Paper Conference

Proceedings of eSim 2020: 11th Conference of IBPSA-Canada


Forecasting building energy performance using machine learning methods: To what extent should we trust it?

Mahnameh Taheri, Colin Parry, Agnieszka Hermanowicz, Alan Wegienka
arbnco Ltd., United Kingdom

Abstract: Rapid and reliable energy performance predictions using building performance simulations (BPS) has been one of the main concerns of the building science community. Machine learning techniques use nonlinear regression models that are trained to quickly approximate the building performance. This contribution explores the utility of dilated convolutional neural networks (dCNNs) for forecasting time-series of energy data. Energy and weather data are used for model training. The performance of dCNNs is examined in terms of efficiency and accuracy for forecasting time series of data. This paper explores whether, for certain applications, similar techniques could be used as an alternative to BPS.