Journal of Electronic Systems Hajar Ait Lamkademe 16 3 2026 https://doi.org/10.6025/jes/2026/16/3/163-178 https://www.dline.info/jes/fulltext/v16n3/jesv16n3_3.pdf Accurate short-term load forecasting is essential for the efficient and reliable operation of modern smart grids. This study develops and comprehensively evaluates a high-resolution, interpretable load forecasting framework using 10-minute smart meter data from Laâyoune Zone 1, Morocco. To capture complex temporal dependencies, we engineered advanced features, including lag variables, rolling statistics, and cyclical time encodings. Multiple machine learning models were systematically compared, with the Random Forest (RF) ensemble achieving the best overall performance (RMSE: 5.42, R²: 0.87), effectively capturing nonlinear demand dynamics. Notably, the unexpectedly strong performance of the Persistence baseline highlighted the significant short-term autocorrelation inherent in high-frequency residential and industrial load data. Multi-step forecast horizon analysis revealed exceptional accuracy for 10 to 60 minute predictions, making the models highly suitable for real-time grid balancing and transformer management, although uncertainty gradually accumulates beyond two hours. Residual diagnostics further confirmed minimal systematic bias and stable model behavior across varying demand conditions. Furthermore, SHAP-based interpretability analysis confirmed that recent lagged load measurements and temporal indicators are the primary drivers of predictions, overcoming the traditional “black-box” limitations of machine learning. Ultimately, this explainable ensemble framework provides a robust, transparent, and computationally efficient tool for utility operators, facilitating demand response, anomaly detection, and sustainable energy management in modern smart grid environments.