Stacked Hybrid Ensemble Learning for Enhanced Short-Term Load Forecasting in Developing Power Grids

Accurate short-term load forecasting (STLF) is critical for reliable and cost-effective power system operation, particularly in developing countries with fragile grids and volatile demand. This paper presents a hybrid ensemble forecasting framework that integrates linear regression, feedforward neural networks (NN), and a stacked meta-model learner for enhancing the accuracy of electricity demand predictions. Using hourly load, amperage, and weather data collected from load data from Geometric integrated limited, Aba, Nigeria, over a seven-month period (September 2024–March 2025), the hybrid approach is benchmarked against traditional methods including linear regression, neural networks, and Elman Recurrent Neural Networks (ERNN). The hybrid model consistently achieved superior performance across multiple statistical measures, recording the lowest RMSE (0.2608), MAE (0.1994), and MAPE (14.80%), while explaining 70.8% of variance in actual load (R² = 0.7079). Visual analysis of historical (Weeks 38–39) and future forecasts (Week 14) further confirmed its capacity to capture peak/off-peak variations more effectively than the baseline models. The results demonstrate that ensemble learning through stacked generalization offers robust advantages in contexts characterized by nonlinear load dynamics, climatic variability, and data irregularities. The findings provide a practical forecasting tool for utilities in Nigeria and other emerging economies, where accurate STLF is crucial for demand-side management, operational planning, and economic dispatch. Keywords — Short-term load forecasting, hybrid ensemble learning, neural networks, linear regression, Nigeria power systems.