Experimental Demonstration of Frequency Regulation by Commercial Buildings - Part II: Results and Performance Evaluation
Evangelos Vrettos, Emre C. Kara, Jason MacDonald, Göran Andersson, Duncan S. Callaway
EExperimental Demonstration of FrequencyRegulation by Commercial Buildings – Part II:Results and Performance Evaluation
Evangelos Vrettos ∗ , Emre C. Kara † , Jason MacDonald ‡ ,G¨oran Andersson § , and Duncan S. Callaway ¶ Power Systems Laboratory, ETH Zurich, Switzerland National Accelerator Laboratory (SLAC), California, US Grid Integration Group, Lawrence Berkeley National Laboratory (LBNL),California, US Energy and Resources Group, University of California, Berkeley, USMay 19, 2016
Abstract
This paper is the second part of a two-part series presenting the resultsfrom an experimental demonstration of frequency regulation in a commercialbuilding test facility. In Part I, we developed relevant building models anddesigned a hierarchical controller for reserve scheduling, building climatecontrol and frequency regulation.In Part II, we introduce the communication architecture and experimentsettings, and present extensive experimental results under frequency regula-tion. More specifically, we compute the day-ahead reserve capacity of the testfacility under different assumptions and conditions. Furthermore, we demon-strate the ability of model predictive control to satisfy comfort constraintsunder frequency regulation, and show that fan speed control can track thefast-moving RegD signal of the Pennsylvania, Jersey, and Maryland PowerMarket (PJM) very accurately. In addition, we report the observed effectsof frequency regulation on building control and provide suggestions for real-world implementation projects. Our results show that hierarchical control isappropriate for frequency regulation from commercial buildings. ∗ [email protected] † [email protected] ‡ [email protected] § [email protected] ¶ [email protected] a r X i v : . [ m a t h . O C ] M a y cronyms AHU
Air Handling Unit AS Ancillary Service
CWS
Central Working Station
HVAC
Heating, Ventilation and Air-Conditioning
MAE
Mean Absolute Error
MAPE
Mean Absolute Percentage Error
MPC
Model Predictive Control PI Proportional-Integral
PJM
Pennsylvania, Jersey, and Maryland Power Market
RMSE
Root Mean Squared Error
SAT
Supply Air Temperature
In Part I of this two-part paper, we performed a detailed literature review on the-oretical, simulation-based, and experimental work on frequency regulation withcommercial buildings. Furthermore, we presented the test facility for our experi-ment (FLEXLAB), developed relevant building models, and designed a hierarchi-cal controller for reserve scheduling (level ), building climate Model PredictiveControl (MPC) (level ) and frequency regulation (level ).In Section 2 of Part II, we summarize the control and communication architec-ture, as well as the experiment settings. In Sections 3, 4, and 5 we report extensiveexperimental results in FLEXLAB for each level of control over a period of oneweek. We summarize some important findings and suggestions for future work inSection 6, whereas Section 7 concludes. We implement the reserve scheduling problem of level (solved once a day) andthe MPC of level (solved every minutes) in Matlab, whereas we calculate thefan speed setpoints of level in Python (every seconds) and communicate them2 easurements HVAC system MeasurementsSetpoints
Python script (periodicalexecution)Forecast.io
Weather forecasts Python script (periodical execution)
Level 2 : Room climate controller
Air mass flow setpoint calculation every 15 min in MatlabModel predictive control
Level 3 : Frequency regulation controller
Fan speed control every 4 sec in Python
Model-based feedforward controlPI feedback control reserve.csvFrequency signalforecast .csvmeasure .csv Central working station
Scripting interface
Default controller setpoint.csv
Kalman filter
Level 1 : Reserve scheduler
Reserve capacity determination on a daily basis in MatlabRobust optimization Kalman filter
PJM
SiemensJetstream (LBNL - gateway)
DNP3 translationSQL database Reg signal push
Reg signal client
DNP3 measure .csv
LBNL regulation server
LBNL hierarchical controller
LBNL FLEXLAB
Figure 1: The developed control and communication architecture for building cli-mate control and frequency regulation in the FLEXLAB test facility.to the Central Working Station (CWS) of FLEXLAB. We used a file-based com-munication between Python and Matlab based on comma-separated-values (csv)files.The reserve scheduler stores the computed reserve capacity in the reserve.csvfile. A Python script periodically queries the CWS and stores the building mea-surements in the measure.csv file. Another Python script periodically queries thepublicly available database of forecast.io and stores the weather forecasts in theforecast.csv file. The MPC’s feedback from the building is obtained from mea-sure.csv and the weather forecasts from forecast.csv. The optimal air flow ratesetpoint calculated by Matlab is stored in the setpoint.csv file. The fan speed set-point is determined in Python by accessing the setpoint.csv and reserve.csv files,and based on the frequency regulation signal.Most of the experiment was performed using archived data of the RegD signalfrom the Pennsylvania, Jersey, and Maryland Power Market (PJM) from December to January . Although the signal was available with a resolution of sec-onds, we down-sampled it to seconds due to the expected communication delays.In addition, a connection with PJM was established based on the DNP3 protocoland using a Siemens Jetstream gateway that provided us with the RegD signal inreal-time. At the FLEXLAB side, the received data were translated, saved in anSQL database, and pushed by a “RegD signal server” to a “RegD signal client”.The complete communication architecture from PJM to FLEXLAB is graphicallyshown in Fig. 1. However, network issues at FLEXLAB made the connection un-reliable, and therefore we chose to run the live connection with PJM only for onecontinuous hour. Since FLEXLAB is not occupied, we emulated the internal heat gains from occu-pants and equipment using electric heaters as plug loads. The total internal heat Only ambient temperature forecasts are obtained from forecast.io. The solar radiation forecastsare obtained from a clear-sky radiation model, which turned out to be sufficient for the weatherconditions during the experiment. ov 16 2015 Nov 17 2015 Nov 18 2015 Nov 19 2015 Nov 20 2015 Nov 21 2015Time24002600280030003200 H ea t e r po w e r ( W ) Actual power Scheduled power
Figure 2: The heater schedule and the actual power consumption. The heat gain ishigh during working hours and low during non-working hours.
