Professional Documents
Culture Documents
com/science/article/pii/S0360132317301920
Manuscript_8057ba49cb0f46ea9a419304428146d2
7 Abstract
8 Studies have shown that managing the setpoints and schedules of rooftop units (RTUs; air-conditioners
9 and heat pumps) will result in up to 20% energy and cost savings. Another problem associated with RTUs
10 is short cycling, i.e., when an RTU goes through ON and OFF cycles too frequently. Excessive cycling
11 can result in decreased average efficiency (up to 10%), even if there are no physical failures in the
12 equipment. Many small- and medium-size commercial buildings (SMBs) use rudimentary controls that
13 are mostly manual and have limited scheduling capability and no monitoring or failure management.
14 Therefore, many of these buildings are operated inefficiently and unnecessarily consume excess energy.
15 SMBs typically use packaged RTUs that are controlled by individual thermostats. Ensuring the correct
16 use of the zone setpoint and eliminating too frequent cycling of RTUs, thereby leading to persistent
17 building operations, can significantly increase the operational efficiency of the SMBs. The work reported
18 in this paper describes two algorithms for detecting the zone setpoint temperature and RTU cycling rate
19 that can be deployed on low-cost infrastructure. These algorithms only require zone temperature data for
20 detection without any additional sensor installation or supervised learning. The algorithms have been
21 tested and validated using field data from 24 RTUs from six buildings in different climate locations.
22 Overall, the algorithms were successful in accurately detecting the setpoints and ON/OFF cycles.
23
24 Keywords: Rooftop unit, Setpoint, Cycling, Zone temperature, Signal processing, and Peak detection.
∗
Corresponding author: Tel.: +1 509 372 4281; E-mail: Srinivas.Katipamula@pnnl.gov.
© 2017 published by Elsevier. This manuscript is made available under the Elsevier user license
https://www.elsevier.com/open-access/userlicense/1.0/
25
26 Nomenclature
[ea
a number of actual ON/OFF cycles
]
[ea
a number of ON/OFF cycles perdition
]
[%
mean absolute percent error
]
n integer
[ea
n a number of predictions
]
R region
[°
s standard deviation of prediction
C]
[°
Tn normalized temperature
C]
[°
Tmax local maximum temperature
C]
[°
Tmin local minimum temperature
C]
2
[°
Tpeak array of maximum temperature
C]
[°
Toa outdoor air temperature
C]
[°
Tsp setpoint temperature
C]
[°
Tvalley array of local minimum temperature
C]
t time series
x discrete-time signal
̅
[°
mean of predicted setpoint temperature
C]
z complex number
27 Subscripts
1 current
2 new
avg average
i ith element
oa outdoor air
3
pre prediction
28 Greek
ω class
29
30 1. Introduction
31 Several studies have documented that commercial buildings consume between 10% and 30% excess
33 heating, ventilation, and air-conditioning (HVAC) system operations can lead to significant reductions in
34 energy use and carbon dioxide emissions. Small- and medium-size (<100,000 sf) commercial buildings
35 (SMBs) in the United States (U.S.) represent about 65% of total commercial building area and consume
36 about 60% of the total site electricity associated with the U.S. commercial building stock [6]. These
37 buildings employ rudimentary controls that are mostly manual, have limited scheduling capability, no
38 monitoring or failure management, and generally do not have a dedicated building operator or an energy
39 manager. Therefore, the HVAC equipment tends to be serviced as a result of occupant complaint or when
40 the units fail [7]. Even the buildings that get periodic maintenance have a number of operational problems
41 that go undetected. The operational problems result from improper control or incorrect commissioning,
42 which leads to inefficient operation, increased energy use, and reduced equipment life [8–10].
43 SMBs typically use packaged rooftop units (RTUs) that are controlled by an individual thermostats
44 and have limited ability to monitor or trend the data necessary to detect system degradation or to perform
45 supervisory controls [7]. These RTUs are often poorly maintained, and degradation of performance and
46 faults are only addressed when occupants complain or a unit fails [11, 12]. Based on a survey and analysis
47 of 503 RTUs conducted by Cowan [11], 54% of RTUs were found to have problems—42% had improper
4
48 airflow, 72% had improper refrigerant charge, and 20% had failed sensors. These problems lead to an
49 estimated excess energy consumption of 8%. Another study evaluated 109 RTUs in the field and found
50 that 89 of them had fault conditions—31 had two or more faults [12]. The average energy efficiency ratio
51 for the units increased from 6.6 before servicing to 7.0 after servicing—an average increase of more than
52 6%.
53 Some SMBs use programmable thermostats that can vary the temperature setpoint based on the time
54 of day and day of the week, allowing for night setup or setback (during unoccupied time periods).
55 However, the thermostat has limited ability to monitor or trend the data necessary for detecting system
56 degradation or for use in control optimization. For instance, the thermostat often uses a single setpoint for
57 each operating mode (e.g., occupied, unoccupied, heating, cooling, etc.). Many SMBs either do not have
58 programmable thermostats or the thermostats are not programmed accurately to reflect the desired
59 schedule or setpoint. Enforcing setpoints and schedules can result in energy savings of more than 20%
60 [7].
