Probabilistic Forecasting of Household Loads: Effects of Distributed Energy Technologies on Forecast Quality

Abstract

Distributed energy technologies introduce new volatility to the edges of low voltage grids and increase the importance of short-term forecasting of electric loads at a granular level. To address this issue, first probabilistic forecasting models for residential loads have been developed in recent years. However, knowledge is lacking about how well these models perform for households with different endowments of distributed energy technologies. Therefore, we first create a new semi-synthetic data set which contains not only conventional residential loads, but net loads of 40 households differentiated regarding heating type (electric space heating, no electric space heating), and rooftop solar installation (solar, no solar). Second, we develop a novel probabilistic forecasting model based on Gated Recurrent Units that uses data from weather forecasts and calendar variables as external features. We apply the developed model, and three benchmarks, to the new data set and find that the GRU model outperforms the other models for households with electric heating, with solar, and with both technologies, but not for households without distributed energy technologies.