Nov 16 2015 Nov 17 2015 Nov 18 2015 Nov 19 2015 Nov 20 2015 Nov 21 2015Time15.015.516.016.517.017.518.018.519.0 S upp l y a i r t e m p e r a t u r e ( C ) Actual SAT Scheduled SAT
Figure 3: The SAT setpoint and actual values during one week.gain in both cells was kept lower than the chiller cooling capacity. The heaters’consumption profile was fixed according to the red curve of Fig. 2 using digitaltimer sockets. The actual heater power (blue curve) fluctuates around the profiledue to voltage variations.Before the start of the experiment, we fixed the manually controlled inlet dampersin the rooms to fully open positions. In addition, we fixed the return air damperto a opening and the outside air damper to a opening, i.e., the return airwas fully recirculated. The speeds of primary and secondary chilled water pumpswere fixed to and of their rated speeds, respectively. Moreover, we de-activated the existing floor heating system and the heating coil at the Air HandlingUnit (AHU).We set the temperature comfort zone to − ◦ C during working hours. Anexisting Proportional-Integral (PI) controller regulates the Supply Air Tempera-ture (SAT) to ◦ C by controlling the position of a cooling valve. The gains ofthis controller had been tuned for a conventional building operation; therefore,we modified them to achieve a tighter control and reduce the fluctuations of SATaround its setpoint during frequency regulation. The resulting SAT profile duringthe experiment is shown in Fig. 3. The mean deviation from the SAT setpoint is . ◦ C (there is a small bias to larger SAT values) and the mean absolute deviationis . ◦ C.Recall that the facility has two building cells with identical construction: cell4 R e s e r v e c apa c i t y ( i n % o f f an r a t ed po w e r) Symmetric reserves with setback on 20 November 2015
Up−reserves Down−reserves R e s e r v e c apa c i t y ( i n % o f f an r a t ed po w e r) Asymmetric reserves with setback on 21 November 2015
Up−reserves Down−reserves
Figure 4: The hourly reserve capacities as a percentage of the nominal fan powerfor November (left) and November (right).1A is used for the frequency regulation experiment, whereas cell 1B serves asa benchmark. Applying the same air flow rate in both cells and recording thetemperature, we verified that the two cells are thermally very similar. However, weobserved that that the same fan speed setpoint induces a slightly different air flowrate in the two cells due to small differences in the AHUs. To compensate for this,we fitted different fan models for the two cells (the parameters for cell 1A are givenin [1, Table IV]).The electricity cost was assumed equal to c k = 0 . e /kWh, whereas the re-serve capacity payment was fixed to a higher value, i.e., λ k = 0 . e /kWh.The goal of this experiment is to demonstrate the technical feasibility of reserveprovision from commercial buildings; therefore, we chose a relatively high capac-ity payment to incentivize reserve provision. The experiment was organized into two parts. The first part took place from to November and relied on an “older” building model identified with datafrom June-July (see [1, Table II]). On November the experiment waspaused and a new building model was identified using the recently collected data(see [1, Table III]), which was used in the second part of the experiment from to November.
