A Hybrid LSTM-XGBoost Model for Multi-Horizon Short-Term Load Forecasting in Smart Power Distribution Systems

Short-term load forecasting is essential for reliable, efficient, and cost-effective operation of modern power distribution systems. This paper addresses hourly load prediction by proposing a hybrid Machine Learning (ML) framework that integrates Long Short-Term Memory (LSTM) networks with eXtreme Gradient Boosting (XGBoost). The framework aims to improve forecasting accuracy across multiple horizons (1 h, 6 h, 12 h, and 24 h) using two years of real data from 1,200 smart meters in a medium-sized U.S. urban utility. The hybrid model combines seasonal trend decomposition (STL) with feature engineering, including weather, lag, and calendar features, to capture temporal dependencies and nonlinear feature interactions. Experimental results show that the hybrid model consistently outperforms baseline methods, including LSTM, XGBoost, Random Forest (RF), Support Vector Regression (SVR), and AutoRegressive Integrated Moving Average (ARIMA). At the 24-hour horizon, it achieved the lowest RMSE (1.93), MAE (1.62), and MAPE (4.3%), compared with LSTM (2.21, 1.85, 4.9%), XGBoost (2.35, 2.01, 5.2%), RF (2.42, 2.10, 5.5%), SVR (2.67, 2.35, 6.1%), and ARIMA (2.80, 2.50, 6.4%). RMSE increased gradually across horizons, but the hybrid model maintained the slowest error growth from 1.53 (1 h) to 1.93 (24 h), demonstrating resilience to long-term uncertainty. Statistical tests confirmed that all improvements were significant (p < 0.05). Seasonal validation confirmed robustness across Winter (1.83), Spring (1.87), Summer (1.95), and Autumn (1.89). Ablation studies showed that removing weather and lag features increased RMSE to 2.25, indicating these features together contributed approximately 23% to variance reduction. Overall, the proposed hybrid LSTM XGBoost model delivers superior accuracy, robustness, and statistical reliability. By combining LSTM’s sequential learning with XGBoost’s nonlinear feature optimization, the framework offers a scalable, generalizable solution for real-world short-term load forecasting, supporting improved demand planning, operational flexibility, and economic efficiency in smart grid applications.