Improving generalization of artificial neural network model for thermal load prediction

Thermal load prediction is essential for optimal operations of heating, ventilation, and air conditioning (HVAC) systems. Usually, the building thermal load is predicted by using artificial neural network (ANN) model based on environmental input variables. Unfortunately, it is not obvious that how many the input items should be or what preprocessing of inputs are best, which can cause significant overfitting and hurt ANN performance. The artificial neural networks existed for thermal load prediction has poor generalization ability. Two methods for improving generalization of ANN are introduced in this paper, which are correlation analysis of the historical data and principal component analysis of input data. ANN input items can be determined reasonably by correlation analysis of the historical data. And the dimension of ANN model will be reduced by principal component analysis. Using the two methods, ANN performance will be better than before.