In this section, we present results relevant to the reserve scheduler. Two mainfactors that determine the amount of reserves are the building’s energy capacityand the symmetry of reserve capacity. Apart from the physical properties of thebuilding, the energy capacity depends also on the comfort zone’s width. In thisexperiment, we specifically address the effect of enlarging the comfort zone duringunoccupied hours to − ◦ C (the so-called night setback). We performed sixfull-day experiments with symmetric (equal up- and down-reserves) or asymmetric5 R e s e r v e c apa c i t y ( i n % o f f an r a t ed po w e r) Symmetric reserves with setback on 15 November 2015
Up−reservesDown−reserves3.59% R e s e r v e c apa c i t y ( i n % o f f an r a t ed po w e r) Symmetric reserves w/o setback on 17 November 2015
Up−reserves Down−reserves
Figure 5: The hourly reserve capacities as a percentage of the nominal fan powerfor November (left) and November (right).reserves, and with or without night setback. Note that the same price is assumedfor up- and down-reserves in the asymmetric case.Fig. 4 shows the results for November when symmetric reserve capacitieswere assumed and for November when asymmetric capacities were used (inboth days night setback was applied). The capacities are reported in % of the fanrated power ( W). The reserve capacity is maximized at night when the com-fort zone is enlarged, and during the hottest part of daytime. In case of symmetricreserves, the maximum capacity is slightly less than of the rated fan power. Incase of asymmetric reserves, the maximum up-reserve capacity is approximately , whereas the down-reserve capacity is more than . These experimental re-sults are in agreement with relevant simulation results in [2,3], and show that down-reserves (consumption increase) are preferable for commercial buildings equippedwith energy-efficient controllers, because down-reserves can be provided withoutincreasing the baseline consumption and the energy cost.Fig. 5 compares the experimental results of November when setback wasused with those of November when no setback was applied (the reserve capaci-ties were symmetric in both dates). With setback most reserve is provided at night,whereas without setback the reserve provision coincides with the highest coolingload in the middle of the day. Although the experiment was conducted with setbackand symmetric reserves both on and November, the capacity profiles duringdaytime are considerably different due to different weather conditions and buildingmodels.To have a fair comparison under the same external conditions, we simulatedthe reserve capacity scheduling for all combinations of symmetry and setback us-ing the building model. The simulation and experimental results are shown inTable 1. The capacity ranges from low values below to high values nearly , and it heavily depends on reserve symmetry, setback, and weather condi-tions. The night setback increases the capacity by . on average for sym-metric reserves, by . for asymmetric up-reserves, and by . for asym-metric down-reserves. If setback is already used, adopting asymmetric capacitiesinstead of symmetric capacities reduces the up-reserves by . but increases the6able 1: Experimental (bold) and simulated (normal font) daily average reservecapacities in % of fan nominal power ( W).Date Symmetric,Setback Asymmetric,Setback Symmetric,No setback Asymmetric,No setback R u R d R u R d R u R d R u R d / .
61 15 . .
45 26 .
15 1 .
91 1 .
91 1 .
64 4 . /
11 9 .
09 9 . .
90 14 . .
74 0 .
74 0 .
90 1 . /
11 11 .
44 11 .
44 5 .
39 21 . .
24 2 . .
85 4 . /
11 16 .
70 16 .
70 7 .
78 31 .
75 3 .
94 3 . .
28 9 . / .
95 28 . .
89 49 .
66 15 .
07 15 .
07 7 .
82 28 . /
11 22 .
10 22 . .
72 46 . .
60 13 .
60 6 .
79 26 . down-reserves by . , and so the net effect is an increase of . in the totalcapacity. Experimental results for November are shown in Fig. 6, where the top plotshows the temperature trajectories in cells 1A and 1B, the middle plot presents theforecasts and actual values for ambient temperature and solar irradiance , whereasthe bottom plot shows the SAT and the air flow rate in cell 1A. The comfort zoneis indicated with red: the actual upper limit (red solid line) is ◦ C, but a tighterlimit of ◦ C (red dashed line) is used within the MPC to account for modelingand prediction errors. Similar results for , , and November are shown inFigs. 7, 8, and 10, but without including the SAT and air flow rate plot due to spacelimitations.In Figs. 6 and 7 the cell 1B is under energy efficient operation and the temper-ature remains close to the upper limit of the comfort zone. On the other hand, inFigs. 