You are on page 1of 5

Instrumentation and Measurement Technology Conference - IMTC 2007 Warsaw, Poland, May 1-3, 2007

Data Acquisition and Sharing for a Distributed Intelligent DSM System


DIASS - Politecnico di Bari - V.le del Turismo 8, 74100 Taranto - Italy [email: andriagmisure.poliba.it,dileccegpoliba.it]

Gregorio Andria and Vincenzo Di Lecce

Abstract - Artificial intelligence has devised solutions for a wide class of scheduling problems that can be applied also to power management and in general to building-automation. This paper presents one of these methods, based on the Multi Agent System (MAS) paradigm for a demand-side management system.
I. INTRODUCTION

Energy management is not a new application. Several researches, both for industrial and domestic purposes, have been carried out on this topic from building automation to ship application focusing on different aspects of the issue. This paper proposes the analysis of artificial intelligence application to power load management. Power load management (or load management for short) consists in controlling loads at the customer side in order to manage energy systems more efficiently. Considering the fact that energy systems are very common and heavily distributed (typically including millions of loads in an area covering counties or countries). In general a Demand-Side Management (DSM) System refers to measures implemented by utilities that modify end-use electrical energy consumption, either reducing overall consumption through energy efficiency or using load management to reduce demand at times when the cost of reducing demand is less than the cost of servicing it [4]. Load management programs try to achieve an approximately constant energy consumption that meets the characteristics of power stations. It is well known that the typical energy consumption is often very far from meeting this requirement. In industrial usage the power consumption is related to the specific activity, in domestic usage the power consumption is typically related to the number of persons present. To cover this gap, load management systems involve reducing loads on a utility's system during periods of peak power consumption or allowing customers to reduce electricity use in response to price signals. Such systems use policy like interruptible load tariffs, timeof-use rates, real-time pricing, direct load control, and voluntary demand response programs. In the past many manufacturing industries had proposed simple DSM-solutions based on current sensor and programmable timers, but they cannot be used for complex problems and they are essentially static [6]. The power management system can be already used in the case of power co-generation. Different types of energy can be produced to satisfy different user
1-4244-0589-0/07/$20.00 2007 IEEE

requirements, for example thermal energy from solar panels, low voltage electric in c.c. from photovoltaic equipments and industrial voltage in a.c. from wind power. This study proposes a model of an intelligent short term DSM system based on a multi agent system (MAS) [1] as in Figurel . The proposed system is based on autonomous units composed of a current measurement system and a processing node. The current measurement system is based on a Halleffect current-to-voltage converter. The process element, based on a small industrial microprocessor, reads the local current information, "sees" the general and all the local load conditions, and finally proposes its optimal solution. Then, through progressive brokerage acts, the DSM system finds the final solution (loads to supply and loads to suspend).

Ix

EC2 bus
,

system

tt

Figure 1: electric wiring and DSM integration

In order to effectively evaluate and design decentralized energy controllers, several processes have to be implemented and integrated (i.e. data logging, digital signal processing of gathered raw data,

ELECTRICAL NETWORK

LCDAD CURRENT /k k-th


vout

ADO

Figure 2 - Block Diagram of Current Measurement System

implementing and evaluating physical device simulations, design and evaluation of several types of controller topologies, etc.). The aim of this work is to evaluate the effects of a decentralized measurement system finalized to inform a M\AS of the local and global load condition and to study the behavior of the M\AS while it is handling critical conditions. For this reason, a simulation was realized by means of Matlab Simulinkg. II.
CURRENT MEASUREMENT SUB-SYSTEM

In order to conveniently reach the fixed objectives, according to the conditions described above, each significant load of the network under control has to be simply characterized. Then, all these loads have to be easily identified and the instantaneous values of their relevant load currents have to be correctly measured and converted real-time, into a digital input (in a suitable format) to the DMS system without losing information. These requirements can be achieved successfully by following these steps: i) providing the load plug of a specific LIC (Load Identification Code); ii) measuring the load current by a suitable Hall-effect sensor located into the same plug-point; iii) converting the voltage signal returned by the sensor in the numerical data MCD (Measured Current Data), by a suitable A/D converter; iv) finally, transmitting the total string (LIC & MCD) to DMS system and to the other load points using EC2like protocol. A diagram-block scheme illustrating all the suggested procedure is shown in Fig.2. Subsequently, some details of the above listed steps are reported. The LIC relevant to the k-th load plug is assigned in the phase of network planning and DMS design. It assumes generally the format of a 8-bit digital word, that is transmitted continuously from a suitable transmission block ("k-th LIC") to a sum down block. Obviously, this code can be easily modified, owing to particular requirements of both users and control systems. In this case, the network planning needs reset procedure. The load current at the k-th plug-point is real-time

measured by a suitable Hall-effect sensor integrated into the same plug. The sensor is also complete with the necessary down electronic circuitry of amplification and filtering ("conditioning block") for the measured analog signal and - a suitable 8-bit flash A/D converter performing its conversion to another 8-bit digital word. Also this word (carrying information of the instantaneous current value) is transmitted from a suitable transmission block ("k-th MCD") to the abovementioned summation down block. Most recent linear Hall ICs provide a ratiometric output voltage. The quiescent (i.e., null) voltage is (nominally) 5000 of the applied, stable supply, as shown in Fig.3. This quiescent output voltage signal equates to no applied magnetic field and, for current sensing, is equivalent to zero current flow.