61 Studies have documented the impact of the zone setpoint temperature on the energy consumption of
62 HVAC equipment [13] and occupant comfort [14]. A study conducted by Hoyt et al. [13] indicates that
63 the desired zone setpoint temperature could lead to significant reduction in energy use without any
64 upgrade to HVAC hardware. The simulation was performed under seven different climate conditions and
65 results indicated that an average of 29% of the cooling energy and 27% of the total HVAC energy could
66 be saved by increasing the cooling setpoint from 22.2°C (72°F) to 25°C (77°F) in San Francisco,
67 California. Based on a survey and analysis of nine office buildings conducted by Karjalainen and
68 Koistinen [14], the thermal comfort was no better in the buildings equipped with individual thermostats
69 than in the buildings with no option for individual zone temperature control. The survey results indicated
70 that the individual temperatures were not controlled to maintain the zone temperature within a desired
5
72 Another problem associated with RTUs is short cycling—an operational mode during which an RTU
73 goes through ON and OFF cycles too frequently. Short cycling can be caused by equipment oversizing,
74 poor thermostat location, low refrigerant charge, clogged air filters, and other reasons. Excessive cycling
75 of RTUs can lead to excessive wear and premature failure of the compressor or its components [15–17].
76 Short cycling can also result in a significantly decreased average efficiency, even if there are no physical
77 failures in the equipment. According to Jacobs [15], the short cycling of RTUs can cause an average
78 energy efficiency penalty of approximately 10% compared to the expected values for fault-free operation.
79 Henderson et al. [16] estimated the energy penalty for short cycling to be roughly 11%. Another study
80 represented the average and maximum energy penalty to be 20% and 50% due to short cycling [17].
81 Hence, detecting the zone setpoint and too frequent cycling can be used in analyzing system performance,
83 The work reported in this paper describes two algorithms for detecting the zone setpoint temperature
84 and RTU cycling that are compatible for deployment on low-cost controls infrastructure. These
85 algorithms only require the zone temperature data for detection. In addition, the output of the algorithms
86 can be used as the control variable for analyzing system performance and developing more robust control
87 methodologies that can be incorporated into existing building systems. The algorithms have been widely
88 tested and validated using field data from 24 RTUs in six buildings. The remainder of the paper describes
89 the algorithms, the results from testing the algorithms, and conclusions.
90
91 2. Methodology
92 Two algorithms were developed to detect the zone setpoint temperature and number of ON/OFF
93 cycles of RTUs. Although the approach for detecting the zone setpoint temperature and number of
94 ON/OFF cycles uses proven mathematically processes, the innovative or unique aspect of the work
95 reported in this paper is that the algorithms only use zone temperature measurement to make conclusions.
96 While detection of the ON/OFF cycling can only be applied to single-zone RTUs, the setpoint detection
6
97 algorithm can be applied to single- and multi-zone RTUs with variable capacity modulation because these
98 systems employ a single thermostat to control the zone temperature. Furthermore, the setpoint detection
99 algorithm can also be applied to built-up air-handling systems. A detailed description of each of the
101
103 This section describes the development of the setpoint detection algorithm using only zone
104 temperature measurement as an input. The algorithm consists of three parts: preprocessor, peak detection,
105 and classification. During preprocessing, the noisy temperature data are filtered using a low pass filter.
106 Next, the peak detection process is used to detect peaks in the data and to measure their position, height,
107 and width. The output of the peak detection is processed by a Bayesian classifier, which compares two
108 adjacent setpoints to confirm whether they are the same or different.
109
111 The presence of random noise in the measured zone temperature data can result in incorrect detection
112 of setpoint temperatures. Therefore, a low pass filter is employed during preprocessing [18, 19] to remove
113 the noise. In signal processing, low pass filtering is a process that removes frequencies higher than a
114 certain preset frequency threshold. The temperature measurement in the time domain is converted to a
115 frequency domain signal in which the independent variable is frequency rather than time.
116 The Z-transform is performed on the time signals. This process converts a discrete-time domain
117 signal into a complex frequency domain representation. The Z-transformation of a discrete-time signal
118 x(n) was the formal power series X(z) defined as Equation (1):
=
!" 1)
121 frequencies higher than the cutoff frequency. Selection of a large cutoff can lead to more stable states, but
122 fewer temperature input data for setpoint detection. On the other hand, a smaller cutoff increases the
123 uncertainty of the setpoint detection outputs. Therefore, it is necessary to find cutoffs that minimize the
124 uncertainty of the setpoint detection while maximizing the use of input data. In this case, the appropriate
125 cutoff that minimizes the setpoint prediction error was chosen using a trial-and-error method. The
126 temperature frequency domain signal was then multiplied by the frequency response of a digital low pass
127 filter that removes high-frequency components. The inverse Fourier transform then recovered the filtered
129
131 The detection of peaks in signals is an important step in many signal processing applications. The
132 peak detection algorithm is used in nuclear monitoring [20], mass spectrometry [21], processing [22, 23],
133 and electronic systems [24]. Peak (or valley) points denote the significant events where the function graph
134 changes from increasing (decreasing) behavior to decreasing (increasing) behavior in time series [25].
135 The identification of these behaviors is important for analysis of the time-series data.
136 Figure 1 shows an example of the filtered zone temperature of an RTU during the summer season.
137 When the zone temperature exceeds the cooling temperature setpoint (plus dead band), the zone
138 temperature will reach a local maximum (denoted as Tmax, represented by the blue circles in Figure 1). As
139 the RTU turns ON and provides cooling to the zone, the temperature will decrease until the zone
140 temperature falls below the cooling temperature setpoint (minus dead band), at which time the RTU will
141 stop cooling (local minimum denoted as Tmin, represented by the red circles in Figure 1). The heat pump
142 operation in the heating mode is similar to the operation in the cooling mode. The difference is the
143 sequence of local maximum and minimum. For example, the zone temperature will reach a heating
144 temperature setpoint minus a dead band (local minimum denoted as Tmin), and then as the heat pump
8
145 system provides heating to the zone the temperature will rise until the zone temperature is above a heating
146 temperature setpoint plus dead band (denoted as Tmax), at which time the heat pump system will stop the
148
149 Figure 1: Zone Temperature and Four Criteria Used for Setpoint Detection
150 Peak detection uses the approach of finding the time locations and temperature amplitudes of Tmin
151 and Tmax. Let Tzone be a given array of zone temperatures that represents the time-series data [26]. A way
152 to detect Tmax in Tzone is to use the property that a peak must be greater than its immediate neighbors. For
153 example, given array Tzone with n elements, the algorithm finds the index i of peak element Tzone[i] as
9
155 For elements, i = 1 or i = n on the boundaries of the array, the element only needs to be greater than
156 or equal to its lone neighbor to be considered a peak. The array of maximum temperatures (Tpeak) is
158 In contrast, we define the array of minimum temperatures (Tvalley) as shown in Equations (5), (6), and
159 (7).
