Short-Term Industrial Power Load Forecasting Based on XGBoost and Multi-Dimensional Feature Engineering

Abstract

Industrial power load exhibits severe fluctuations with multi-scale seasonality coexisting with discrete operating states, making it difficult for traditional single models to accurately capture its complex dynamics. To address these issues, this paper proposes a short-term power load forecasting method based on eXtreme Gradient Boosting (XGBoost) and multi-dimensional feature engineering. By constructing a five-dimensional feature engineering framework that integrates temporal features, multi-scale lagged variables, rolling statistics, and electrical domain parameters, the time series forecasting problem is transformed into a structured supervised learning problem. Time-series-aware cross-validation and grid search strategies are employed to achieve systematic hyperparameter optimization. Validated on full-year 2018 operational data at 15-minute resolution from a steel enterprise, the optimized model achieves a Mean Absolute Percentage Error (MAPE) of 7.3% on the test set, representing a 48.6% improvement over the naive persistence baseline. Ablation experiments and feature importance analysis reveal the independent contributions of each feature category and the core driving factors of load variation, providing a practical solution that balances accuracy, efficiency, and interpretability for short-term load forecasting in industrial scenarios.