Implementation of hybrid CNN-LSTM with HGB model for short-term load forecasting in local distribution utility

Abstract

Short-term electricity load forecasting provides anticipatory information for operation, energy procurement, and network-constraint assessment in distribution systems. This work develops and evaluates a hybrid methodology for the Ibarlucea Energy and Consumption Cooperative in Santa Fe, Argentina, using real active-power measurements recorded every 15 minutes from July 2024 to April 2026 at the point of interconnection with the upstream grid, meteorological data from Rosario Airport station, and temporal descriptors. The multi-horizon forecast combines a CNN-LSTM sequence-to-sequence architecture with a Histogram Gradient Boosting model that estimates the neural predictor residuals. Hourly and 15-minute resolutions, one- to three-day horizons, and two-, four-, and seven-day historical windows were evaluated. For a two-day window and a one-day horizon, the 15-minute configuration reduced MAE from 64.46 to 51.83 kW and RMSE from 102.57 to 94.33 kW; peak MAE decreased from 102.29 to 83.79 kW. Overall, residual correction improved both global and high-demand accuracy, although an underestimation bias remained. Finer temporal granularity was advantageous, whereas extending the historical window did not produce systematic improvements.