Novel Predictive Factors to Boost Short-Term Electricity Demand Forecasting Using Machine Learning Models

Abstract

Accurate Short-Term Load Forecasting (STLF) is essential for effective operational planning, particularly for optimizing maintenance schedules, managing power generation capacity, and ensuring grid stability. This paper presents a robust data-driven framework for STLF, focusing specifically on forecasting daily morning and evening peak electricity demand in Jordan. Multiple Machine Learning (ML) models are investigated, including Random Forest (RF), XGBoost, CatBoost, Support Vector Regression (SVR), Gradient Boosting Machines (GBM), Gaussian Process Regression (GPR), and a hybrid STL decomposition approach. Each model is accurately evaluated using Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) across training and testing data sets. A key contribution of this research is the integration of diversity of features that significantly enhance prediction accuracy. In addition to temperature, the feature set includes load growth rates and comparison with last year loads, daily usage data, population growth trends, and yearly day order. Using a five-year dataset (2015–2019), GPR emerges as the most accurate model, achieving a testing MAE of 4.71 MW and MAPE of 0.20% for morning peaks, and an MAE of 4.31 MW with a MAPE of 0.15% for evening peaks. To enhance model transparency, SHAP (SHapley Additive exPlanations) analysis is applied to tree-based models (RF and CatBoost), while permutation importance is used for kernel-based models such as GPR and SVR. This enables interpretable insights into feature contributions. In summary, the proposed methodology offers a robust and reliable tool for STLF. By accounting for both daily fluctuations and trends.