1、Junya Nishiguchi, Tomohiro Konda and Ryota Dazai are research engineers in the Technology Development Headquarters, Yamatake Corporation, Japan. Adaptive Optimization Method for Energy Conservation in HVAC Systems Junya Nishiguchi Tomohiro Konda Ryota Dazai ABSTRACT As a way to reduce the energy con
2、sumption in the buildings HVAC system, a practical method for optimal control has been developed. Since the proposed method identifies the approximated objective function with the past operational data, it automatically optimizes the settings in accordance with the change in the outside environment
3、and building operation. As the result of a validation with an actual hotel building, we were able to demonstrate that the proposed method realized energy conservation by increasing the chilled water temperature appropriately. INTRODUCTION In a central air-conditioning system, the various HVAC facili
4、ties (such as chillers, cooling towers, and pumps) should be cooperatively controlled to satisfy the required chilling load with the minimum total energy consumption. However, most settings of individual facilities are usually not adjusted optimally but independently defined as the fixed design cond
5、itions. Thus, the optimal operation with the existing facilities has a benefit for the costumers due to unnecessary for large facility investments. As optimal operation methods in the HVAC systems, a vast amount of physical model based approach has been proposed (Braun et al. 1989, Ito et al. 2008,
6、Takagi et al. 2009, Low et al. 2009). These methods require the detailed simulation models describing the thermal and physical relationships between each facility, and the model parameters are adjusted with the facility specifications. If the constructed models are sufficiently accurate for describi
7、ng real phenomena, the physical model based approach may be an ideal method for energy conservation. However, since the HVAC system structure in the real buildings is widely varied, it usually takes a huge effort to identify such relationships even if recent modeling tools (Sugihara et al. 2007) are
8、 used. In addition, it is difficult to revise the model parameters on-line to adapt the current environmental conditions, system structure, and operations policy. As a result, most of the conventional methods have not been used in practice. Recently, the historical operational data in the HVAC syste
9、ms, which contain the energy consumption characteristics, have been easily accumulated with the Building and Energy Management System (BEMS). While these accumulated data are actively utilized for performance evaluation and improvement (Brambley et al. 2009), there has not been an on-line optimal co
10、ntrol method with the measured data. In this paper, we propose a data-driven optimal control method, which constructs and revises the model adaptively with the past operational data. In this method, model input variables are selected as related to energy consumption, such as chilled water and conden
11、ser water temperatures and flow rate, and chilling load; the output variable is the total energy consumption of the whole HVAC system. The optimal settings are derived from this model by determining at what settings total energy cost is minimized. Also, by successively adding input variables and ene
12、rgy consumption data to the model, the model is serially revised. Therefore, this method greatly simplifies the process of constructing and revising the model, which has been a stumbling block for conventional methods, and it becomes possible to optimally control HVAC systems in accordance with outs
13、ide environmental fluctuations and building operation changes. LV-11-C068 2011 ASHRAE 5492011. American Society of Heating, Refrigerating and Air-Conditioning Engineers, Inc. (www.ashrae.org). Published in ASHRAE Transactions, Volume 117, Part 1. For personal use only. Additional reproduction, distr
14、ibution, or transmission in either print or digital form is not permitted without ASHRAES prior written permission.METHOD An optimal control procedure is to find the setting that minimizes the objective function describing energy consumption characteristics. Our method identifies the approximated ob
15、jective function with the measured data as shown in Figure 1. In order to accumulate the large amount of the measured data as the compressed model, Topological Case-Based Modeling (TCBM) is used. The most prominent feature of the TCBM is to preserve the data itself as a model called “Case-Base” acco
16、rding to the given model precision. This feature is similar to that of the Just-In-Time modeling methods (Fujiwara et al. 2009), which have become recently popular in the industrial field, and leads to reduction of the computational cost of the model generation and revision. Thus the TCBM has been a
17、pplied to many applications (such as soft-sensing, performance monitoring (Tsutsui et al. 1996), and load prediction (Konda et al. 2001). In this method, if the objective system satisfies continuity on the input-output relationships, the Case-Base structure and similarity between each case are defin
