Transformer networks for short-term forecasting of electricity prosumption

Abstract

Efficient energy management is accomplished by ensuring sustainable production of power and usage with minimized energy wastage. Electricity prosumption refers to electricity production and consumption. The forecasting of electricity prosumption is a complex problem as it is to model non-linear phenomenon. A deep neural network provides powerful algorithms for tackling tasks like time-series forecasting. However, there are still challenges in memory and computation for long sequence time series. The transformer network, specifically the informer model, was proposed to forecast electricity prosumption in various prediction horizons. The experiments were performed based on real datasets obtained from LUT University. In most cases, the forecasting error of the proposed method achieved much better accuracy compared with the naive methods in various prediction horizons, and the performance was relatively consistent with the growth of the forecasting prediction horizon.