8 and 10 the cell 1B is in a “regulation-ready” operation mode, namely theconsumption of the Heating, Ventilation and Air-Conditioning (HVAC) system isscheduled identically to cell 1A to allow reserve provision, but no regulation signalis received. For this reason, the temperature trajectories of the two cells are veryclose to each other for most of and November. The total global irradiance is shown, which includes the long-wave radiation losses from thebuilding envelope to the atmosphere, and it can be negative at night. This effect is known as nighttimeradiation cooling [4]. The discrepancies from . to . on November are due to the calibration differencesbetween the fan models of the two cells (see Section 2.2). The discrepancies from . to . on November are because of interruptions in the hierarchical control in cell 1B due to serverconnection timeout error from approximately . to . . When the server was unresponsive, R oo m t e m p e r a t u r e ( C ) Cell 1ACell 1B Comfort zoneMPC bounds00:00:00 03:00:00 06:00:00 09:00:00 12:00:00 15:00:00 18:00:00 21:00:00 00:00:008101214161820 A m b i e n t t e m p e r a t u r e ( C ) Actual temperatureForecasted temperature 2001000100200300400500600700 S o l a r i r r a d i a n ce ( W / m ) Actual ir radian ce Forecasted ir radian ce A i r f l o w r a t e ( c f m ) Air flow rate 15.516.016.517.017.518.018.519.0 S upp l y a i r t e m p e r a t u r e ( C ) Supply air temperature
Cell 1A
Figure 6: Experimental results for the room climate controller under frequencyregulation on November (symmetric reserves, without night setback).The temperature trajectory of cell 1B in Figs. 6 and 7 remains mostly in theband [24 − ◦ C, which illustrates the necessity of tightening the comfort zoneconstraints in the MPC to compensate for modeling errors. The temperature trajec-tory of cell 1A is more variable and it follows the scheduled reserve and air flowrate. On November (Fig. 6) frequency regulation is provided while respectingthe comfort zone.However, on November (Fig. 7) the comfort zone is violated from . to . in cell 1A, but not in cell 1B. This happens because: (i) the ambienttemperature is higher than the day-ahead forecast from the beginning of the dayuntil . , and (ii) asymmetric reserves are used. The asymmetry allows for amore aggressive scheduling with a larger down-reserve capacity on Novemberin comparison with November when symmetric reserves are used (see Table 1).The control performance is significantly better on and November (Figs. 8and 10) despite the large discrepancies between the day-ahead ambient temperatureforecasts and the actual values. The improvement is due to the recently calibrated the cell was controlled by an existing fallback controller. R oo m t e m p e r a t u r e ( C ) Cell 1A Cell 1B Comfort zone MPC bounds00:00:00 03:00:00 06:00:00 09:00:00 12:00:00 15:00:00 18:00:00 21:00:00810121416182022 A m b i e n t t e m p e r a t u r e ( C ) Actual temperatureForecasted temperature 2001000100200300400500600 S o l a r i r r a d i a n ce ( W / m ) Actual ir radian ce Forecasted ir radian ce A i r f l o w r a t e ( c f m ) Air flow rate 15.516.016.517.017.518.018.519.019.5 S upp l y a i r t e m p e r a t u r e ( C ) Supply air temperature
Cell 1A
Time
Figure 7: Experimental results for the room climate controller under frequencyregulation on November (asymmetric reserves, without night setback).building model (see Section 2.3). No comfort zone violations occur and more-over the temperature is below the MPC constraint of ◦ C for most of the time.Therefore, periodic model calibration (for example on a weekly or daily basis) isimportant to account for seasonality and eliminate systematic errors.These results show that if the model and weather predictions are sufficientlyaccurate, the robust reserve scheduler allows a commercial building to bid in day-ahead markets for frequency regulation. On the other hand, if the modeling andprediction errors exceed the controller’s robustness margin, reserve provision forfrequency regulation might have an adverse effect on occupant comfort.The temperature trajectory in Figs. 8 and 10 is typical for a building with anight setback. The controller chooses to overcool the space at night in order togenerate higher revenue by offering a larger reserve capacity. In contrast, the re-serve capacity is smaller during working hours and the room temperature is higher.A comparison of the temperature trajectories in cells 1A and 1B shows that track-ing the RegD signal has little effect on room temperature due to the signal’s limitedenergy content.
Fig. 9 compares the out-of-sample performance of the older model (left plot) andthe new model (right plot). The blue curve is the measured room temperature,the green curve is the estimated temperature with the Kalman filter, the red curvecorresponds to a day-ahead model prediction, whereas the orange curve shows thestep-ahead temperature predictions. Clearly, the new model outperforms the older9 R oo m t e m p e r a t u r e ( C ) Cell 1ACell 1BComfort zoneMPC bounds00:00:00 03:00:00 06:00:00 09:00:00 12:00:00 15:00:00 18:00:00 21:00:0081012141618202224 A m b i e n t t e m p e r a t u r e ( C ) Actual temperatureForecasted temperature 2001000100200300400500600 S o l a r i r r a d i a n ce ( W / m ) Actual ir radian ce Forecasted ir radian ce A i r f l o w r a t e ( c f m ) Air flow rate 15.516.016.517.017.518.018.519.019.520.0 S upp l y a i r t e m p e r a t u r e ( C ) Supply air temperature
Cell 1A
Time
Figure 8: Experimental results for the room climate controller under frequencyregulation on November (symmetric reserves, with night setback). R oo m t e m pe r a t u r e ( C )
17 November R oo m t e m pe r a t u r e ( C )
20 November
MeasurementKalman filter1−day model pred.1−step model pred.
