Prediction of Long Periods Heating Demand at Small Time Intervals for a Single Story Building using a Black Box Method

Abstract

Machine learning algorithms have been extensively implemented in building physics to predict demand profiles of different variables such as heat input or electric power. Most applications focus on consecutive time periods with no seasonal variations. This study relies on a black box method to predict the heat demand profile for a single story building based on data from different seasons. The models produced satisfying results. The Neural Networks (NN) models in general, produced higher accuracy, however, at higher computational cost compared to the Support Vector Regression (SVR) models. The multi-hidden layer NN's led to over fitting and deficiency. When the target variable is near constant, the prediction accuracy of the models substantially decreases. The input temperature profiles have higher influence on the accuracy of the prediction, up to 6 times maximum, in comparison of the other variables.