Empirical validation of a data-driven heating demand simulation with error correction methods

Abstract

Advanced control concepts for building energy systems, such as Model Predictive Control, often require models that forecast the energy demand of a building. Such models are commonly based on first principles, however the cost and effort required to develop such models may be prohibitive for real-life applications. As an alternative, we introduce and validate a data-driven simulation approach based on Artificial Neural Networks to forecast the heating demand of buildings. The forecast is enhanced with the help of two correction methods, based on online learning and forecast error auto-correlation. Validation results based on data from four office buildings suggest that our method shows better forecasting performance than a fitted 5R3C building model.