Figure 9: Model and Kalman filter performance. Left: Results for Novemberwith the older model. Right: Results for November with the new model.one, especially for the day-ahead predictions. This is why the performance of thelevel controller is much better on November than on November in terms ofcomfort zone violations.The effect of model accuracy on MPC operation is shown in Fig. 11. On November the model mismatch is large, which results in a significant discrepancyin the scheduled air flow rate and fan power between levels and . The MPCreacts on the modeling error by reducing the cooling power in level during nighthours and increasing it during daytime. In this way, the MPC provides the sameamount of electric reserve in daytime with less change in air flow rate by takingadvantage of the nonlinear fan curve. On the other hand, the model mismatch issmall on November and so the air flow and fan power schedules of level andlevel are similar. In fact, level consistently schedules less cooling power thanlevel because the air flow constraints are relaxed [1, Equations 15, 24], and the10 R oo m t e m p e r a t u r e ( C ) Cell 1ACell 1BComfort zoneMPC bounds00:00:00 03:00:00 06:00:00 09:00:00 12:00:00 15:00:00 18:00:00 21:00:00810121416182022 A m b i e n t t e m p e r a t u r e ( C ) Actual temperatureForecasted temperature 2001000100200300400500600700 S o l a r i r r a d i a n ce ( W / m ) Actual ir radian ce Forecasted ir radian ce A i r f l o w r a t e ( c f m ) Air flow rate 15161718192021 S upp l y a i r t e m p e r a t u r e ( C ) Supply air temperature
Cell 1A
Time
Figure 10: Experimental results for the room climate controller under frequencyregulation on November (asymmetric reserves, with night setback). A i r f l o w r a t e ( c f m ) F an po w e r ( W )
17 Nov − Level 1 17 Nov − Level 2 20 Nov − Level 1 20 Nov − Level 2
Figure 11: Left: Air flow rate schedule in level and level . Right: Fan powerschedule in level and level .reserve scheduling in level is robust and thus conservative. We present results on the dependence of SAT and cooling valve opening on fanspeed in Fig. 12, where the blue points are measurements and the red trend is apolynomial fit on them. As expected, the trend in cooling valve opening is increas-ing because the higher the fan speed the more cooling is required from the chilledwater loop. The trend in SAT is a flat line for fan speeds up to . However,for speeds above (and especially above ) there is a clear increasing SATtrend despite the increased cooling valve opening.These results lead to an interesting observation: the heat gain due to the rotationof the fan is significant at high speeds and it cannot be effectively rejected by ex-11 S upp l y a i r t e m p e r a t u r e ( C ) MeasurementsPolynomial fit 10 20 30 40 50 60 70 80 90Fan speed (%)020406080100 C oo li ng v a l v e op e n i ng ( % ) MeasurementsPolynomial fit
Figure 12: The dependence of SAT and cooling valve opening on fan speed.changing heat with the chilled water loop. According to Fig. 12, if the fan operatesat a speed or higher, the SAT will likely have a steady-state deviation from thesetpoint ◦ C that can be as high as ◦ C. Steady-state SAT deviations might resultin comfort zone violations, because the controller assumes the SAT fixed to ◦ C.This did not create problems in our experiment because the scheduled fan speed bythe MPC was at most . A major concern when providing Ancillary Services (AS) with commercial build-ings is the effect on energy consumption. Ref. [5] reported a round-trip efficiencyof when a building responded to demand response events in an experiment.There are two types of efficiency losses relevant to frequency regulation: “reserveavailability efficiency loss” and “reserve utilization efficiency loss” [6]. The firstone is the efficiency loss due to scheduling the consumption in an energy subopti-mal way to be able to provide frequency reserves, if requested. The second one isthe additional efficiency loss while tracking the frequency regulation signal.We report efficiency results in Table 2 for: (i) - November when the cell1B was under energy efficient operation (to quantify the reserve availability effi-ciency loss); and (ii) November when cell 1B was in regulation-ready operationmode (to quantify the reserve utilization efficiency loss). The efficiency loss is cal-culated comparing the energy consumption of cell 1A with that of the benchmarkcell 1B. We use two different definitions of energy consumption: (i) electric energyconsumption of the fan, and (ii) thermal cooling power consumption of each cell.The latter is calculated based on the chilled water flow rate ( ˙ m cw ), as well as thesupply ( T ch,s ) and return ( T ch,r ) chilled water temperatures using P cool = ˙ m cw · ( T ch,r − T ch,s ) . (1)Based on the results of Table 2, the reserve availability efficiency loss is sig-nificant and equal to approximately in terms of fan consumption and interms of cooling power from the chiller. However, the additional consumption incell 1A is not entirely wasted because it results in a lower average temperature.When cell 1B is in the regulation-ready mode, the cell 1A consumes less energy12able 2: Effect of frequency regulation on energy consumption1B operationmode Energy efficient / - / Regulation-ready / ( - h) / ( - h)Cell 1A 1B 1A 1B 1A 1BFan energy (kWh) .
22 16 .
23 19 .
24 20 .
91 9 .
40 10 . Cooling (gpm · F) .
92 1800 .
55 772 .
16 788 .
69 328 .
78 331 . Mean temp. ( ◦ C) .
85 24 .
43 21 .
45 21 .
26 21 .
05 21 . M P C c o m pu t a t i on t i m e ( s ) Symmetric reservesAsymmetric reserves
Figure 13: The average MPC computation time depending on the time of the dayand on reserve symmetry.than cell 1B despite frequency regulation. The non-negligible difference in theaverage temperature of the two cells is due to the imperfections in fan model cali-bration and the limited temperature sensor accuracy. However, even from . to . when both average cell temperatures are . ◦ C, the consumption of cell1A is still lower than that of cell 1B. This result indicates that the “reserve utiliza-tion efficiency loss” is negligible while tracking a fast-moving regulation signallike RegD.
