Improved day ahead heating demand forecasting by online correction methods

Abstract Novel control strategies to reduce the heating and cooling energy consumption of buildings and districts are constantly being developed. Control on higher system levels, for example demand side management, usually requires forecasts for the future energy demand of buildings or entire districts. Such forecasts can be done with Artificial Neural Networks. However, the prediction performance of Artificial Neural Networks suffers from high variance. This means that two parameter-wise identical networks fitted to the same training data set perform differently well in forecasting the testing set. Here, we use two correction methods, one based on the forecasting error autocorrelation, and one based on online learning, to obtain reliable forecasting models. The approach is tested in the frame of day-ahead sub-hourly heating demand forecasting in a case study of a complex building, which has properties of a district heating system. It is demonstrated that the methods significantly reduce variance in prediction performance and also increase average prediction accuracy. When compared to other grey-box and black-box forecasting models, the approach performs well.