1、4747 Neural-Based Air-Handling Unit for Indoor Relative Humidity and Temperature Control Q. Zhang S.C. Fok, PhD ABSTRACT Humidity is an importantfactor contributingto one s ther- mal sensation and comfort. It also afJects onespercepiion of the air quality. Air-conditioning systems often encounter hu
2、midityproblems at low cooling load. Thispaperproposes an intelligent controller for an air-handling unit to control the temperature while limiting the humidity below 70%. The proposed scheme is based on the back-propagation-through- time approach. It uses artiJicia1 neural networks to develop an emu
3、lator to learn on-line the plant dynamics and a controller to control the fan speed and chilled water valve opening in real time. The neural-based controller was implemented on an industrial air handler for performance validation purposes. The implementation results show that the intelligent control
4、ler could efectively control the temperature and humidity within the operating range investigated. The results also indicate the potential of intelligent controllers as practical alternatives for controlling nonlinear and complex air-conditioning systems. INTRODUCTION Indoor humidity is a major cons
5、ideration in the design of an air-conditioning system. It may not only cause discomfort, but it may also be associated with poor health. Addendum 62x to ASHRAE Standard 62-200 1 (ASHRAE 2003) has recom- mended 65% as the upper limit for the relative humidity for habitable spaces. A high humidity sit
6、uation normally happens when the sensible cooling load is low. To avoid such a situa- tion, Singapores Code ofpractice for Mechanical Ventilation and Air Conditioning in Buildings (Singapore 1999) requires an air-conditioning system to be designed such that the indoor relative humidity does not exce
7、ed 70% when the sensible cool- ing load is at half of its design value. Y.W. Wong Fellow ASHRAE T.Y. Bong, PhD Fellow ASHRAE Air-conditioning systems are designed to supply air at either a constant flow rate or a variable flow rate, hence the names constant-air-volume (CAV) system and variable-air-
8、volume (VAV) system. In both systems, the air is treated using a cooling coil andfor heating coil housed in an air-handling unit (AHU) where the fan to supply the air is located. In CAV systems, the indoor temperature is maintained constant by varying the rate of chilled water flowing through the co
9、oling coil. VAV systems perform favorably to meet part-load condi- tions (Cappellin 1997). McQuiston and Parker (1 994) had proposed that the water-side control technique be used in conjunction with VAV and face-and-bypass dampers. However, all these techniques can still lead to an indoor rela- tive
10、 humidity exceeding 70% at certain part-load operation. The failure of air-conditioning systems in regulating both temperature and relative humidity is associated with the controller design. For example, if the cooling load falls below its design value, the sensible capacity of the cooling coil has
11、to be reduced to maintain the indoor temperature, leading to an increase in the indoor humidity above its design value. This is a common occurrence in hot and humid regions such as Singapore. One way to overcome this problem is to overcool the supply air so that more moisture can be removed. The air
12、 is then reheated to the required temperature to maintain the indoor temperature and humidity. The disadvantage of this approach is that extra energy and equipment are needed in overcooling and reheating the air. Currently, most air-conditioning systems are controlled using the PIPID technique, whic
13、h optimizes and fixes the controller gains based on a design condition. Since most oper- ating conditions seldom exactly match the design condition, Qi Zhang is a research student, Y. W. Wong is an associate professor, and T.Y. Bong is an associate professorial fellow in the School of Mechanical and
14、 Production Engineering, Nanyang Technological University, Singapore. S. C. Fok is an associate professor on the Faculty of Engineering and Surveying, Universiy of Southern Queensland, Toowoomba, Australia. 02005 ASHRAE. 63 the performance of the controller may not be optimal during operation when t
15、here are changes in the operation parameters. In the past a few years, there has been a growing interest in the use of artificial intelligence approaches such as neural networks for the control of practical industrial processes. Neural networks have been investigated in many areas of heat- ing, vent
16、ilating, and air-conditioning (HVAC) systems. Correctly configured neural networks could learn to produce the required outputs even though the relationship between inputs and outputs is difficult to describe. This paper aims to develop a scheme to control the indoor air temperature, together with re
17、lative humidity, in a VAV system. The controller involves two feed-forward neural networks to regulate the temperature and limit relative humid- ity within 70%. The project was implemented in tropical Singapore, where only cooling operation is required. Two control variables are the cooling water va
18、lve opening and the fan speed (driven by a variable-speed driver, VSD). The neural-based controller was developed on a PC and imple- mented on an industrial AHU for performance validation purposes. AHU MODEL DEVELOPMENT The design of a controller is dependent on the dynamics of the plant to be regul
19、ated. Although an intelligent controller should be able to self-learn and adapt to the plant changes, a plant model is still needed to initiate the design of the intelli- gent controller. The habitable space under consideration is an office building with floor space of about 38 m x14 m x 2.6 m (124.