01
0

Oi UTPUT

VOLTAGE

CL
SATUPAT 0O

MAGNETIC FIELD (GAUSS)

Fig. 3 - A Linear Hall Sensor Transfer Curve

A south polarity field induces a positive voltage transition (toward Vcc), and a north polarity results in a transition toward ground (0 V). Output saturation voltages are (typically) 0.3 V (high/sourcing) and 0.2 V (low/sinking) and are measured at 1 mA. Each linear Hall-effect IC integrates a sensitive Hall element (also called a 'plate'), a low-noise (bipolar) amplifier, and sink/source output stage [5]. Any systems problems associated with low-level signals and noise are minimized by the monolithic integration of magnetic sensor, amplifier, output, and allied signal processing circuitry. The existing very stable, linear silicon HEDs (Hall Effect Devices) exploit dynamic quadrature offset cancellation circuitry and utilize electronic switching to change the current path in the sensor element. In the
2

The output of these sensors is independent from the as shown in Fig.4. Then, it can be outlined that in a wide frequency range, the device behaviour can be considered linear and characterized also by a limited

applications. frequency,

Table I, some characteristics of the most used Hall-effect sensors are reported. As it is easily to be viewed, flexible sensors for AC curr ent are very suitable for DMS

Currenit measuremendt

error phase.
1.2577 X

~~~Phase Degree
k k~~~~ W

I 4

2
0-

uncertainty050 < 3to < 7gs Response time

Linearity

Table I - Characteristics of more used Hall-effect sensors Open-loop Closd-loop Flexble sensor for AC current sensors sensors Measuring 0-18 kA 0-15 kA 0-60 kA range 8 Hz-100 kHz Bandwidth 0-25 kHz 0-200 kHz 10 100 Typ. accuracy 0.50 at 250C
.0.60

%06%

temperature

Operating

-20o00

25t

7'C

<1 g~s -40 to 85

< 5Og~s

-20 to 85 'C

1.00 ii
0.9

Once the two k-th components LIC and MCD are added, the total 16-bit k-th string is transmitted to DMS system.
Output

-84
400

8 10

40

100

1k

4k

10k

40k

100k

LIC

MCD

Frequency (Hz)

Fig. 4 - Dynamic behavior of flexible sensors for AC current

Measurement uncertainty of these devices depends generally on both electrical parameters, and ambient or operating temperature. To evaluate uncertainty due to-ambient temperature, the following factors are particularly significant: - offset voltage; - loop gain; - linearity. Moreover, other important factors, to evaluate the increasing measurement uncertainty depending on operative temperature of the device, are the variations of both offset and gain.
.

The same string is also transmitted to the other "intelligent" load points using Ec2-like protocol. Now the curr ent information at each load-plug is ready to be processed, as illustrated in the following section.

III. PROCESSING NODE


This work proposes a DSM solution based on a 1\4AS. The system is based on three different functional blocks arranged to obtain two main functions (see fig.3) in a distributed network: 1. curr ent sensor (with or without) clock and timing generator. This function is realized using the sensor described in section II, a 16 bit A/D converter, a small memory (one word for each load, each word of 4096 cells, 2 bytes in length, (curr ent sampled

----value at the

current time

wDC
each ed 5 each

~current sensor

~Hall effect

Loa current

If a neiw load is supplied, all the value are set the last sampled Ieff If the Iioad is known the new value in LM in corrniputed as mean value between new samplies I,f and stored

Figure 5 - Block Diagram of Processing node

each 5 seconds) called Loads Memory (LM), and a Ec2-like interface. 2. processing node (PN), hosting one independent agent; it is able to manage a single electric load allowing or disallowing it to supply power. Each intelligent agent knows the load curve of the appliance connected to it and those of the active appliances connected to all the other current sensors [7]. At each instant of time, the load curves of all the active appliances are stored in LM and replayed using Ec2 protocol on all the LM present in the system. This situation can be considered similar to the communication room in [8]. A fundamental constrain for this system is that not all the devices have to react at each event, but as many as there are necessary to avoid a critical condition. A key point is how the system chooses what devices have to be switched off. This choice has to achieve two main targets: 1. the overall power request has to be less than the critical load. On the other hand, the selected choice should have an overall power request as close as possible to this limit. 2. giving to user the maximum power usability. This means that the system should satisfy as many user requests as possible. In order to find the load configuration that achieves these targets, each agent evaluates different strategies.
IV.

greater than 3 kW. This situation is the most limitative and stressful for the proposed system. The used configuration is: * Number of soket/PN equal to five * Each PN handles a single socket tap * On each socket tap only a household appliance at time can be active * T ( observing time) is equal to twenty*5s (100 seconds) * Base load variable between 0 and 1.05 kW as shown in Figure 6. For a residential consumer this base load could not be constant in the time [3]. For example, just thinking of a house with only the refrigerator (that is a variable load) and some other small appliances on (lamps, alarm, appliances in stand-by and so on). In the proposed model, the base load is modeled using an uninterruptible variable power request.

