Neural-network-based Intelligent Control Of VAV Air-Conditioning Systems

Variable air volume (VAV) systems can reduce energy-consumption significantly and meet requirements of heat-moisture comfort. However, there are a lot of problems for controlling VAV systems. Firstly, there are several control loops in VAV systems and these loops couple with each other. In addition, control loops and equipment, such as actuators and transducers, have strong nonlinear characteristics, which make the control process more complex. Moreover, changing outdoor meteorological parameters and inside disturbances have a bad influence on VAV systems. As a result, it is very difficult for traditional PID control scheme to achieve good control performance or low energy consumption.A neural-network-based nonlinear predictive optimal control system was developed in this dissertation. A multi-layer forward neural network acted as the optimal feedback controller, which was trained with optimization algorithm based on the Hamilton-Jacobi-Bellman(HJB)and Euler-Lagrange(EL)equations as well as multi-step predictive performance function. After been trained, the neural network controller can approximate the optimal feedback solution without complexities of computation and storage problems. During control period, other neural networks were used for predicting time-varying parameters.Simulation was made before experiments, and the results show that the scheme proposed can get desired control performance for large-lag nonlinear MIMO systems. Then experiments were carried out in the Building Automation Systems(BAS)Laboratory of Beijing University of Civil Engineering and Architecture. Advantech AMAM6500 network controllers were used to acquisition parameters, including temperature, humidity, solar radiation intensity, and so on, for the air-conditioning zone of BAS. On studying the generalization method of feed-forward Artificial Neural Networks(ANN), a predictive model of air-conditioning zone had been identified with a multi layer feed-forward ANN. Based on the established model, the algorithm was used to control the VAV air-conditioning system in practice, in which cost function integrated energy consumption and thermal comfortability.The results show that:1) As the predictive control scheme has characteristics of rolling optimization and feedback correction, it can handle many problems in control systems, such as model inaccuracy, non-linearity, time-varying uncertainty, and so on. The experimental results prove that this system still has good performance under disturbances. And under the same condition, the control parameters will fluctuate for a long time with PID algorithms. This fluctuation will cause frequent actuator movements. As a result, operation life of the actuators will be shortened.2) Because that the cost function integrated comfortability and energy consumption, the control system can save energy consumption more than 6% compared with PID control methods.3) The intelligent control algorithms based on neural networks were implemented with Advantech network controllers, in which the FLASH storage space occupied was only 1.1MB, and the process memory size is 1.3MB. This control scheme can be implemented with distributed control systems, such as FCS, without the complexities of computation and storage problems of conventional neural-network-based intelligent control algorithms.