6 n 3.1 2.2 2.2 2.2 2.2 2.2 0.9 0.9 2.69 2.69 2.25 2.25 1.80 1.80 3.1 3.1 3.1 3.1 o 0.13 0.13 0.12 0.13 0.12 0.13 0.7 0.7 0.24 0.46 0.24 0.46 0.24 0.46 0.6 0.6 0 0.6 p 0 0 0 0 0 0 0 0 0.1 0.1 0.1 0.1 0.1 0.1 0 0 0.4 0 q 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 r 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 s 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 t 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 u 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 v 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
where a: number of vertices; b: number of holes; c: x coordinate of the 1st vertex; d: y coordinate of the 1st vertex; e: z coordinate of the 1st vertex; f: x coordinate of the 2nd vertex; g: y coordinate of the 2nd vertex; h: z coordinate of the 2nd vertex; i: x coordinate of the 3rd vertex; j: y coordinate of the 3rd vertex; k: z coordinate of the 3rd vertex; l: x coordinate of the 4th vertex; m: y coordinate of the 4th vertex; n: z coordinate of the 4th vertex; o: reflectance: p: transmittance; q: diffuseness of reflectance; r: specularity of reflectance; s: phone exponent for specular reflectance; t: diffuseness of transmittance; u: specularity of transmittance; v: phong exponent for specular transmittance
Table B.3: Properties of the windows installed on the western facade of the test bay
a 0 0 0 0 0 0 b -0.3 -0.3 0.3 1.4 2.9 4.0 c 2.2 0.9 0.9 0 0.9 0.9 0 0 0 0 0 0 d e 4.6 0.3 1.4 2.9 4.0 4.6 f 2.2 0.9 0.9 0 0.9 0.9 0 0 0 0 0 0 g h 4.6 0.3 1.4 2.9 4.0 4.6 3.1 2.2 2.2 2.2 2.2 2.2 i 0 0 0 0 0 0 j k -0.3 -0.3 0.3 2.9 2.9 4.0 l 3.1 2.2 2.2 2.2 2.2 2.2 m 0.7 0.62 0.62 0.7 0.62 0.7 n 0.84 0.61 0.71 0.5 0.71 0.61 o 0.13 0.13 0.12 0.13 0.12 0.13
where a: x coordinate of the 1st vertex; b: y coordinate of the 1st vertex; c: z coordinate of the 1st vertex; d: x coordinate of the 2nd vertex; e: y coordinate of the 2nd vertex; f: z coordinate of the 2nd vertex; g: x coordinate of the 3rd vertex; h: y coordinate of the 3rd vertex; i: z coordinate of the 3rd vertex; j: x coor-
5: View point specification for glare calculation a 1. e: ground reflectance.2 1. f: building azimuth Table B. b: y coordinate of the eye location.4: Site properties of Pittsburgh a 72 b 80. d: height above sea level. c: latitude. k: y coordinate of the 4th vertex. b: longitude of time meridian.3 . l: z coordinate of the 4th vertex.6 b 1. o: reflectance Table B.22 c 40.5 d 373. B3 .6 shows a sample of the Luminaire Matrix used for electrical lighting calculation in HISSTO. d: altitude of the line of sight. c: z coordinate of the eye location.1 c 0 d e 180 where a: x coordinate of the eye location.2 Luminaire Matrix sample Table B. e: azimuth of the line of sight B.1. m: transmittance.dinate of the 4th vertex.2 e f 283 where a: longitude. n: frame factor.
MOi: total solid angle (modified) subtended by the window at the eye. Lb: background luminance. Ei: illuminance value of the ith reference point.Table B.6: Luminaire Matrix sample where. Ei: indirect illuminance on the vertical surface of the eye. Lmax: luminance value of the patch on the computer screen having maximum luminance B4 . Ed: direct illuminance on the vertical surface of the eye. Lsi: luminance of the ith glare source (luminaire) in the field of view. Li: dimming level (0 -10) of the ith luminaire in the test bay.
