Dynamic Ensemble Short-Term Load Forecasting for Residential Building with Low Data Availability: SARIMA vs. LSTM

The field of short-term load forecasting has been a growing area of research in the context of energy systems. The recent increased adoption of renewable energy sources and energy storage systems at the residential scale has led to increased importance of load forecasting, yet it is faced with practical problems such as low data availability. This study proposes a dynamic ensemble technique as a potential solution, which has been tested on two kinds of forecasting methods, namely SARIMA and LSTM. The proposed approach dynamically splits the input data into temporal subsets using the MOSUM statistical tool, thus keeping the approach universally applicable. To validate the results of the proposed methodology, tests were conducted on input data spanning from 40 to 120 days. The results were compared to those obtained using standard SARIMA and LSTM techniques, with the latter not employing the proposed methodology. The proposed ensemble approach improves the forecasting performance for weekdays by up to 5% RMSE and 7.5% MISE for low input days. For weekends and holidays, the ensemble LSTM approach showcases much better performance compared to ensemble SARIMA. Based on the experiments conducted a clear trend of diminishing advantage of the ensemble approach with an increase in the number of input days is observed, suggesting the advantage of the ensemble technique for forecasting with low input data.