Feature-Driven Electrical Load Prediction Using Interpretable Machine Learning Models and Meta-Ensemble Learning: Evidence from Mosul’s 132 kV Grid Energy

Accurate forecasting of electrical load plays a critical role in enhancing the efficiency, reliability, and operational planning of power distribution systems—particularly in regions facing infrastructure limitations and volatile consumption patterns. This study proposes a robust forecasting framework based on real daily load data obtained from the 132 kV Al-Intisar substation in Mosul, Iraq, covering the years 2022 to 2024. The dataset, sourced in collaboration with the Nineveh Electricity Distribution Directorate, provides a practical foundation for modeling realistic demand behavior in emerging grid environments.A domain-informed and carefully engineered set of 21 predictive features was designed by the authors, integrating temporal dynamics, statistical indicators, and thermo-sensitive variables such as load-to-temperature ratios, multi-scale lag features, and moving averages. Nine machine learning models were selected to represent diverse algorithmic paradigms—ranging from tree-based methods and classical regressors to ensemble learners. Among them, the Gradient Boosting model achieved the highest accuracy Coefficient of Determination= 0.986, Mean Absolute Error= 0.260 MW, followed closely by the Meta-Learner (Coefficient of Determination = 0.983), strategically introduced in this study as a novel stacked ensemble, which demonstrated strong generalization by integrating multiple high-performing base learners. Explainability was addressed through Shapley Additive Explanations analysis and feature importance visualization, confirming the predictive value of the engineered features and enabling interpretable insights into model behavior. The findings affirm that hybrid and tree-based models, when coupled with purposeful feature design and interpretability tools, offer scalable and high-performing solutions for long-term load forecasting in real-world power systems.