160 The key issue in the peak detection algorithm is the fact that peaks occur with different temperature
161 amplitudes and at different scales, which results in a large number of false positives among detected
162 peaks. To filter out valid peak information or fail to reject false peaks, the algorithm can measure the
163 position, height, width, and sequence of each peak. These measurements are compared to preset criteria to
164 reduce the identification of false peaks and ensure that the time locations and temperature amplitude
165 results are reasonable while finding peaks and valleys. For example, it is possible to detect only the
166 desired peaks and ignore peaks that are too small, too wide, or too narrow. A description of each
168
170 All RTU compressors need a minimum of 5 minutes of continuous run time to saturate bearings,
171 warm surfaces, and equalize refrigerant side pressure. The compressor ON/OFF signals are valid if the
172 time difference between the ON signal and OFF signal is at least 5 minutes. Therefore, the minimum time
173 period is set as 5 minutes as the default value (purple line shown in Figure 1). The time period between
10
174 the Tmax and neighboring Tmin must be greater than the minimum time period. The minimum time period
175 may have to be adjusted based on the thermostat’s time delay or the equipment’s internal setting.
176
178 Compressor equipment ON-OFF cycling should be a sequential order of events (e.g., OFF signal →
179 ON signal→ OFF signal). For instance, a valid Tmax point will be adjacent to two Tmin points as shown in
180 Figure 1. The same signals (two adjacent Tmax) that are repeated can be removed from Tpeak and Tvalley. For
181 example, the repeated ON signal detection is considered an indication of continuous compressor running
183
185 Tmax can be ignored if the temperature amplitude between Tmax and neighboring Tmin is less than a
186 minimum temperature amplitude (red line shown in Figure 1). When the temperature amplitude is kept
187 very small, all of the peaks including Tmax that resulted from noise in the temperature can be detected.
188 When the difference is kept very large, it will be difficult to detect Tmax. The suggested default for the
189 minimum difference is 0.2°C. For example, Tmax is valid if the temperature difference between Tmax and
190 neighboring Tmin is higher than 0.2°C. The minimum temperature amplitude can be adjusted based on the
192
194 The minimum cycling number is a minimum number of Tmax and Tmin per day required for accurate
195 setpoint detection. If the number of Tmax and Tmin is less than this value, any Tmax and Tmin are ignored. The
196 minimum cycling number of 5/day is the suggested default as shown in Figure 1. For example, if the
197 minimum cycling number is less than 5/day, then the setpoint detection indicates that the minimum
200 detection algorithm uses a fixed-length moving average for detecting setpoint temperature. The moving
201 average is obtained by taking the averages of fixed subsets of the number series. In this approach, the
202 number of fixed subsets (n) is set as 5. For example, each Tsp,pred can be determined by averaging the zone
203 temperature reading at the Tmax and Tmin. Five Tsp,pred in the fixed subsets are replaced by their average
204 value (Tsp,avg) of the data points calculated using Equation (8). Then, the subset is modified by excluding
205 the first number of the series and including the next number following the original subset in the series.
206 This process is repeated for the entire data series. In this example, Tsp,avg for the fixed window is 22.1°C
E# 4 #F +, H
∑+!J D F&G, I
2 ∑+!JE#A$,$K%L, H
#A$,&:B = = 8)
208
209 Table 1: Calculated Setpoints Based on an Example of the Zone Temperature Profile
210
212 When the building is occupied, the thermostat controls the RTU to maintain the zone temperature
213 within a desired comfort level. When the building is unoccupied, the programmable thermostat allows the
12
214 zone temperature to deviate from the occupied setpoint to an unoccupied setpoint, called night
215 setup/setback. Therefore, the setpoint detection should identify all possible setpoint values.
216 A statistical classifier can distinguish the different setpoint temperatures more effectively. There are
217 several possible classifiers to identify to which a set of categories (e.g., occupied cooling setpoint and
218 occupied heating setpoint, etc.) a new setpoint belongs. A Bayesian classifier is optimal with respect to
219 minimizing the classification error associated with current and new normal distributions [27]. Based on a
p ( x ω i ) p (ω i ) (
p (ω i x ) =
p(x ) 9)
221 where p(ωi|x) is the conditional probability of ωi having accounted for evidence x, p(ωi) is prior
222 probability, p(x|ωi) is the class conditional probability of x, and p(x) is the mixture density function.
223 Because p(x) is positive and common to both sides of the inequality, the Bayes decision rule of Equation
(
ω1 ( Current ) : p ( x ω 1 ) p (ω 1 ) ≥ p ( x ω 2 ) p (ω 2 )
10)
(
ω 2 ( New ) : p ( x ω1 ) p (ω1 ) ≤ p ( x ω 2 ) p (ω 2 )
11)
225 Figure 2 shows two conditional probabilities P(x|ωi), i = 1, 2, as functions of x in each of the classes.
226 The dashed line at x0 is a threshold partitioned into two regions, R1 and R2. According to the Bayesian
227 decision rule, for all x values in R1 the classifier decides ω1 and for all x values in R2 it decides ω2.
