A Comparative Evaluation of Ensemble and Statistical Models for Electricity Consumption Prediction

Abstract

Accurate electricity consumption forecasting is essential for effective energy management, demand planning, and reliable power-system operation. This study investigates machine learning and statistical approaches for predicting electricity consumption using a dataset collected from four major cities in Pakistan. Five forecasting models are developed and evaluated, including two ensemble learning models, XGBoost and LightGBM, the statistical time-series model ARIMA, and two hybrid models combining XGBoost with ARIMA and LightGBM with ARIMA. Several other predictive models were also examined; however, their performance was comparatively inferior, and therefore they were excluded from the final analysis. The models are evaluated using four widely adopted metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), coefficient of determination $(\mathrm{R}^{2})$, and Mean Absolute Percentage Error (MAPE). The experimental results demonstrate that all five selected models achieve strong predictive performance for electricity consumption forecasting. The hybrid models effectively combine the linear time-series modeling capability of ARIMA with the nonlinear learning capabilities of XGBoost and LightGBM, enabling them to capture different patterns in electricity consumption data. The results demonstrate the potential of ensemble and hybrid approaches for accurate electricity consumption forecasting and provide useful insights for energy demand management in major urban areas of Pakistan.