20、7 ft x 46 ft x 8.5 fi). The AHU involved is a VAV system (central chilled water system), a horizontal draw-through type with 78.64 kW cooling capacity. A building automation system provides monitoring and controls for the main mechanical and electrical services for the whole system. To develop the p
21、lant model, essential parameters were measured under different operating conditions after the orig- inal PI controller was disabled. The measured parameters include the responses of the indoor air temperature and relative humidity under different combinations of chilled water valve openings and fan
22、speeds. They are labeled asyl,y2, ul, and u2, respectively: yl is the air temperature (OC), y2 is the relative humidity (%), u1 is the two-way valve actuation opening signal, and u2 is the fan speed. The first two variables are the outputs to be controlled and the latter two constitute the control i
23、nputs. The valid range of u1 and u2 is from O to 100%. For ul, it corresponds to O to 100% opening of the cooling water valve. For u2, it is linearly associated with 25 to 40 Hz VSD output frequency. The lower limit meets the minimum requirements for the ventilation rate. To simplisl the model, the
24、dynamic interaction was treated as disturbance. The dampers in the VAV boxes were fixed based on their average values associated with normal operating conditions in a week before the data collection. The data collected were used to develop a simple model relating the plant input to output relationsh
25、ips. The AHU model parameters were identified based on the measured data. Eight hundred sets of data were initially collected at 30-second intervals. The data sets were analyzed to determine the optimum sampling rate without compromis- ing the sensors accuracies. A two-minute sampling interval was f
26、ound to be appropriate, and this sampling interval was used for all the subsequent data acquisition and implementa- tion. An auto regression with the exogenous (ARX) model was built from 200 data sets collected on a separate day. The mathematical model is -_- where yl, y2, ul, u2 are the mean values
27、 for yl, y2, ul, u2, respectively, andyIl , y2, ulI , and u2 are the deviations from their respective mean values: yl =yl -yl =yl -21.8“C (71.2F) y2=2-Y2=248% 1- 1= 1-52% - - 2 = u2 - 2= u2 - 51.6% k =time step el and e2 = white-noise terms Figure 1 shows a comparison of the actual data sets with th
28、ose obtained from the developed model. A simplified square waveform representing random binary input signals was used to excite the system. The results show that there is very little difference between the measured data and the model outputs. The model identified gives the deviations of the temper-
29、ature and humidity about the average conditions, which can fluctuate from day to day. To examine the effects of the model with respect to changing conditions, data sets collected on another day were used. Again, the inputs were operated in random binary mode. Table 1 gives some statistical compari-
30、son for the two-day data sets with the model outputs. The root- mean-square of the deviations of yl and y2 from the means based on the model prediction for the two separate days are similar. The differences in the mean values for the two days show that the ambient conditions and internal loading pat
31、terns on the two separate days are different. However, the prediction errors are consistent. NEURAL NETWORK CONTROLLER SCHEME: NEURAL NETWORK STRUCTURE The preliminary plant model developed in the previous section is used to design the neural network controller. Many different neural network archite
32、ctures are available for control applications. For example, Psaltis et al. (1988) proposed two methods referred to as generalized learning and specialized learning. Another notable neural control architecture utilizing 64 ASHRAE Transactions: Research Table I. Comparison of Two-Day Data Set Outcomes
33、 Parameters Number of samples yi mean value y2 mean value Outdoor air temp. mean value Outdoor air RH mean value Root-mean-square ofy, Root-mean-square ofy2 First Day Data Second Day Data 20 1 204 21.8“C (71.2“F) 21.4“C (70.5“F) 68.10% 68.30% 26.1“C (79.O“F) 29.1“C (84.4“F) 81.29% 75.31% 0.03“C (32.
