1、4739 Verification of a Neural Network-Based Controller for Commercial Ice Storage Systems Darrell D. Massie, Ph.D., P.E. Member ASHRAE Jan F. Kreider, Ph.D., P.E. Member ASHRAE Peter S. Curtiss, Ph.D. Member ASHRAE ABSTRACT Thispaper describes the validation andperformance ofan optimal neural networ
2、k-based controller for an ice thermal storage system. The controller self-learns equipment responses to the environment and then determines the control settings that should be used. As such, there is minimal need to calibrate the controller to installed equipment. Results are verijed using computer
3、simulation as well as with the opera- tion of a full-scale HVAC laboratoy. These results demon- strate the robustness of a neural network-based controller and its ability to develop an optimal solution with minimal human interaction. INTRODUCTION Massie et al. (2004) developed a neural network-based
4、 optimal controller for commercial ice thermal storage systems. The controller consists of four neural networks, three of which map equipment behavior (Massie et al. 1998) and one that acts as a global controller. When combined, these networks self-calibrate to model installed cooling plant equip- m
5、ent and then determine the sequence of control actions that minimizes total cost over a planning window. This paper demonstrates the robustness of the neural network-based controller through computer simulation and through actual operation of a full-scale HVAC cooling plant with thermal energy stora
6、ge (TES). COMPUTER SIMULATION RESULTS A computer simulation was conducted to determine how the neural network-based controller handles a variety of price structures and to identify potential problems prior to operating the real plant. The two price structures investigated are tradi- tional with a de
7、mand charge but no ratchet and real-time pric- ing, where energy rates vary hourly, but without a demand charge. Assumptions The chilledice plant was modeled using trained neural networks as described by Massie et al. (1998). Results of this section are generated using the assumption that actual pla
8、nt operations behave exactly as the models predict and that the chiller and ice tank models have no associated error. This was done so that results could be compared with results of Henze et al. (1997). Although control of an actual plant will be infe- rior to the results shown here, this study prov
9、ides an indication as to whether or not the supervisory controller obtains an opti- mal solution for controlling TES-equipped cooling plants. The assumption is also made that future building loads and weather conditions are perfectly known in advance. The building load for all simulations is the sam
10、e and varies throughout the day (Figure 1). It is typical of office buildings, which, in general, have relatively constant cooling loads. Afternoon loads are slightly higher than morning loads, and loads during the first and last hour of the on-peak period (7:OO a.m. to 8:OO a.m. and 6:OOp.m. to 7:O
11、Op.m.) are reduced. There is no building load from 7:OO p.m. to 7:OO a.m. Traditional Rate Structures The first simulation tests the controllers ability to provide optimal control when placed under a traditional utility price structure. The price structure found in Equation 1 is a two- period struct
12、ure that consists of two time-of-day demand peri- ods plus time-of-day energy charges and no ratchet clause. If the ratios of on-peak to off-peak demand R, and energy Darrell Massie is director, Mechanical Engineering Research Center, and associate professor, United States Military Academy. Jan Krei
13、der is founding director, Joint Center for Energy Management (JCEM), and professor, University of Colorado, Boulder. Peter Curiss is owner, Curtiss Engineering, Boulder, Colo. 02004 ASHRAE. 471 35 , I 0000000 O000 000000080000 Zc?!C?OfOBM oem Hour of Day Figure I Building loadprojle. charges R, are
14、large, then the price structure is termed “strong,” and if the ratio is closer to unity, it is termed “weak.” ci 24 2 J = c c P(k)r,(k)At+ c PrnOX,” Yd,” (1) 1 k=I Y= 1 where p is the number of days in the month and k is the hour of each day. P(k) is the total power consumption due to the cooling an
15、d non-cooling load at hour k, re is the energy charge at hour k of the month, and At is the unit time step, which has been set to one hour in this study. Demand charges are computed by taking the product of the maximum power consumption Pmm, of the demand period v (typically 15 minutes) and the dema
16、nd rate rd, that is incurred during that hour of the month. Henze et al. (1 997), Kintner-Meyer and Emery (1 995), and Braun (1 992) all demonstrated that a strong price structure favors the formation of ice during the off-peak period for use during the on-peak period. Results shown here are based o
17、n a strong price structure where the utility rate had an on-peak demand and energy charge that was five times greater than during the off-peak period demand (i.e., R, = 5 and R, = 5). This rate structure was used because it is easy to determine if the controller is working as expected. The on-peak p