F0 n
800W

1000\aooA t0X

a!
IL

50ow

TESTS AND EVALUTIONS

In order to test the system, a lot of experiments were carried out using different loads. The Italian provider offers contracts that allow domestic consumers to use electric power till a limit of 3 kW. If a consumer exceeds this limit, the outcome is an outage, that will last until the user's power request is

12

time(h)

24

Figure 6 - Example of daily profile

3kW

uninterrupted washing nachine cye


1 .5kW
.

water heating
water pump and spin-dryer wasi'ng m tor

[t.

3 kW

i .. =
.. .. ...........
..

DSM eftecton

ad

1 .5kW . .......

.......n[.

Lnu.

consent to load fee (load over i 50 watts)

OF

.o

120'

180,

240' time

Fig. 7 typical load diagram of a household application with and without DSM
-

For the proposed evaluation different household appliances (washing machine, dishwasher, vacuum cleaner, tumble dryer, iron and steam iron, oven) have been characterized considering different working conditions. The acquired data are used to train the system and to evaluate the behavior of the system. Figure 7 shows the effect of the proposed DSM on a washing machine. The top plot of figure7 shows the load diagram of a washing machine without DSM. The whole washing cycle takes about 2 hours. The used power is up to 2.6 kW, close to the maximum before shutdown. If the devices is switched-on at 11.00 a.m. the sum of the base load and the device load will overcome the 3kW threshold. For this reason the DSM temporally stops the power supply to the device according to the choice policy.
V. CONCLUSION

REFERENCES
[1] Nadel, S. 1992 ."Utility Demand-Side Management Experience and Potential - A Critical Review." Annual Review of Energy and the Environment, 17:507-35. Palo Alto, CA: Annual Reviews, Incentive. [2] INS. Wooldridge, M. and Jennings, N.R., editors. (1995). Intelligent Agents - Theories, Architectures and Languages. Lecture Notes on Artificial Intelligence, volume 890, Springer-Verlag: Berlin. [3] Joanicjusz Nazarko, Zbigniew A. Styczynski, "Application Of Statistical And Neural Approaches To The Daily Load Profiles Modelling In Power Distribution Systems", IEEE ISBN:0-7803-55156/99 1999. [4] Leonard, M. "Networked Controllers talk over power lines", in Electronic Design, pp. 73-79, 17 September 1992. [5] LonWorks Technology Device Data, Motorola Inc., USA 1995 [6] Nadel, S. 1992. "Utility Demand-Side Management Experience and Potential - A Critical Review." Annual Review of Energy and the Environment, 17:507-35. Palo Alto, CA: Annual Reviews, Incentive. [7] V. Di Lecce, C. Pasquale and V. Piuri, "A Basic Ontology for Multi Agent System Communication in an Environmental Monitoring System", CIMSA 2004, Boston, MA, USA, 14-16 JULY 2004 [8] V. Di Lecce I. Guarino, A. Guerriero, C. Pasquale: "Multi-Agent Approach in Environmental Application:An Evaluation", Second IEEE on ADVANCED International Workshop ENVIRONMENTAL SENSING AND MONITORING TECHNOLOGIES - Villa Olmo, Como, Italy July 24 - 25, 2003 [9] A.Amato, M. Calabrese, V. Di Lecce and V. Piuri "An Intelligent System for Decentralized Load Management". CIMSA 2006 - IEEE International Conference on Computational Intelligence for Measurement Systems And Applications 12-14 July 2006 La Corufia- Spain

This work proposes a model of an intelligent shortterm DSM system based on a MAS. Considering a physical implementation of the proposed system, it seems that there are no significant technology problems. Small and cheap processors are available and the technologies enabling information interchange among the agents are accessible and well tested [9]. The Hall-effect transducer is very small and gprocessor compatible. Simulations are performed using both Matlab and Labview in order to evaluate system performance and decision strategy. The main advantages of this system in comparison with a system based on the simple user defined priority
are:

appliance. Nowadays load curves, sampled using real household appliances, are used as input to test the system. Since one household appliance, such as a washing machine, can have more than one load curve, future works will be about the ability of sampling and storing many load curves to each agent. In this way the agent can try to recognize an ongoing power request and it can associate a stored load curve with it. Therefore, the table of loads will be dynamically and automatically built.

* It has a great flexibility, indeed user can define a set of devices as uninterruptible loads. * It gives a big power reliability because it is able to avoid the critical conditions. * Its behavior is not predefined by the user, but it is dynamically determined by the competition among various agents. * The proposed consumption strategies aim to fulfill both user's whishes (defined by means of the user defined priority) and the intrinsic constrains introduced by the state of work of each household