34、1“F) 0.03“C (32.1“F) 0.2% 0.2% measured simulated (i-a)y 70% then wt = O and wh = 1. However, this definition would cause a discontinuity in the cost function. This is prohibited in the gradient training algorithm. To overcome this problem, wt and wh are modified so that the gradient is valid over t
35、he entire range, i.e., 1 Wh = 80 - 200y, 1 +e (3) 66 ASHRAE Transactions: Research 1.0, rz Figure 4a Emulator network on-line training error. All the parameters used in the network training are normalized, so the signal for both the valve opening and the fan speed can only vary from O to 1. Once the
36、 limit is hit, a zero is sent to inhibit further learning, and controller outputs must be squashed to within this range, i.e.: If u1 o* u2 1 thenul or2= 1. Simulations were used to test the controller performance before actual implementation. The simulated results indicated that the controller was f
37、easible for tracking changes in the setpoint. According to the preset control strategy, the control- ler could determine which input should be trained (i.e., u1 or uz), by comparing the humidity with its upper limit. Following the simulation, preliminary on-line investiga- tions were conducted to va
38、lidate the emulator on-line training. Real-time data from the AHU were fed into the computer through a data acquisition (DAQ) board. The AHU was under the regulation of the PI controller during this testing process. Five consecutive historical data sets were used for every emulator update, the rate
39、of which was at two minutes. Figure 4 shows the error result of the on-line emulator training. The results indicate that the emulator can success- fully predict the temperature and humidity based on the measured fan speed and the valve actuation signal. The emula- tor output error for temperature is
40、 within f0.2“C (+_32.36“F), and the error for humidity is within *OS% for most of the time. The emulator outputs match the system outputs well and the errors are gradually minimized as the learning progresses. The results demonstrate that the emulator network is able to predict the system outputs tw
41、o minutes in advance, Le., in one sampling interval, based on the on-line learning at every time step. The emulator will be linked with the controller network to form a complete neural network controller in the actual implementation. IMPLEMENTATION RESULTS AND DISCUSSIONS Implementation Results The
42、intelligent controller was realized on a PC. The DAQ board was responsible for data acquisition and control signal generation, while the neural networks were coded into matrix for calculation. In the actual implementation, the DAQ board 32. IS 5 32 OC E z E d . -321 -33 - Temp mor RII error Figure 4
43、b Emulator network on-line training error was hooked up to the direct digital control (DDC) box of the AHU system. The original PI controller of the DDC was disabled. The fan and the valve were fully controlled by the signals from the DAQ board. The measured temperature and humidity were fed into th
44、e PC for on-line network training. The training result was recorded before the outputs were sent out to modulate the valve opening and fan speed. Figure 5 shows a typical result of the performance of the controller with varying temperature setpoints and a constant upper limit of 70% for relative hum
45、idity. The actual valve opening (ui) and the fan speed (u2) determined by the control- ler network are also plotted. The initial setpoint was at 24C (75.2“F). The controller response was allowed to settle. The results indicated that the temperature settled to the desired setpoint with humidity well
46、below 70%. This corresponds to the region from the O to 22nd time step in Figure 5. The controller achieved these desired conditions by opening the cooling water valve at 55% while maintaining the fan speed at 42%. The valve opening and fan speed were comparable to the average operation values when
47、under PI control, Le., based on data collected on three separate days prior to implementation, under normal operation. To test the intelligent controller, the setpoint was changed to 23C (73.4“F) at the 22nd time step. The intelligent control- ler responded by fully opening the valve three sampling
48、inter- vals later. This is clearly the optimal approach to maximize the cooling capacity. As during these adjustments the humidity was well below 70%, the controller did not adjust the fan speed (this is similar to the operation of a CAV system). The valve remained fully opened for the next twenty t
49、ime steps until the temperature gradually decreased to 23OC (73.4“F). During this time, humidity dropped slowly to 58%. After the temperature had reached the new setpoint, the controller slowly reduced the valve opening to stabilize the temperature. This corre- sponds to the region from the 40th to the 67th time step in Figure 5. At the 67th time step, the temperature setpoint was suddenly raised from 23C (73.4“F) to 25C (77F). The intel- ligent controller responded by quickly closing the valve a time step later. This is again the optimal approach to increase