228 However, there is overlapping probability, which is equal to the total shaded area under the curves
229 belonging in R1 and R2, shown in Figure 2. The shaded area is the Bayesian classification error probability
231 (12)
xo −∞
13
232 One of the most common probability density functions in practice is the normal probability density
233 function. Where P(x|ωi) is the normal distribution with mean μi and covariance Σi, Equations (13) and
p ( x ω1 ) ~ N (µ 1 , Σ 1 ) (
13)
p ( x ω 2 ) ~ N (µ 2 , Σ 2 ) (
14)
235 where x is a vector of current residuals, μ1 is the mean describing the distribution of the new setpoint, Σ1
236 is the covariance describing the uncertainty of the new setpoint, μ2 is the mean describing the distribution
237 of the current setpoint, and Σ2 is the covariance describing the uncertainty of the current setpoint.
238 The Bayesian classification error is calculated by integrating the overlap area between probability
239 distributions that fall within each class region of the domain, as shown in Figure 2. The classification
240 error probability decreases as the error of estimated value becomes more significant and is therefore a
241 useful measure for distinguishing the current setpoint from the new setpoint temperature. The thresholds
242 for the classification error were established by evaluating the statistical significance of a match or
243 mismatch between the current and the new setpoint temperatures. For this research effort, the threshold
244 was determined to be a 0.1 classification error using existing field data.
14
245
246 Figure 2: Bayesian Decision Rule for Minimum Error (ω1: new setpoint & ω2: current setpoint)
247
249 This section describes the methodology used to detect RTU ON/OFF cycling based on zone
250 temperature measurement. The RTU can provide mechanical cooling (direct expansion vapor
251 compression) from cool outside air (economizing) or a combination of mechanical cooling and
252 economizing (when economizing alone cannot meet the cooling needs of the conditioned zone). The
253 ON/OFF cycling algorithm is useful when the thermostat’s cooling command value is not available.
254 ON/OFF cycling detection can be used to quantify the number of times an RTU is ON or OFF in a given
255 time period without monitoring the RTU compressor. If the number of detected cycles is higher than a
256 predefined threshold during a certain time period, it indicates a RTU short-cycling problem. The RTU
15
257 cycling is a function of a number of variables including oversizing and temperature differences between
258 the indoors and outdoors. The RTU is expected to cycle more often when the temperature difference
259 between indoors and outdoors is high or if the unit is significantly oversized.
260 In a zone temperature profile, the peak indicates “ON” and the valley indicates “OFF.” The peak
261 detection (2.1.2 Peak Detection) can detect the number of peaks (ON)/valleys (OFF) in the daily zone
262 temperature time-series data. The algorithm reads Tzone, searches for valid peaks and valleys, and keeps
263 track of the total number of peaks and valleys found. The overall number of ON/OFF cycles in a day can
264 be calculated by summing the number of peaks. The algorithm can detect excessive equipment ON/OFF
265 cycling and equipment that remains in an ON or OFF state for significant periods of time. If the RTU runs
266 for a very long time and then shuts off for just a few minutes, this type of information could be used as
267 evidence of control or mechanical faults. Overall, the proposed technique does not require any supervised
268 learning or extra additional sensor installation, which is costly and time consuming. Two thresholds,
269 which are user adjustable configuration parameters, are described below.
270
272 The maximum number of cycles represents the maximum number of cycles expected per day. A
273 cycling problem is detected when the number of cycles exceeds the maximum cycling number (default
274 value 60 cycles/day). If the number of detected cycles is higher than a predefined threshold during a
276 The minimum cycling number represents the minimum number of cycles per day. A cycling problem
277 is detected when an RTU remains either in the ON or OFF state for significant period of time (default
278 minimum value is 0 cycles/day). To avoid false detection, the minimum cycling detection should only be
279 employed when the outdoor conditions are significantly different from the zone setpoint. For example, the
280 RTU may not cycle when the outdoor conditions are close to the setpoint.
281
16
282 2.3. Metrics Used for Validation of the Two Algorithms
283 Validation of the algorithms is critical for ensuring that they can reliably estimate the setpoint and
284 the cycling rate. The methodology used to validate the two algorithms follows.
285
287 The confidence interval with t-distribution is used to quantify the accuracy of the setpoint detection.
288 The width of the interval depends upon the confidence level and the precision of the prediction as shown
̅/ ∙ 1 1 ̅4 ∙
(
M M
√ √ 15)
290 where M is t-value for corresponding confidence level, ̅ is a mean of prediction, is a number of
292 The confidence level is used as an accuracy/validation metric for measuring the success of the
293 algorithm. The confidence level is the probability of how surely the setpoint prediction lies within a
294 confidence interval [28]. Depending on the confidence level chosen, the interval margin of error and
295 respective range also change. The difference between the identified setpoint and the actual setpoint is
296 used as a residual input to identify an interval in which the prediction will lie. The goal of the algorithm is
297 to identify the setpoint temperature to within +/-0.6°C and with 90% confidence.
298
300 This section describes the metric used for the ON/OFF cycling detection. The mean absolute percent
301 error (MAPE) estimate is used to quantify the accuracy of the ON/OFF cycling detection algorithm. The
302 MAPE between the identified and actual value is a widely used accuracy metric. The MAPE can be
303 calculated as the average absolute percent error, as shown in Equation (16). The ON/OFF cycling
304 algorithm estimates the number of ON/OFF cycles over a 24-hour period. The goal of the ON/OFF
17
305 cycling detection algorithm is to identify the number of ON/OFF cycles with a MAPE value of less than
306 20%.
