An optimized machine learning approach for short-term electricity demand forecasting considering environmental and historical influencing factors

Abstract

This study presents an optimized machine learning framework for short-term electricity demand forecasting to support energy purchase planning and operational reliability of power distribution systems—particularly under market restructuring and climate variability. The proposed method employed a Random Forest (RF) algorithm enhanced through Grid Search hyperparameter tuning. Temperature data, as a key environmental factor, along with previous-day electricity demand, were used as input features. The model was evaluated using real-world data from a distribution substation in an Iranian city and was compared with conventional forecasting approaches, including Exponential Smoothing (ES), Seasonal ARIMA (SARIMA), and the standard RF. The optimized RF model achieved a Mean Square Error (MSE) of 0.01 and an R² of 0.894 in winter, outperforming the other methods across all seasons. These results confirmed that systematic hyperparameter optimization can significantly enhance the predictive performance of machine learning-based models for day-ahead electricity load forecasting.