Home Energy Management using End-to-End Learning

Home Energy Management Systems (HEMS) are used to control devices in households to improve among others efficiency, costs, and comfort. This paper investigates the use of an end-to-end learning pipeline to operate a HEMS that controls a battery to minimize costs. The end-to-end pipeline optimizes forecasting neural networks to minimize the objective of a downstream linear program. The end-to-end pipeline is compared against a classical pipeline which is trained solely to minimize forecast error, as well as against the optimal performance computed assuming perfect forecasts. The result of this comparison shows that end-to-end learning reduces the gap between traditional forecasting and the optimal performance using perfect forecasts by 9.25%, while also being more stable. This demonstrates the potential of end-to-end learning in the context of Home Energy Management.