!+
1 /
% = Q1 / R RS × 100
(
!J 16)
307 The MAPE is scale-sensitive and should not be used when working with small size test data.
308 Because "Actual" is in the denominator of the Equation (16), the MAPE is undefined when “Actual” is
309 quite small and the MAPE will often take on extreme values.
310
312 The two algorithms were validated using field data from 24 RTUs from 6 buildings in various
313 climate zones. First, the data used to validate the algorithms is presented, followed by validation of the
315
317 To test and validate algorithms, field test data from 24 RTUs (air-conditioners and heat pumps) from
318 six different locations in the U.S. were used, as listed in Table 2. RTUs can operate in either heating or
319 cooling modes. Building A and Building B are office buildings located in suburban of Seattle,
320 Washington, and Berkeley, CA. Building C is a mechanical shop building with offices located in
321 Richland, Washington. Building D is a machine product factory located in a suburb of South Paris, ME.
322 Building E and Building F are fitness centers located in Miami, FL, and Cupertino, CA. Buildings in
323 different climate zones and local conditions were selected to evaluate the performance of the algorithms.
324 A data set was prepared by selecting measurements over a period of one week based on the date
325 information. For example, the summer data set was selected during peak summer conditions in July and
18
327 The different data sets from each RTU were selected from existing field data based on the outdoor
328 air temperature conditions (i.e., spring/fall, summer, and winter). Each data set contains five
329 measurements (zone temperature, setpoint temperature, supply fan status, cooling/heating command, and
330 outdoor temperature) from a one-week period. Although the algorithm used only zone temperature, the
331 remaining sensor data are used as ground-truth data to verify the results of the algorithm. For example,
332 the cooling/heating command is used to estimate how long the RTU was ON or OFF.
333
334
19
335 Table 2: Validation Data Set for Six Different Locations
338 The setpoint detection approach is explained using data from RTU-1 in Building B during the
339 spring/fall-Day 2. The blue and green lines indicate the raw zone temperature and filtered zone
340 temperature, respectively, as shown in Figure 3. The noise in the raw zone temperature (blue) is reduced,
341 while the peak remains the same in the filtered zone temperature (green), making it easier to measure the
20
342 peak position, height, and width.
343
344 Figure 3: (a) Zone Temperature (top) and (b) Filtered Zone Temperature (bottom) Using Data from
346 Figure 4 shows zone temperature (blue) and the detected setpoint temperatures (green) over a 24-
347 hour period. In this example, the zone has two distinct temperature setpoints: the first is the occupied
348 cooling temperature setpoint (22.0°C [71.5°F]), and the second is the unoccupied cooling temperature
349 setpoint (24.0°C [75.0°F]). The blue and red circles indicate Tmax (blue) and Tmin (red) estimated by the
350 setpoint detection algorithm. The total number of Tmax and Tmin points were 28 for this day. The Tsp was
351 calculated using the average of corresponding Tmax and Tmin. The setpoint detection algorithm uses a fixed-
21
352 length moving window average that creates a series of average setpoint temperatures of different subsets
353 of the full Tsp set. The size of subsets used for validation of the algorithm was 5 points. The first setpoint
354 temperature is calculated by averaging the first five points from the Tsp, as shown in Figure 4. Then the
355 subset is modified by excluding the first point from Tsp and adding the sixth point from Tsp. The average
356 for points 2 to 6 represents the second setpoint temperature. This process is repeated over the full Tsp set.
357 In this example, there are 24 average setpoint temperatures (given that there are 28 points).
358
360 The second step in the setpoint detection process, is to distinguish whether two adjacent detected
361 setpoints are distinct and different. The probability distribution function is used to isolate distinct
362 setpoints. Figure 5 shows an example of the probability distributions for two setpoints that correspond to
363 the first and the eighth moving window subset. The mean for the first point (red dashed line) and eighth
364 setpoint temperature (blue dashed line) are 22.1°C (71.8°F) and 22.2°C (71.9°F), respectively. As
22
365 discussed previously, the Bayesian classifier is used to estimate the classification error (overlap area)
366 between the first and eighth distributions. The significant overlap between the first and eighth probability
367 distributions indicates that there is no difference between the two sets. The classification error based on
368 the residuals between first and eight setpoint is 0.59, indicating that there is no significant statistical
370
371 Figure 5: Overlapping Normal Distribution between 1st and 8th Setpoint Temperature Prediction
373 The third and the final step in setpoint detection is to compare the estimated setpoint with the
374 ground-truth data. As discussed in the previous section, the confidence level is used as an accuracy metric
375 for measuring the success of identifying the setpoint detection. The difference between the predicted and
376 the actual setpoint is used as residual input to the confidence interval to identify an interval in which the
377 prediction will lie. As stated earlier, the goal of the setpoint algorithm is to identify the setpoint
378 temperature within ±0.6°C with 90% confidence. Figure 6 shows the probability density function. Using a
23
379 sample data set, the algorithm identified the setpoint within 0.2°C of the actual setpoint. The confidence
380 interval for this sample data set at the 90% confidence level is between 0.02°C and 0.44°C. Because the
381 upper bound (0.44°) for the sample data set is smaller than 0.6°C, the identified setpoint is considered to
383
386 Now that the first and the eighth setpoints are considered to be same and also meet the accuracy
387 metric, the next step is to compare the distribution of the first and the ninth setpoints. This process of
388 comparing the distributions continues for all setpoints. Figure 7 compares the first and the eighteenth
389 moving window subsets and shows a much larger difference between the first and the eighteenth setpoint
390 probability distributions, indicating that the setpoints are distinct and different. The Bayesian error
391 between the first and the eighteenth setpoint temperature is 0.07, which is less than the established
394 Figure 7: Overlapping Normal Distribution between 1st and 18th Setpoint Temperature Prediction
396
397 Figure 8 shows the probability density function for the second distinct setpoint. The absolute
398 difference between the predicted (23.9°C [75.0°F]) and the actual (24.0°C [75.1°F]) setpoint value is
399 0.1°C. The interval corresponding to the 90% confidence level lies between 0.0°C and 0.5°C. Because the
400 upper bound of the confidence interval (0.5°C) is less than 0.6°C (1°F), the predicted setpoint is
402
25
403
406
407 Table 3 provides an overall summary of the building data used for validation of the setpoint
408 detection algorithm and results. For Building A, the evaluation was performed using sample data from
409 four RTUs spanning seven days of each season (summer, winter, and spring/fall) from four RTUs. To
410 evaluate the performance of setpoint detection, the predicted setpoint was compared to a target value. The
411 analysis of data from Building A indicated a total of 66 detections, including 61 correct and 5 incorrect
412 detections.
