Cross-Zonal Short-Term Load Forecasting in Resource-Constrained Distribution Grids: A Multi-Model Benchmarking Study

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Background: Accurate short-term load forecasting (STLF) is essential for modern power distribution, yet selecting an optimal predictive architecture across diverse feeder profiles remains a key challenge for local grid operators. Method: This paper presents a comprehensive multi-feeder comparative evaluation of machine learning architectures using high-resolution hourly load data spanning January 1, 2024 to April 30, 2026 collected across six distribution feeders from two major operational zones under Bangladesh Rural Electrification Board (BREB): Gazipur PBS-1 and Mymensingh PBS-2. The forecasting models used here are classical statistical baselines (ARIMA, SARIMA), Linear Regression, tree-based ensembles (Random Forest, XGBoost, LightGBM), and recurrent deep learning structures (LSTM, GRU). Advanced feature engineering—incorporating cyclical temporal encodings, meteorological parameters, and temporal lags—was applied to capture localized spatial-temporal load dynamics. Results: Experimental results demonstrate that tree-based ensemble models consistently outperform both traditional statistical baselines and recurrent neural networks across all six feeders. Specifically, Random Forest achieved the highest overall predictive performance, attaining a peak average network accuracy of 97.66% (with a lowest feeder MAPE of 0.83% on Sreepur-8 F2). Meanwhile, LightGBM demonstrated superior operational efficiency, achieving a comparable 97.44% average accuracy with training time of only 3.47 seconds and a minimal memory footprint of 1.2 MB. Conclusion: These findings highlight that tree ensembles combined with feature engineering offer an optimal balance of accuracy, scalability, and computational efficiency for real-time distribution grid operation across neighboring utility zones.