The MPC computation time is sufficiently low for our demonstration. As shownin Fig. 13, the longest computation time is s for symmetric reserves and sfor asymmetric reserves. The computation time for asymmetric reserves is lowerbecause the problem is simpler and smaller [1, Section IV.D].The computation time decreases at the end of the day because a reducing MPChorizon is used. After the th time step, when the MPC prediction horizon issmaller than time steps ( . hours), the computation time is less than s. There-fore, Fig. 13 can be used to select the prediction horizon’s length depending on themaximum allowable computation time. Since the computation time grows expo-nentially with the number of variables of the nonlinear optimization problem, ashorter prediction horizon might be necessary for larger buildings.13 Regulation Signal Tracking (Level 3)
In this section, we present results from level and evaluate the tracking perfor-mance of the regulation signal. The following metrics are used e t ,k = e c ,k /P d ,k , e c ,k = P d ,k − P f ,k (2) e r ,k = (cid:40) e c ,k /R u ,k , if w k < e c ,k /R d ,k , if w k ≥ (3) e me = (1 /N exp ) · (cid:88) N exp − k =0 e c ,k (4) e mae = (1 /N exp ) · (cid:88) N exp − k =0 | e c ,k | (5) e rmse = (cid:114) (1 /N exp ) · (cid:88) N exp − k =0 e c ,k (6) e t,mape = (1 /N exp ) · (cid:88) N exp − k =0 | e t ,k | (7) e r,mape = (1 /N exp ) · (cid:88) N exp − k =0 | e r ,k | , (8)where P f ,k denotes the instantaneous fan power, w k denotes the normalized regula-tion signal, and N exp denotes the experiment duration. The metrics e t ,k and e r ,k arerelative instantaneous errors but the normalization is performed using the desiredfan power P d ,k in e t ,k , and the up- ( R u ,k ) or down-reserve capacity ( R d ,k ) in e r ,k .The mean error e me is used to measure any biases in the control response, whereas e mae is the Mean Absolute Error (MAE) during the experiment. The Root MeanSquared Error (RMSE) e rmse penalizes more the large control errors, for exampledue to overshoots and undershoots. The metric e t,mape is the tracking Mean Abso-lute Percentage Error (MAPE), and e r,mape is the reserve MAPE. We use the metric e r,mape because it describes the relative size of the control error with respect to thereserve capacity.In addition, we use the score proposed by PJM for evaluating the performanceof frequency regulation. The total score S tot consists of three parts, namely the cor-relation score S c , the delay score S d and the precision score S p , which are definedas [7] S c = max τ ∈ [0 , min ] ( R cor ) (9) S d = (cid:12)(cid:12)(cid:12) τ ∗ − min min (cid:12)(cid:12)(cid:12) , τ ∗ = argmax τ ∈ [0 , min ] ( R cor ) (10) S p = 1 − (1 /n ) · (cid:88) nk =1 (cid:12)(cid:12)(cid:12) e c ,k ¯ P d , h (cid:12)(cid:12)(cid:12) (11) S tot = (1 / · S c + (1 / · S d + (1 / · S p . (12)14he correlation score is the maximum correlation R cor of the desired power P d ,k and fan power P f ,k , and τ ∗ is the time shift at which the correlation is maxi-mized ( τ takes a value from to min with a step of s). We calculate the delayscore based on the time shift with maximum correlation. In the precision score cal-culation, we normalize the absolute control error by the average hourly value of thereference signal ¯ P d , h , whereas the total score is a weighted sum of the individualscores. In Fig. 14 we present results from the operation of level controller from 18.30to 19:30 on November . The duct pressure is quadratic to the fan speed,as expected from the fan laws. Since the duct system is designed to sustain thepressure corresponding to the maximum fan speed, and because the fan speed doesnot exceed its maximum value ( ) during frequency regulation, pressure con-straints were not necessary in the reserve scheduling and MPC formulations in ourexperiment.The RegD signal changes direction very often and has a limited energy content.During periods of time when the RegD signal is relatively flat, or the reserve capac-ity is low, the PI controller is active. On the other hand, whenever the changes infan power are rapid, the control switches to the model-based feedforward control.The tracking of the RegD signal is generally very good. However, when largerapid changes in fan power are requested, overshoots or undershoots might appear.In addition, if the reserve capacities change significantly at the beginning of eachfull hour, temporarily large errors might occur. In general, the instantaneous per-centage errors e t ,k and e r ,k are higher at a low operating fan power and low reservecapacity. The performance metrics (2)-(8) for the days of Table 1 are presented in Table 3. e r,mape is larger than e t,mape because small reserve capacities are offered for a largepart of the experiment. The mean error e me has a negative bias, which means thatthe fan power is more often higher than the desired setpoint because the controlovershoots are larger than the undershoots.We investigate the dependence of control performance on the minimum re-serve capacity, which we call “reserve threshold” and denote by R thr . The metrics e mae , e rmse , e t,mape and e r,mape are recalculated considering only the time steps when R u ,k ≥ R thr if w k < , and R d ,k ≥ R thr if w k ≥ . We repeat this procedure fordifferent R thr values and present the results in Fig. 15.In contrast to e t,mape , e r,mape decreases rapidly as R thr increases in the range [0 , W. This happens because (for the same absolute control error) e r ,k de-creases if R u ,k or R d ,k increase. On the other hand, e mae and e rmse generally in-crease as R thr increases because the higher the reserve capacity the larger the fan15 F a n s p ee d ( % ) A i r f l o w r a t e ( c f m ) D u c t p r e ss u r e ( p s i ) R e g D s i gn a l (-) R e s e r v e ca p ac it y ( W ) Up-reserveDown-reserve18:39:00 18:49:00 18:59:00 19:09:00 19:19:00 19:29:0001 P I ac ti v a ti on (-) F a n po w e r ( W ) Actual Setpoint18:39:00 18:49:00 18:59:00 19:09:00 19:19:00 19:29:00Time804004080 C on t r o l e rr o r ( % ) Tracking error Reserve error
Figure 14: Fan control and RegD signal tracking for a period of hour.power change, and thus the higher the errors due to the overshoots and under-shoots. Fig. 15 can provide us with bounds on the reserve capacity from a trackingperformance point of view.In Table 4 we report the PJM scores calculated separately for the period - November and the period - November. Different scores are calculated foreach hour (only if the reserve capacity is non-zero) [7], and the values in Table 4are hourly averages. The frequency regulation performance is exceptional duringthe whole experiment. For comparison, the highest possible total score is S tot = 1 and the minimum S tot accepted by PJM is . . The scores are slightly higherfor − November because the building provides a larger reserve capacitycompared with − November. 16
100 200 300 400 500 600 700 800 900Reserve threshold (W)2468 M A P E ( % ) Tracking MAPEReserve MAPE 0 100 200 300 400 500 600 700 800 900Reserve threshold (W)1020304050 T r ac k i ng e rr o r ( W ) MAERMSE
Figure 15: Dependence of tracking and reserve errors on the reserve threshold.Table 3: RegD tracking performance metrics during the experimentMetric e me e mae e rmse e t,mape e r,mape Value − . W . W . W .
