Abstract
Heating and cooling degree hours (HDH and CDH) are weather-based technical indexes designed to describe the need for energy requirements of buildings. Their calculation is the simplest method to estimate energy demand, providing the pattern of internal temperature variations in a building in response to weather conditions. The aim of the study is HDH and CDH prediction for Wrocław, Poland, based on outdoor air temperature using machine learning methods: artificial neural networks and support vector regression (ANN and SVR). The key issues raise in the study are: a detailed analysis of the most significant temperature lags (from 1 to 24 past hours) serving as predictors for modelling and an assessment of the impact of the database clustering on its accuracy. The best results are obtained with the clustering approach. The best predictor is the outdoor temperature observed 1 and 24 h before forecast demand (R(2) = 0.981 and 0.904 for heating degree and cooling degree hours indices, respectively). Models with the highest quality are created using ANN, and the lowest with SVR. Prediction of heating/cooling degree hour indices provides building demand in advance, does not require knowledge about its characteristics, and expresses the possible impact of regional climate modifications.