International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Probabilistic Short-Term Load Forecasting Using GBDT-Based Sister Forecasts and Ensemble Methods

Authors :

Hossein Shahinzadeh¹ Hamed Nafisi² Amirafshin Zamani³ Saiedeh Mehrabani-Najafabadi⁴ Arezou Mahmoudi⁵ Farshad Ebrahimi⁶

1- Amirkabir University of Technology (Tehran Polytechnic) 2- Amirkabir University of Technology (Tehran Polytechnic) 3- Islamic Azad University Najafabad Branch 4- Islamic Azad University Najafabad Branch 5- University of Tabriz 6- University of Houston

Keywords :

Probabilistic load forecasting،Gradient Boosting Decision Trees،Sister Forecasts،Learning-based ensemble،Short-term load forecasting

Abstract :

Short-term load forecasting (STLF) plays a vital role in the operational scheduling and expansion planning of modern power systems. The inherent uncertainties in electricity demand necessitate reliable probabilistic forecasting approaches. In this paper, an enhanced hybrid framework is proposed for probabilistic load forecasting by leveraging state-of-the-art Gradient Boosting Decision Tree (GBDT) models, including LightGBM, XGBoost, and CatBoost, as base learners for generating sister forecasts. The sister forecasts are obtained by training GBDT models on different feature subsets and temporal lag structures derived from historical load and weather data. To aggregate these sister forecasts, both conventional statistical methods (Trimmed and Winsorized averaging) and a learning-based stacking ensemble approach are employed. The proposed method is evaluated using the publicly available GEFCom2014 dataset, covering multiple years of hourly load and meteorological data. Experimental results demonstrate that the GBDT-based sister forecasts consistently outperform traditional decision tree and quantile regression benchmarks in both point and probabilistic forecasting accuracy. Furthermore, the learning-based ensemble method yields superior performance over conventional averaging techniques, achieving improved reliability in predictive distributions and peak load estimation.