413 The setpoint detection performed well under a wide range of outdoor temperature conditions. On 23
414 occasions no detection was possible because the number of ON/OFF cycles were less than 5. Because the
415 setpoint detection requires a minimum number of peaks and valleys to predict setpoint temperature, it had
416 difficulty when the number of peaks and valleys was less than 5. For Building A, the weekday occupied
26
417 schedule is 6:00 a.m. to 5:00 p.m. The RTUs are in an unoccupied mode the rest of the time, including all
418 day on weekends. Most of the days when the ON/OFF detection was not possible were either weekends
420
421 Table 3: Summary of Building Data Set Used for Setpoint Detection and Results
423 For Building B, the setpoint detection performed well with a total of 96 detections, including 88
424 correct and eight incorrect detections. On 44 occasions no detection was possible and most of these
425 occasions were either during a weekend or during unoccupied weekdays (the occupancy period for
426 Building B is 7:00 a.m. to 5:00 p.m. Monday through Friday and closed on weekends). The number of
27
427 “no” detections was higher than for Building A, because Building B is open space and thus only some of
428 the six RTUs were necessary to maintain the comfort in the space.
429 Based on data from eight RTUs from Building C, the analysis included a total of 116 detections,
430 including 95 correct and 21 incorrect detections. The number of incorrect detections was relatively higher
431 than the other two buildings, and the reason for this is explained below. Figure 9 shows the zone
432 temperature profile and cooling command for the HP4 system in Building C during the cooling season.
433 The blue, red, and black lines represent the zone temperature, setpoint temperature, and the cooling
434 command signal, respectively. Of the 21 incorrect detections, in some cases the system was cycling
435 several times before it reached the setpoint temperature; therefore, the predicted setpoint value was higher
436 than the actual zone setpoint temperature. The setpoint detection predicts 22°C (75°F) when the actual
437 setpoint is 21°C (70°F). On 60 occasions no detections were possible because Building C was occupied
438 from 5:30 a.m. until 5:00 p.m., Monday through Friday and closed on weekends. Because most of the
439 RTUs were turned OFF during the weekends, most of the “no” detection data were from weekends.
440 For Building E and Building F, the setpoint detection algorithm was evaluated using sample data
441 from two RTUs spanning seven days. For Building D, no winter data were used because both RTUs are
442 air-conditioners only. The analysis included 22 correct detections and 1 incorrect detection and on 5
443 occasions detection was not possible. The analysis of data from Building E indicated a total of 42
444 detections, including 41 correct detections and 1 incorrect detections. For both buildings, all setpoints
445 were detected correctly with one exception because the building was a grocery store with a single setpoint
446 all week. Overall, the setpoint algorithm showed good performance under a wide range of operating
447 conditions.
448 For Building F, the analysis included 37 detections—29 correct and 8 incorrect. Five setpoints could
449 not be detected because the system was cycling several times before it reached the setpoint temperature
28
451
453
454 Overall, the setpoint detection algorithm has been shown to provide accurate setpoint temperature
455 estimates using just the zone temperature data from a number of RTUs in diverse climate locations and
456 building types. The setpoint detection algorithm correctly detected 354 out of 399 setpoint temperatures.