18、ed only with the past data without any prior knowledge. As the Case-Base structure, the input space is discretized according to the given model precision. Each set of the input and output data is converted into a case corresponding to the discretized input space. When using model, the TCBM finds cas
19、es whose inputs are the most similar to the new inputs, and averages the outputs of the similar cases as the corresponding new output. After the real output is obtained, it simply adds the new case to the Case-Base in order to adapt the new situation. As an optimization method, the multivariate spli
20、ne approximation (Kaseda 2004) is used. It is a general approximation method for nonlinear objective functions with multidimensional input, and suitable for function approximation on real-world data because it has an advantage on smooth features for extrapolation. The spline approximation is more pr
21、actical method than other nonlinear approximation methods such as Artificial Neural Networks, in that the calculated result is independent with initial calculation conditions, and parameter adjusting is unnecessary. Once approximated objective function is identified, optimal setting is calculated by
22、 the simple gradient descend methods (such as quasi-Newtons method). Note that the resulting solution is one of local optima instead of global optimum. However, since the energy characteristics of the HVAC systems can be adequately approximated as a convex function (Ann et al. 2001), local optimal c
23、ondition is considered to have sufficient performance for the energy conservation. Chilled watertemperature CLoadGJTotalenergyWhNew dataAdaptive LearningChilled watertemperature CLoad GJOptimizationOptimum(a)(b)TotalenergyWhTotalenergyWhTotalenergyWhFigure 1 Overview of proposed optimal control. (a)
24、 Adaptive learning with Topological Case-Based Modeling. (b) Optimization with multivariate spline approximation. 550 ASHRAE TransactionsCASE STUDY Optimization of Chilled Water Temperature A chiller in the HVAC system usually produces chilled water at about 7C as a design condition, but if it suppl
25、ies the water at a higher temperature this will improve the Coefficient Of Performance (COP) and reduce the energy consumption for each chiller. Considering the HVAC system as a whole, however, if the temperature of the chilled water supply is increased, the cooling capacity of Air-Handling Units (A
26、HUs) decreases. As a result, the flow rate of chilled water, as well as the conveyance power in chilled water pumps, increases. There is thus a trade-off between chiller and chilled water pumps energy consumption as shown in Figure 2. Furthermore, since various external factors, such as outdoor temp
27、erature and chilling load, also affect the efficiency of the HVAC system, there is a need to find the optimal chilled water temperature setting for minimizing energy consumption by taking into account the trade-off between energy consumption of facilities, and the effect of external factors. Althoug
28、h optimizing the chilled water temperature setting is considered to be an effective way to HVAC energy conservation (ASHRAE 2007), it has been difficult to understand the complicated characteristics of chillers, coils, and pumps, which are widely varied among buildings. Cooling towersAHUsChillersChi
29、lled water pumpsCondenser water pumps Supply airChilled waterCondenser waterEnergyconsumptionChilled water temperaturePumps / FansChillersTotalOptimum(a) (b)EnergyconsumptionEnergyconsumptionFigure 2 Chilled water temperature optimization. (a)Process diagram of typical chilling system. (b)Trade-off
30、between chillers and pumps energy. Target Building The proposed method was applied to an optimal control of chilled water temperature at an actual hotel building. The target building has a total floor area of 15,000m2including guest rooms, banquet rooms, and restaurants. The chilling system contains
31、 several facilities as shown in Table 1, and is equipped with the two pump system, namely with the chilled water pumps and condenser water pumps as shown in Figure 3. The process data and energy consumption data around the chilling system are continuously monitored with Building Automation (BA) Syst
32、ems. In this case study, the only chilled water temperature setting of the inverter-driven turbo-refrigerating machine (R-1) (which is the primary chiller in this system) was optimally controlled. 2011 ASHRAE 551Table 1. Rated Output of Chilling / Heating Facilities Facilities Rated Output Inverter-
33、driven turbo-refrigerating machine (R-1) Chilling ; 250 USRT Absorption chiller/heater (R-2) Chilling ; 240 USRT, Heating ; 773kW Heat exchanger (HX-1) 313kW One-through boilers (B-1, B-2) 752kW R: ChillerHX: Heat exchangerCP: Chilled water pumpHP: Hot water pumpB: BoilerCT: Cooling towerCHP: Chille
