Load Forecasting in Cellular Low-Voltage Grids for Day-Ahead Optimization Using Ensemble Learning

As the centralized energy system is transformed into a decentralized one, caused by the electrification of the transportation and heating sectors and the increasing integration of renewable energy, accurate forecasting of increasing loads becomes more and more important, especially in the context of predictive cellular grid management. We implement various novel forecasting models, including Long Short-Term Memory, Gated Recurrent Units, Categorical Boosting, Light Gradient Boosting Machine, and N-BEATS, in addition to integrating them using stacking ensemble learning with a novel dynamic voting meta-model. Dynamic voting is a method developed in this paper that is based on statistical methods to increase accuracy. Using standardized datasets from SimBench and current forecasts, we evaluate model performance using the MAE and RMSE. The results show that the ensemble model achieves superior accuracy for daily forecasts, improving the reliability of the forecasts. In contrast, the Long Short-Term Memory model achieves excellent results for 15-minute forecasts as it effectively captures temporal dependencies. This work demonstrates the effectiveness of ensemble learning methods in improving forecasting processes and addresses the challenges associated with the growing integration of renewable energy and increasing demand on the distribution grid.