18、eriod ran from Monday through Friday, starting at 8:OO a.m. and ending at 7:OO p.m. The off-peak period encompasses all other hours including all hours on Saturday and Sunday. In order to compare results to those of Henze et al. (1997), a non-cooling load was omitted for this portion of the study. L
19、ater in the paper a non-cooling load example will be provided. Results of a weak price structure will be addressed below. Strong Price Incentive. This section demonstrates the optimal solution to a traditional price structure with a strong price incentive to shift the cooling load to the off-peak pe
20、riod. Since this is a well-studied price structure where the optimal solution is known, it is a good case for testing the algorithm. Maximum cost savings would be achieved if the ice tank were large enough to shift the entire cooling load. Since it is not, the optimal solution is one where the power
21、 consumption is leveled for both the on- and off-peak periods. As a benchmark for comparison, the cooling load shown would require a maxi- Hour of Day Figure 2 Optimal solution under a strong price incentive. mum on-peak demand charge in excess of 52 kW and would cost the building owner $700 per mon
22、th if thermal storage were not available. In analyzing the solution shown in Figure 2, ice storage is fully charged at the start of the on-peak period and fully discharged at the end, and the ice tank has, therefore, under- gone a full charge/discharge cycle. It is also interesting to note that afte
23、rnoon chiller loads are reduced, as power consumption remains constant. This is because the ASHRAE temperature model (ASHRAE 1997) assumes warmer afternoon outside air temperatures that lead to lower chiller efficiencies. Chiller power consumption is nearly constant during the on-peak period, has a
24、maximum of 27 kW, and is within one-half kilo- watt during all hours. The total cost is reduced by more than 50% from $700 to $342, due to storage in this example. In general the relative amount of cost reduction will vary with the size of the ice storage, building load, electric rate structure, and
25、 chiller size. For weekend days when there is no price incentive, ice is not made during the prior night and the chiller is operated so that it just meets the load. This is the least cost solution since there is a performance penalty for operating the chiller at low temperatures and no price incenti
26、ve to form ice. These results from the neural network-based controller are in complete agreement with the benchmark established by Henze et al. (1 997). Weak Price Incentive. A price incentive is termed weak when the on- and off-peak demand and energy ratios are near unity (i.e., Rd = 1 and Re = 1).
27、 Weak price structure incentives, in general, do not favor ice thermal storage (Braun 1992; Henze et al. 1997). This is due to decreased plant efficiencies associated with low evaporator temperatures required for making ice coupled with the small cost incentive. For this analysis the building load r
28、emains the same as that used for the strong incentive price structure, shown in Figure 1. In formulating a solution, the NN-based controller did not form ice during the off-peak period for use during the on-peak period. Had it done so, the overall cost would have 472 ASHRAE Transactions: Research O
29、.6 0.5 1 0000000 OO00000 ONpmzc O P 0 r Hour of Day Figure 3 Real-time pricing rate structure. I .I 1,1 I,I 888 moN r“ increased. The controller operated the chiller with the maxi- mum allowable setpoint temperature so that chiller efficiency was maximized. In determining the economic threshold for
30、using TES, it was found that a demand ratio (Rd N 1.3) and energy ratio of near 1.3 (i.e., Rd z 1.3, Re N 1.3) is the minimum ratio for economically using an optimally controlled TES system. These results are in agreement with Henze et al. (1997), Kint- ner-Meyer and Emery (1 995), and Braun (1 992)
31、. The precise threshold varies slightly depending on ambient conditions and chiller part-load conditions. In conclusion, the controller works as expected under both high and low traditional rate structures. If the utility company offers or does not offer an economic incentive, the neural network-bas
32、ed controller finds the cooling system setpoints that minimize total cost. Real-Time Pricing A common real-time pricing (RTP) structure consists of an hourly energy charge ($/kWh) with no demand charge. The cost to purchase electricity is adjusted according to supply and demand and is subject to cha
33、nge every day. Building owners are informed of the pricing structure one day prior to rates going into effect, allowing owners time to develop a strategy to offset cost during high-cost hours if desired. A ypical rate might look like that shown in Figure 3. In this rate structure purchase of nightti
