Hybrid Attention-Based LSTM and XGBoost Model for Short-Term Residential Load Forecasting

This study introduces a forecasting framework that combines deep learning with ensemble techniques to enhance predictions of residential AC load. The devised approach merges an attention-augmented long-short-term memory (LSTM) network with an Extreme Gradient Boosting (XGBoost) model, effectively capturing long-term temporal dependencies alongside nonlinear residual patterns found in electrical load data. The attention mechanism enables the LSTM model to dynamically prioritize significant time steps, improving its predictive accuracy. The initial LSTM-based forecast is further refined using XGBoost to model complex residual patterns, enhancing the reliability of the short-term forecast. The model is evaluated on a real-world AC load data set with high-resolution $\mathbf{1 5}$-minute intervals over one year. Experimental results indicate that the hybrid approach outperforms standalone deep learning and boosting models in terms of mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE). The findings highlight the robustness of hybrid forecasting techniques and their potential for optimizing residential energy management and grid reliability.