7: Initial systems states for testing HISSTO’s various controllers Year 1998 Month 3 Day 14 Hour 9 Iglob 81 Idiff 78 Eglob 9379 Louver 0 L1[%] 30 L2[%] 30 L3[%] 30 L4[%] 30 There were two sub-test categories: The first sub-test category establishes the rating of the different louver angles based on their predicted daylight performances.15. it could not pick up the angle more than 15 degrees as the desirable louver state unlike the other predictors.B.7. Table B. multiple instances of the daylight responsive lighting control scenario were tested with the same initial systems states illustrated in Table B. Most predictors suggested similar system control states for the louver and the luminaires in the test bay. The second sub-test focuses on the 10 best control actions for both the louver angle and the luminaire dimming levels suggested by each predictor. Since pure simulation method has limited louver state search range bounded between one step below and above the initial state. B5 . The result of this test is illustrated through Table B.2 Test of HISSTO as a lighting control system B.8 to Table B.2.1 Control system performance test with different predictors Based on the predictions made by the different predictors.
Table B.801768 0.814496 0.821825 0.694854 0.952111 0.667 418.440089 Table B.004 221.766531 0.818703 0.22415 2.813567 0.766502 0.74782 6.25197 2.8 29 40.6162 6.472181 0.27312 2.796577 0.18997 2.953007 0.8: Order of the best louver positions suggested by the SMC0NN predictor θn+1 (lvr)[degree] 105 90 75 60 45 30 15 0 Em[lx] 287.812854 0.89488 CGI 0 0 0 0 0 0 0 0 GCRT 0.48 268.417 417 416.82349 6.8 29 0.952648 0.631 UE 0.694548 0.694345 0.69578 5.954444 Q[W] 2.83727 0.462685 0.471109 0.694345 0.796135 DGI CGI GCRT 0.800836 0.806306 0.951931 0.765359 0.72392 5.66999 P[W] 0 0 0 0 0 0 0 0 U 0.14646 2.9: 10 best control actions suggested by the SMC0NN predictor θn+1 (lvr)[degree] 105 105 105 105 105 105 105 105 105 105 L1[%] L2[%] L3[%] L4[%] Em[lx] 432.917 402.99775 1.341 258.457668 0.765007 B6 .917 446.766144 0.765323 0.71686 6.75212 5.6 46.333 431.951931 0.804694 0.72392 5.80374 0.694447 0.8 34.805872 0.765587 0.85802 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 40.5 UE 0.6 34.766096 0.765133 0.95229 0.72034 6.812422 0.25 431.09116 1.843112 Q[W] P[W] U 20 20 20 20 10 10 10 10 20 20 20 10 10 20 20 10 20 10 20 10 20 20 20 20 20 20 20 20 10 10 10 10 20 20 10 10 20 20 10 10 6.465678 0.78046 5.94 285.95229 0.643 282.694243 0.804283 0.6 34.333 417.695669 DGI 5.823905 0.795618 0.61357 6.469647 0.81773 0.51194 6.69465 0.61017 6.765705 0.467803 0.71416 6.8 40.879 274.719 278.4 34.69578 5.75 403.70983 5.
817124 0.333 411.09928 5.65934 6.774168 0.808351 0.98815 7.55665 6.938526 0.808987 DGI CGI GCRT 0.934378 0.957695 Q[W] 2.818311 0.262 282.8 40.825616 0.77449 0.466285 0.472897 0.842 UE 0.475736 0.76727 6.083 UE 0.930783 0.82446 0.77453 0.703352 0.48 265.10: Order of the best louver positions suggested by the WFC0NN predictor θn+1 (lvr)[degree] 105 90 75 60 45 0 30 15 Em[lx] 297.816726 0.25 440.583 440 425.797103 0.462334 0.24498 2.802397 0.31251 2.801449 0.774128 0.36755 2.16976 1.826647 0.457648 0.6815 5.774936 0.77398 0.949966 0.702941 0.79505 6.781553 0.719 274.816065 0.703147 0.8 29 34.5 411.4 34.774814 0.87156 6.820447 0.786238 0.8 40.703558 0.019 290.8 29 0.727 239.825686 Q[W] P[W] U 10 20 10 20 20 10 10 20 20 10 20 20 20 20 20 20 20 20 10 10 10 10 20 20 10 10 20 20 10 10 6.5 454.927389 0.3377 7.453133 Table B.86992 2.833 441.917 426.775302 0.457793 0.662 6.07204 1.8336 0.703764 0.6 34.949254 0.90648 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 34.42033 2.71556 0.6 46.Table B.6 40.702941 0.15554 6.94126 P[W] 0 0 0 0 0 0 0 0 U 0.11: 10 best control actions suggested by the WFC0NN predictor L1[%] θn+1 (lvr)[degree] 105 105 105 105 105 105 105 105 105 105 20 20 20 20 10 10 10 10 20 20 L2[%] L3[%] L4[%] Em[lx] 426.597 7.16993 6.037 244.917 425.361 253.943769 0.469735 0.773935 B7 .703558 DGI 8.790258 0.39348 CGI 0 0 0 0 0 0 0 0 GCRT 0.65591 6.775349 0.785329 0.76375 6.807031 0.76102 6.