34、me electricity is relatively inexpensive ($O.OS/kWh). During the daytime, electric rates increase considerably, in this case up to $O.O/kWh for three hours early in the afternoon. The hours between 2:OO p.m. and 4:OO p.m. will be referred to as peak hours. Other daytime hours have an associated ener
35、gy cost somewhere between night cost and peak cost. The building load profile is the same as that used for the traditional price structures. The optimal solution is shown in Figure 4 and cost is reduced by over 50% from $201 to $94 when compared to direct cooling with no storage. The solution shows
36、that the optimal trajectory is to increase the ice inven- tory during the night when rates are low. During the three peak Figure 4 Plant operating characteristics with real-time pricing. hours of the day when rates are highest, the chiller is turned off completely and all cooling comes from the ice
37、tank. Between low-cost and peak-cost hours, the chiller is operated such that the building load is met and storage is for the most part neither increased nor decreased. The exception to this is during the hours near the peak period (second most costly per kilowatt- hour) when the ice inventory is co
38、mpletely depleted. An inter- esting result is that the chiller model does not show power consumption to run fans and pumps during the three peak hours. This is an inaccuracy that is attributable to chiller model accuracy. Despite this minor discrepancy in equipment models, the controller has found t
39、he least cost solution to a complex problem. Non-Cooling Loads Non-cooling loads in many applications vary throughout the day. As such, they should be included when determining the peak demand. Up to this point, non-cooling electrical loads have been ignored so that results could be verified. The nu
40、mber of possible non-cooling load combinations is infinite. While TES systems cannot reduce the non-cooling loads, cooling plant equipment can be operated around peak non- cooling loads to limit total demand and cost. The following examples demonstrate that the NN control- ler operates the cooling p
41、lant so that overall demand is reduced for any situation. Two examples are presented here: (1) when there is a non-cooling load demand spike and (2) when the non-cooling load varies and is significantly larger than the cooling load. The utility rate structure and building load used for all non-cooli
42、ng load simulations is the same as that used for a strong traditional rate structure. The first case investigated has a one-hour 40 kW non- cooling load spike at 12:OO and a 20 kW spike at 13:OO. All other non-cooling loads are set to zero, In the absence of non- cooling loads, the utility rate stru
43、cture encourages full ice charging during the off-peak period. With the minimum on- peak demand rate (non-cooling load) set at 40 kW, however, there is little incentive for making ice during off-peak periods unless it can be used to reduce the total demand during the on- ASHRAE Transactions: Researc
44、h 473 . . - Building Load - - Chiller Power f -ke Charge Figure 5 Two hours ofnon-cooling demand load, 40 kWat 12:OO and 20 kWat 13:OO. peak period. In this example, demand reduction below 40 kW cannot be affected. Figure 5 shows the optimal solution to this problem. The 40 kW peak demand is not exc
45、eeded by completely turning off the chiller at 12:OO and operating it at partial capacity at 13:00, during the 20 kW demand spike. After 14:00, the chiller carries the bulk of the cooling load while ice storage is used to a much lesser extent. An ice charge of 70% was required SO that apower consump
46、tion of40 kW was not exceeded. Forma- tion of more ice during the off-peak period would have increased the off-peak demand and less ice would have raised the on-peak demand. The second non-cooling-load case consisted of a rela- tively large non-cooling load when compared to the cooling power consump
47、tion. Cooling loads and rate structures are the same as in the last example. Power profiles for the solution to this problem are shown in Figure 6. We observe that the total power consumption is 200 kW during both the on- and off- peak periods. The utility rate structure encourages the forma- tion o
48、f ice so that the on-peak demand can be reduced. To accomplish this, the chiller setpoint temperature is lowered to the maximum possible extent (26F) between 22:OO and 7:OO. During these hours, charging the ice tank does not increase overall demand. However, the tank cannot be fully charged in such
49、a short period due to the limited rate at which the ice tank can accept a charge. Therefore, charging also occurs from 18:OO to 22:00, increasing overall off-peak demand. From 7:OO to 12:OO the chiller operated such that it provides cooling for the full building load and the ice tank is neither charged nor discharged. The ice tank is discharged between 13:OO and 1 8:OO and the chiller operates at part load with reduced power consumption and overall demand reduction. LABORATORY RESULTS The above simulations (and many others) demonstrate the robustness of an NN controller for
