Optimizing Smart Grid Load Forecasting via a Hybrid Long Short-Term Memory-XGBoost Framework: Enhancing Accuracy, Robustness, and Energy Management

As renewable energy sources and distributed generation become more integrated into modern power systems, accurate short-term electricity load forecasting is increasingly critical for effective smart grid management. Most statistical techniques used in the analysis of time series models, conventional statistical models, often fail to account for temporal dependencies and inherent non-linear patterns found in real-world energy time series. Methods: To this end, merging the power of both the ML approaches, namely Long Short-Term Memory (LSTM) networks and XGBoost, into hybrid frameworks has become a powerful solution. This work aims to develop a new compound model of LSTM for time series pattern extraction from the temporal data and XGBoost for outstanding predictive performance. To assess the performance of the proposed model, we used the Elia Grid dataset from Belgium, which includes load data recorded every 15 min throughout 2022. Results: When compared to individual models, this hybrid approach outperformed them, achieving a Root Mean Square Error (RMSE) of 106.54 MW, a Mean Absolute Percentage Error (MAPE) of 1.18%, and a coefficient of determination (R 2 ) of 0.994. Discussion: In addition, this study implements an ensemble learning strategy by combining LSTM and XGBoost to improve prediction accuracy and robustness. An experimental attempt to integrate attention mechanisms was also conducted, but it did not enhance the performance and was therefore excluded from the final model. The results extend the literature on the development of fusion-based machine learning models for time series forecasting, and the future work of energy consumption analysis, anomaly detection, and resource allocation in SM grids.