Modeling and predicting buildingʹs energy use with artificial neural networks: Methods and results

This paper discusses how neural networks, applied to predict energy consumption in buildings, can advantageously be improved, guided by statistical procedures, such as hypothesis testing, information criteria and cross validation. Recent literature has provided evidence that such methods, commonly used independently, when exploited together, can improve the selection and estimation of neural models. We use such an approach to design feed forward neural networks for modeling energy use and predicting hourly load profiles, where both the relevance of input variables and the number of free parameters are systematically treated. The model building process is divided in three parts: (a) the identification of all potential relevant input, (b) the selection of hidden units for this preliminary set of inputs, through an additive phase and (c) the remove of irrelevant inputs and useless hidden units through a subtractive phase. The predictive performance of short term predictors is also examined with regard to prediction horizon. A comparison of the predictive ability of a single-step predictor iteratively used to predict 24 h ahead and a 24-step independently designed predictor is presented. The performance of the developed models and predictors was evaluated using two different data sets, the energy use data of the Energy Prediction Shootout I contest, and of an office building, located in Athens. The results show that statistical analysis as an integral part of neural models, gives a valuable tool to design simple, yet efficient neural models for building energy applications.