Control of heating, ventilation and air conditioning system based on neural network

Summary form only given. The HVAC (heating ventilation and air conditioning) car unit may be considered as a very complex multiple-input multiple-output system with the basic target to make the car passengers feel comfortable. The system inputs are: air temperature, actual temperature in the passenger compartment, sun load, vehicle speed, desired temperature set by the driver or co-driver, air recirculation request and defrost request. The last two inputs are binary signals. The system outputs are defined by the actuators located in the HVAC unit: defrost door position (it controls the air distribution between the defrost registers and penal-feet registers), mode door position (it controls the air distribution between the penal and feet registers), blend door position that mixtures hot and cold air, recirculation door position that makes a difference between the fresh air input and recirculation and the last output is the blower speed. The system that has to be controlled is quite nonlinear and nonstationary and there are two important criteria that have to be fulfilled. The first of them is the passengers' comfort and this criterion is vague, hardly defined numerically. In order to acquire some training sets many test trips have been made. Car companies usually organize three test trips trying to cover extreme temperature and weather environments. One of them is during summer choosing some hot destinations in Spain, Portugal or North Africa. The second trip is during the winters with the destinations in North Europe and the third trip is the so-called 'mild' trip usually organized in Middle Europe. The set of characteristic training values that may be collected during these trips is usually not enough for training of neural networks. This paper presents an algorithm to design the simulator in order to generate enough data necessary to train the networks. These data must be consistent with the data collected during the trips. Also, the paper proposes an architecture and a training algorithm for the neural networks that control the actuators in the HVAC unit. The obtained results demonstrate the accuracy and the simplicity of the proposed solutions.