457 Although the absolute difference between the predicted and actual setpoint temperature for 45 cases was
458 greater than 0.6°C (1°F), the difference between the prediction and actual value was within ±1.0°C
459 (1.8°F) for most of these cases. The setpoint detection algorithm can be deployed on SMBs that do not
460 have building automation systems using low-cost single board computers (e.g., Raspberry PI,
461 BeagleBone, etc.) with a single inexpensive temperature sensor to measure zone temperature. The
462 algorithm can generate actionable information, including detecting incorrect temperature setpoints
29
464 3.3. Validation of HVAC Cycling Detection Algorithm
465 Excessive (short) cycling of the system can lead to premature failure of the compressor or its
466 components. Short cycling can also result in a significantly decreased average efficiency, even if there are
467 no physical failures in the equipment. First, the cycling detection algorithm is explained using cooling
468 season data from RTU-1 in Building D. Figure 10 shows the zone temperature profile and cooling
469 command for RTU-1 in Building D during the cooling season. The blue and black lines indicate the zone
470 temperature and compressor signal, respectively. The zone temperature swings within the dead band
471 around the Tsp. The RTU is ON when the zone temperature reaches Tma and continues to operate until the
472 zone temperature reaches Tmin. As seen in the figure, the Tmax and Tmin correctly correspond to compressor
473 signals 0 (OFF) and 1 (ON), respectively. The total number of cycles predicted by the algorithm was 35
474 cycles/day and the actual number of cycles was 34 cycles/day. The corresponding MAPE metric is 97%,
476
30
477 Figure 10: Zone Temperature Profile and Cooling Command (Building D-RTU1-Summer/Day1)
478
479 Table 4 presents a summary of the building data set used for validation of the ON/OFF cycling
480 detection algorithm and results. The results are split into two groups: 1) a group of days when the number
481 of ON/OFF cycles per day was greater than 15, and 2) a group of days when the number of ON/OFF
482 cycles per day was greater than 5. The algorithm was more successful in detecting the ON/OFF cycles
483 when they exceeded 15 cycles/day than when the ON/OFF cycles exceeded 5 cycles/day. A previous
484 study [11, 12] found that 2.5 cycles/h is an average maximum cycling rate that happens under hot summer
485 conditions. For Buildings A, B, C, and D, the ON/OFF cycling detections with more than 5 daily
486 ON/OFF cycles were 77%, 80%, 65%, and 87% accurate, respectively. The accuracy of ON/OFF cycling
487 detection increased to 88%, 85%, 88%, and 100% when the data only included days with 15 or more
488 ON/OFF cycles. The performance of cycling detection algorithm is good over a wide range of operating
489 conditions. For Building C, the data with more than 5 daily ON/OFF cycles were 65% accurate, and the
490 accuracy increased to 88% when only data with 15 or more cycles were used.
491 For Building E, the accuracy of detection was 41% with data containing at least 5 ON/OFF cycles
492 per day. The accuracy did not increase with data containing 15 ON/OFF cycles per day, the reason is
493 explained Figure 11. Also, the data from Building C and Building E showed inconsistent cycling
494 frequency (e.g., 1 minute, 10 minutes). This inconsistency led the algorithm to miss several ON/OFF
495 cycles.
496
31
497 Table 4: Summary of Building Data Set and Results for ON/OFF Cycling Detection
499 Figure 11 shows the zone temperature profile and cooling command of the HP3 system in Building
500 C during the cooling season. The blue and red lines indicate the zone temperature and cooling command
501 signal, respectively. As shown, the cooling command signals do not correspond to the zone temperature
502 trend, because several RTUs in Building C serve open space. Therefore, there is unknown but highly
32
503 possible inter-zonal convective coupling due to close interactions between the different RTUs and their
505
506 Figure 11: Temperature Profile and Cooling Command (Building C-HP3-Summer/Day 2)
507
508 Table 5 shows the overall summary results of the ON/OFF cycling detection algorithm. For data with
509 more than five ON/OFF cycles/day, the algorithm detected 203 out of 283 correctly. For data with more
510 than 15 ON/OFF cycles/day, the algorithm detected 133 out of 156 correctly. Because the goal is to detect
511 frequent ON/OFF cycles, lower accuracy of detection when there are fewer cycles than 15 is not critical.
512 The weekly ON/OFF cycling detection was calculated based on a sum of cycling over a one-week period.
513 For data with more than 75 weekly ON/OFF cycles, the cycling detection was incorrect only once.
514
33
515 Table 5: Summary of Daily ON/OFF Cycle Detection Results
517 4. Conclusions
518 Two algorithms for detecting zone setpoints and RTU cycling were developed, tested, and validated.
519 To validate the algorithms, data monitored in a number of zones/RTUs from six buildings in different
520 climate locations were used. The setpoint and cycling detection algorithms use a peak detection technique
521 that only requires one measured temperature to detect peaks/valleys in any given time-series data. These
522 algorithms do not require supervisory learning or additional sensor installation other than the zone
523 temperature. Two different metrics are used to validate the two different algorithms. To evaluate the
524 accuracy of algorithms, three different data sets for each zone/RTU were selected from the field data for
525 different periods in the year (i.e., spring/fall, summer, and winter). Each selected data set contained five
526 measurements (zone temperature, setpoint temperature, supply fan status, cooling/heating command, and
527 outdoor temperature) from a one-week period. Although only zone temperature was used by the
528 algorithm, the remainder of the data were used as ground-truth data to verify the results of the algorithm.
529 Overall, the algorithms were successful in accurately detecting the setpoints and ON/OFF cycles using the
530 peak detection technique. Because these algorithms only use zone temperature for identification, if the
531 RTUs are not actively heating/cooling, the algorithms will not be able to accurately identify setpoints and
34
533 There is increased urgency to improve the operating efficiency of the existing U.S. commercial
534 building stock for many reasons, chief among them being to mitigate the impacts of climate change [31,
535 32]. Many cities (e.g., New York City, Seattle, etc.) are passing ordinances that would require periodic
536 retro-commissioning of commercial buildings. The U.S. climate action plan will give new impetus to
537 many states and cities to follow Seattle and New York City’s examples. Although the traditional retro-
538 commissioning process can be effective, it cannot ensure the persistence of building operations beyond a
539 few months and the process is also perceived to be costly. A technology-based solution can deliver retro-
540 commissioning service at a lower cost and will also ensure the persistence of building operations because
542 The technology-based solution requires software algorithms and applications that can be deployed in
543 the Cloud or on low-cost platforms (Raspberry PI and BeagleBone, etc.) in the buildings. The work
544 reported in this paper covers two such algorithms. These algorithms can be used to ensure the persistence
545 of building operations in SMBs by enforcing setpoints and mitigating the reasons for RTU short cycling.