34、d and hot water pumpCDP: Condenser water pumpFigure 3 HVAC process diagram of target building. Implementation The input variables should be selected to have strong relationship with total energy consumption according to the energy characteristics of each target building. In this case study, the mode
35、l input and output variables are selected with the statistical analysis as listed in Table 2. Note that, the chilling load is calculated with the chilled water flow rate and temperature difference between inlet and outlet chilled water. Table 2. Model Variables Input Variables Output Variable (1) Ch
36、illing Load Total energy consumption (2) Condenser water temperature ( = Chiller energy (3) Condenser water pump speed + Chilled water pump energy (4) Outside temperature + Condenser water pump energy (5) Chilled water temperature + Cooling tower fun energy ) Figure 4 illustrates the prototype syste
37、m employed for the case study. A laptop PC implemented with our optimal control method was connected to the existing BA network. Necessary input data was acquired in real time and this was then used for model revision and optimization calculations. As the pre-processing the moving average function w
38、as used to reduce the effect of the measurement noise, and the outliers were omitted with the pre-defined thresholds. In case that the outliers were found, the optimal temperature calculated last time was used. Finally, the optimal setting for the chiller was output via the BA system. The actual chi
39、lled water temperature is controlled by the local PID controller in the chiller such that it reaches the calculated setting. Note that, lower and upper limit of chilled water temperature setting are also defined, in order to avoid 552 ASHRAE Transactionsfreezing in the evaporator of the chiller and
40、provide adequate dehumidification for the room. ModelLANUniversal Integrated Controller (UIC)Collecting dataCentral monitoring systemDirect Digital Controller (DDC)PCChiller4-20mAOptimizationLearningOptimal chilled water temperature settingsFigure 4 Architecture of prototype system. RESULT Chilled W
41、ater Temperature Setting Figure 5 shows 1-day historical chilled water temperatures after the introduction of our method. From the graph, it can be seen that in the afternoon when AHUs operation is stable, raising the chilled water temperature setting enables the system to increase chiller COP and t
42、hus save energy. In contrast, in the morning when the AHUs are starting up, the supply air temperature becomes lower and as a result it appears that the chilled water temperature setting has also become lower. It represents the method specified optimal chilled water temperature settings according to
43、 the environmental conditions such as chilling load. g16097g16098g16099g16100g16101g16093g16092g16093g16093g16093g16094g16093g16095g16092g16093g16094g16095g16096g16097g16098g16099g16100g16101g16093g16092g16093g16093g16093g16094g16093g16095g16093g16096g16093g16097g16093g16098g16093g16099g16093g16100g
44、16093g16101g16094g16092g16094g16093g16094g16094g16094g16095g16135g16116g16155g16161g16158g16137g16135g16989g16137g16092g16092g16090g16097g16093g16093g16090g16097g16094g16094g16090g16097g16095g16095g16090g16097g16096g16135g16115g16118g16137g16111g16148g16149g16152g16152g16145g16144g16076g16163g16141g
45、16160g16145g16158g16076g16160g16145g16153g16156g16145g16158g16141g16160g16161g16158g16145 g16117g16154g16160g16145g16160g16076g16163g16141g16160g16145g16158g16076g16160g16145g16153g16156g16141g16160g16161g16158g16145 g16111g16148g16149g16152g16152g16145g16144g16076g16163g16141g16160g16145g16158g1607
46、6g16160g16145g16153g16156g16145g16158g16141g16160g16161g16158g16145g16076g16159g16145g16160g16160g16149g16154g16147 g16111g16148g16149g16152g16152g16149g16154g16147g16076g16120g16155g16141g16144Figure 5 Example of optimized water temperature. 2011 ASHRAE 553Adaptive Learning Effect The proposed meth
47、od revises the model function as new data are obtained. Figure 6 visualizes the examples of the characteristics of the data model before and after the adaptive learning (i.e. revising model with the historical data). The graphs show the relationships between chilled water temperature and total energ
48、y consumption, where the other input variables (such as chilling load, outside temperature and condenser water temperature) are fixed at typical values. The model functions before the adaptive learning are far from assumed convex function due to lack of model data at low chilled water temperature (a
49、round 6C). On the contrary, the model functions after the learning with the historical data show the near-convex shape, where each optimal value is found with the gradient decent methods. g16093g16097g16092g16093g16097g16097g16093g16098g16092g16093g16098g16097g16093g16099g16092g16093g16099g16097g16093g16100g16092g16093g16100g16097g16093g16101g16092g16097g16099g16101g16093g16093g16093g16095g16093g16097g16100g16092g16100g16097g16101g16092g16101g16097g16093g16092g16092g16093g16092g16097g16093g16093g16092g16093g16093g16097g16093g16094g16092g16097g16099g16101g16093g1
