Improved day ahead heating demand forecasting by online correction methods

Abstract

To reduce the heating and cooling energy demand of buildings and districts novel control strategies are constantly being developed that require information on the future demand of the controlled entity. Demand forecasting is commonly done with deterministic white box models or fitted grey-box models, however, recently more and more data and machine learning based approaches are being developed. All approaches have weaknesses: white-box models require major modelling effort, grey-box approaches are limited by their model or parameter complexity and machine learning is dependent on hyperparameters, some of which are randomly chosen, and therefore considered unreliable. Here we develop a forecasting approach based on Artificial Neural Networks (ANN) and introduce error correction methods based on online learning and the learned autocorrelation of the forecasting error. We compare the approach to other regression based and grey-box methods in a real case study of a small-scale district energy system with mixed use and unknown lower-level control. We show that the proposed method outperforms the other forecasting methods in terms of average error and coefficient of determination. We further demonstrate that in our case study the error correction methods significantly reduce variance in ANN performance created by randomly initialized parameters in the networks.