58% 8 . % Tracking the fast-moving RegD signal introduces high frequency oscillations onSAT as shown in Fig. 16. The SAT of cell 1A (frequency regulation) oscillatesmore than that of 1B (regulation-ready mode), especially after sudden changes inthe regulation signal that induce sudden changes in the air flow rate. In addition,large excursions in SAT occur in both cells when the MPC changes the air flowsetpoint significantly, for example at hour . . Moreover, the magnitude of SAToscillations is high at low air flow rates, for example from . to . . The fan and the chiller are thermally coupled through the chilled water loop, hence,it is worth investigating if the operation of the chiller is affected while providingfrequency regulation with the fan. In Fig. 17 we present relevant experimentalresults for a duration of hours. The top plot shows the instantaneous and hourly-average electric power of the fan in cell 1A and the chiller. The bottom plot showsthe cooling power in the chilled water loop for cells 1A and 1B calculated with (1).The chiller has two stages and the electric power consumption is relativelyconstant at each stage. The chiller’s cycling depends on the cooling load, which inturn depends on the fan power and ambient conditions. In general, as the fan powerincreases the chiller cycles more often and remains longer at the on state. This isshown in Fig. 17 where the average chiller power (green line) generally follows theaverage fan power (black line).The effect of regulation is visible on the cooling power that fluctuates more incell 1A compared with cell 1B (regulation-ready mode). Whenever the fan powerincreases, the cooling load also increases and the SAT tends to decrease. This issensed by the SAT controller that opens the cooling valve to compensate for theSAT decrease, which in turn increases the cooling power in the chilled water loop.17able 4: PJM scores for tracking the RegD signalScore S c S d S p S tot − Nov. .
89 0 .
97 0 .
96 0 . − Nov. .
96 0 .
99 0 .
98 0 . A i r f l o w r a t e ( c f m ) Cell 1ACell 1B01:00:00 02:00:00 03:00:00 04:00:00 05:00:00 06:00:00 07:00:00 08:00:00 09:00:00 10:00:00 11:00:00 12:00:00Time15.516.517.518.519.5 S A T ( C ) Cell 1A Cell 1B
Figure 16: Air flow rate and SAT in cells 1A and 1B on November.The delay in cooling power response depends on the time constant of the coolingvalve’s controller.Despite the oscillations in cooling power, there is no observable effect onchiller’s cycling and electric power. This happens because: (i) the chilled wateris stored in a tank that provides some inertia; and (ii) the RegD signal is approx-imately zero-mean. Note that the gradual reduction in the hourly-average chillerelectric power from 19.00 to 00.00 in Fig. 17 is mainly the result of a lower cool-ing need due to ambient temperature drop, rather than a side-effect of frequencyregulation.These results indicate that frequency regulation can be provided with fan con-trol without side-effects on chiller consumption. However, this does not necessarilyhold for regulation signals with a larger energy content such as RegA. In addition,chillers with continuous rather than duty-cycle control will likely display a moreobservable impact on consumption while providing frequency regulation, espe-cially if there is no chilled water storage tank. In these cases, the level controllershould be revised, which is an interesting direction for future work. A challenge in this experiment was the communication delays in measurements andactuation, which result in the overshoots and undershoots in fan power in Fig. 14.In Fig. 18 we present a histogram of the experienced delays during the whole exper-iment. The probability distribution of the delay is positively skewed with a meanvalue of . s and a -percentile of . s. In fact, there exist a few very largedelays in excess of s due to temporary unresponsiveness of the CWS, which are18 E l ec t r i c po w e r ( W ) Fan (instant.)Fan (average) Chiller (instant.)Chiller (average)15:00:00 16:00:00 17:00:00 18:00:00 19:00:00 20:00:00 21:00:00 22:00:00 23:00:00 00:00:00Time0102030405060708090 C oo li ng po w e r ( gp m · F ) Cell 1A Cell 1B
Figure 17: Effect of frequency regulation on the chiller and cooling power. P r ob a b ilit y (-) Mean95% percentile
Figure 18: Histogram of communication delays during the experiment.not included in Fig. 18. Despite the fact that the average delay is large comparedwith the time step of level controller ( s), the tracking performance of RegDsignal is very good. Hierarchical control is an efficient way to provide frequency regulation with com-mercial buildings because time-separated tasks are considered individually. Threecontrol layers are essential: (i) a reserve capacity scheduler, (ii) a building climatecontroller to satisfy comfort while leaving enough slack for reserves, and (iii) acontroller to track the regulation signal.