546 The algorithms can also automatically detect the overrides of temperature setpoints to provide actionable
548
549 5. Acknowledgments
550 The authors acknowledge the Buildings Technologies Office of the U.S. Department of Energy’s
551 Office of Energy Efficiency and Renewable Energy for supporting this research and development effort.
552 The authors thank Dr. Marina Sofos and Joseph Hagerman, Technology Development Managers for their
553 guidance and strong support of this work. At PNNL, we acknowledge George Hernandez for his technical
554 guidance, and Susan Ennor for her editorial support in preparing this document.
35
555 6. References
556 [1] M.M. Ardehali, T.F. Smith, Literature Review to Identify Existing Case Studies of Controls-Related
557 Energy-Inefficiencies in Buildings, Prepared for the National Building Controls Information Program
558 (2001).
559 [2] S. Katipamula, M.R. Brambley, Fault Detection, Diagnostics and Prognostics for Building Systems –
561 [3] S. Katipamula, M.R. Brambley, Fault Detection, Diagnostics and Prognostics for Building Systems –
563 [4] M.S. Breuker, J.E. Braun, Common Faults and their Impacts for Rooftop Air Conditioners, HVAC&R
565 [5] P. Jacobs, Small HVAC Problems and Potential Savings Reports, California Energy Commission
567 [6] EIA, Commercial Buildings End-Use Consumption Survey. 2012a. Retrieved October, 2016 from:
568 http://www.eia.gov/consumption/commercial/index.cfm.
570 Brown, S.M. Lanzisera, T. Kuruganti, Small-and Medium-Sized Commercial Building Monitoring and
571 Controls Needs: A Scoping Study, Pacific Northwest National Laboratory, 2012. PNNL-22169.
572 [8] K. Mařík, J. Rojíček, P. Stluka, J. Vass, Advanced HVAC control: Theory vs. reality. IFAC
574 [9] E. Mills, Building commissioning: a golden opportunity for reducing energy costs and greenhouse
575 gas emissions in the United States., Energy Efficiency, 4(2) (2011) 145-173.
576 [10] L. Wang, S. Greenberg, J. Fiegel, A. Rubalcava, S. Earni, X. Pang, R. Yin, S. Woodworth, J.
36
579 [11] A. Cowan, Review of recent commercial rooftop unit field studies in the Pacific Northwest and
580 California, New Buildings Institute, White Salmon, WA, USA (2004).
581 [12] ADM Associates, ADM Associates Market assessment and field M&V study for comprehensive
583 [13] T. Hoyt, E. Arens, H. Zhang, Extending air temperature setpoints: Simulated energy savings and
584 design considerations for new and retrofit buildings. Building and Environment 88 (2015) 89-96.
585 [14] S. Karjalainen, and O. Koistinen, User problems with individual temperature control in
587 [15] P. Jacobs, Small HVAC system design guide: Design Guidelines, California Energy Commission
588 (2003b).
589 [16] H. Henderson, D. Parker, J. Huang, Improving DOE-2’s RESYS Routine: User Defined Functions to
590 Provide More Accurate Part Load Energy Use and Humidity Predictions, Lawrence Berkeley National
592 [17] E. Djunaedy, K. Van den Wymelenberg, B. Acker, H. Thimmana, Oversizing of HVAC system:
593 signatures and penalties, Energy and Buildings, 43 (2) (2011), 468–475.
594 [18] S.J. Orfanidis, Introduction to signal processing. Prentice-Hall, Inc., 1995.
595 [19] S. Smith, Digital signal processing: a practical guide for engineers and scientists, Newnes, 2013.
596 [20] I. Azzini, R. Dell’Anna, F. Ciocchetta, F. Demichelis, A. Sboner, E. Blanzieri, A. Malossini, Simple
597 methods for peak detection in time series microarray data. Proc. CAMDA’04 (Critical Assessment of
599 [21] K.R.Coombes, S.Tsavachidis, J.S.Morris, K.A. Baggerly, M.C. Hung, H.M. Kuerer, Improved peak
600 detection and quantification of mass spectrometry data acquired from surface‐enhanced laser desorption
601 and ionization by denoising spectra with the undecimated discrete wavelet transform, Proteomics, 5(16)
37
603 [22] V.T. Jordanov, D.L. Hall, M. Kastner, Digital Peak Detector with Noise Threshold, Proc. IEEE
605 [23] K. Harmer, G. Howells, W. Sheng, M. Fairhurst, F. Deravi, A Peak-Trough Detection Algorithm
606 Based on Momentum, Proc. IEEE Congress on Image and Signal Processing 4 (2008) 454 – 458.
607 [24] R. Schneider, Survey of Peaks/Valleys identification in Time Series, Department of Informatics,
609 [25] G. Palshikar, Simple algorithms for peak detection in time-series. In Proc. 1st Int. Conf. Advanced
611 [26] T.H. Cormen, Introduction to algorithms, Presses Massachusetts Institute of Technology (2009).
612 [27] K. Fukunaga, Introduction to Statistical Pattern Recognition, Academic Press, West Lafayette, IN
613 (1990).
614 [28] D. Altman, D. Machin, T. Bryant, M. Gardner, Statistics with confidence: confidence intervals and
616 [29] W. Kim, RG. Lutes, Development and Evaluation of Algorithms to Improve Small-and Medium-Size
617 Commercial Building Operations, Pacific Northwest National Laboratory, 2016. PNNL-25996.
618 [30] M. Baechler, J. Williamson, T. Gilbride, P. Cole, M. Hefty, P. Love, Building America best practices
619 series: volume 7.1: guide to determining climate regions by county, Pacific Northwest National
622 http://www.seattle.gov/Documents/Departments/OSE/OSE%20Building%20Tune-Ups%20ORD.pdf
38