Utilizing Hybrid Feature Extraction, Tree-Based Learning and Data-Lightweight Model Re-Training for Accurate Short-Term and Day-Ahead Residential Load Forecasting
Residential load forecasting is one of the most important tasks of the overall supply management process in electrical grids, since it enables smart grid services such as demand response (DR). Hence, several approaches for accurate residential load forecasting have been proposed in the relevant literature. However, most of the existing methods focus on the forecasting performance and neglect other aspects of the problem like training time and model size (i.e. memory usage). In this paper, we introduce a new model for both short-term and day-ahead residential load forecasting. The model synthesizes an heterogeneous feature set, which is constituted by both automatically-selected lagged values from the load time series and manually-extracted temporal features. Then, the tree-based algorithm light gradient boosting machine (LGBM) is fed with the constructed feature set and used as a regression model. Finally, a data-lightweight strategy is used for retraining the proposed model, which leads to both high forecasting accuracy and low training times. The proposed model has been extensively evaluated on a large real-world residential load dataset. The experimental results indicate that the proposed model achieves both higher forecasting performance and lower training times and model sizes compared to state-of-the-art solutions.