Frequency regulation accuracy:
High-quality frequency regulation can be pro-vided by fan speed control. The RegD signal tracking is excellent even withlarge communication delays in the building automation system. A switched con-19roller comprised of a feedforward controller and a PI feedback controller with gainscheduling provides a fast response without compromising stability. This results ina total PJM score as high as . , which is well above PJM’s limit of . . Means to increase reserve capacity:
In our experiment, the fan provided . − . of its rated power as reserve capacity, depending on ambient conditions andreserve assumptions. Allowing asymmetric reserve capacities and using a nightsetback are effective ways to increase the reserve potential from commercial build-ings. In fact, down-reserves are preferable for buildings because the capacity canbe offered without increasing baseline energy consumption. Occupant comfort:
If the building bids in day-ahead AS markets, respectingoccupant comfort might be challenging if the building model and weather fore-casts are not very accurate. Furthermore, asymmetric reserves result in a moreaggressive scheduling that might increase comfort zone violations.
Building model:
An accurate building thermal model is essential for comfortsatisfaction, especially in day-ahead AS markets. Periodic calibration of the build-ing model helps to account for seasonality and eliminate offsets in modeling error.
Advantages of MPC:
Perhaps the most important advantage of MPC is thatit identifies the optimal balance between reserve provision and energy efficiency.MPC additionally provides us with a baseline consumption ahead of real-time op-eration, which is beneficial from a practical point of view. Moreover, due to itspredictive closed-loop nature it reacts to modeling and weather prediction errors ina way that minimizes occupant discomfort.
Robustness measures:
It is important to consider the regulation signal uncer-tainty when scheduling the reserve capacity. A conservative modeling of this un-certainty builds robustness to weather prediction and building modeling errors. Ad-ditional robustness can be obtained by tightening the comfort zone constraints inthe MPC, and allowing a larger fan speed control band in the MPC compared withthe reserve scheduler.
Effects of frequency regulation on building control:
Frequency regulation mightintroduce oscillations in SAT that can be reduced by appropriately tuning the cool-ing valve controller. In addition, if the MPC schedules the fan speed at very highvalues, the cooling loop might not be able to reject the additional heat gain due tofan rotation. On the positive side, there is little effect on the average energy con-sumption of the chiller while tracking an energy-constrained frequency regulationsignal by controlling the fan power. However, the impact of fan control on chillercycling may prevent the building from accurately following the regulation signalwhen measured against a baseline that includes the combined consumption of thefan and the chiller. This is an interesting area of further study.
Energy consumption:
Provision of frequency reserves entails some efficiencyloss. The efficiency loss due to scheduling the HVAC consumption in a suboptimalway compared with an energy efficient building control can be as high as . Onthe other hand, the efficiency loss while tracking frequency regulation signals withlimited energy content is negligible. 20 .2 Outlook
There are several avenues for follow-up work. Two direct extensions are to repeatthe experiment with the RegA signal of PJM, which is slower but has more energycontent, and/or with the heating loop of the AHU enabled. In addition, performingthe frequency regulation experiment using all four buildings of FLEXLAB willleverage the full potential of hierarchical control and verify the scalability of theapproach.In some HVAC systems a duct pressure controller regulates the pressure toa fixed setpoint. The combined operation of this controller and the dampers ofeach zone might reject the frequency regulation action [8]. This is an importantchallenge that could not be addressed in this experiment at FLEXLAB as it requirestesting on a large building.The reserve scheduling optimization problem might be computationally heavyfor buildings with many zones. An alternative is to approximate the nonlinear fanpower curve with a piecewise affine function by introducing binary variables. Thebilinear building dynamics can be approximated with sequential convex optimiza-tion [9], but the convergence is not guaranteed. Finally, the conservatism of thereserve scheduling problem can be reduced by generating scenarios from historicalfrequency regulation signals at the cost of reducing robustness.
In Part II of this two-part paper, we reported experimental results for frequencyregulation from a commercial building test facility (FLEXLAB). The results arevery encouraging: the test building can track fast-moving signals such as RegDreliably, with very high accuracy, and with minimal effect on occupant comfortand the operation of the HVAC system. The results also indicate that a hierarchicalcontrol approach is appropriate for frequency regulation with day-ahead bidding ofthe reserve capacity, and it can be used in field tests and real-world implementationsin larger buildings.
References ∼∼