792796 0.96261 0.8 29 34.167 395.02701 4.792499 0.805083 0.417 UE 0.85054 1.792305 B8 .456912 Table B.793612 0.43301 CGI 0 0 0 0 0 0 0 0 GCRT 0.333 379.6 40.994012 0.79774 0.833 365.792341 0.72 242.462857 0.982542 1.812938 0.Table B.464592 0.466608 0.5377 5.996717 0.793277 0.792073 0.13: 10 best control actions suggested by the SMC12NN predictor θn+1 (lvr)[degree] 105 105 105 105 105 105 105 105 105 105 L1[%] L2[%] L3[%] L4[%] Em[lx] 381.792826 0.87551 1.837503 Q[W] P[W] U 20 02 20 20 10 10 10 10 20 20 10 20 10 20 20 10 10 20 20 10 20 20 20 20 20 20 20 20 10 10 10 10 20 20 10 10 20 20 10 10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 34.724 206.814181 0.00078 1.57059 5.803172 0.6 46.692283 0.82105 0.986152 0.075 226.822253 0.504403 0.25 380.793116 0.69242 4.8152 1.748 188.970907 0.811197 0.812459 0.12: Order of the best louver positions suggested by the SMC12NN predictor θn+1 (lvr)[degree] 105 90 75 45 15 30 60 0 Em[lx] 251.823395 0.689732 0.694691 0.73433 P[W] 0 0 0 0 0 0 0 0 U 0.793079 0.00177 Q[W] 2.82598 1.831319 0.698065 0.999 162.810143 0.2603 4.861 UE 0.43301 4.667 366.787722 0.831026 0.91807 1.481992 0.6 34.4 34.25 379.917 394.833 409.68146 0.669116 0.957514 0.475341 0.034 175.8 40.792434 0.793685 0.97074 1.583 394.8 40.803841 DGI CGI GCRT 0.69062 DGI 5.495861 0.44754 5.811761 0.700491 0.546 168.816627 0.8 29 0.
128317 0.132401 0.169139 0.798804 0.127942 B9 .183285 0.463 95.439186 0.796171 0.42798 0.133068 0.083 265.795959 0.15: 10 best control actions suggested by the MS1NN predictor θn+1 (lvr)[degree] 105 105 105 105 105 105 105 105 105 105 L1[%] L2[%] L3[%] L4[%] Em[lx] 275.795107 0.126169 0.8041 87.128109 0.463 109.2 110.75 265.800104 0.436882 0.7427 x x x x x x x x UE DGI x x x x x x x x CGI x x x x x x x x x x x x x x x x GCRT Q[W] x x x x x x x x P[W] x x x x x x x x U 0.18 101.4 0.583 256.132902 0.425766 0.833 UE 0.132693 0.433364 0.25 266.128192 0.79202 0.2 110.146519 0.2 104.949 106.4 104.797484 0.75 257.167 255.177002 0.4 104.8937 75.137653 0.793098 0.25 256.413354 Q[W] P[W] U 50 50 50 40 50 50 50 50 40 40 50 50 40 50 50 50 40 30 50 40 50 50 50 50 40 50 50 50 50 50 50 40 50 50 50 30 40 50 40 50 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 116 110.420493 0.412473 0.4 104.14: Order of the best louver positions suggested by the MS1NN predictor θn+1 (lvr)[degree] 105 90 75 60 45 30 15 0 Em[lx] 113.189143 0.128651 0.42135 0.333 256.0679181 Table B.6861 40.2 110.794446 DGI CGI GCRT 0.4 104.333 264.Table B.159705 0.426661 0.795998 0.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue listening from where you left off, or restart the preview.