Design of a Hybrid Power Demand Forecasting Model with Uncertainty Quantification under Input Perturbation

Abstract

The study aimed to address the persistent challenge of accurate electricity demand forecasting in modern power systems, where reliability is often undermined by input perturbations such as weather fluctuations, consumer behaviour shifts, and sensor noise. Conventional Machine Learning (ML) and Deep Learning (DL) approaches, while effective in predictive accuracy, rarely incorporate uncertainty estimation, which reduces their robustness in real-world applications. To overcome this limitation, this study designed a hybrid power demand forecasting model with embedded uncertainty estimation. Seven base models, such as Artificial Neural Networks (ANN), Recurrent Neural Networks (RNN), Support Vector Regression (SVR), Extreme Gradient Boost (XGBoost), RandomForest, LightGBM, and CatBoost, were trained and evaluated using R², RMSE, MAE, and MAPE. The three best-performing models (XGBoost, CatBoost, and Random Forest) were fused through weighted averaging based on inverse error contributions. An uncertainty estimation mechanism was then integrated by quantifying variance under perturbed inputs, thereby generating confidence intervals around forecasts. Findings show that the hybrid model achieved high predictive accuracy (R² = 0.9539, with low error values: RMSE = 1.7128, MAE = 1.2270, MAPE = 3.1178) while also producing reliable uncertainty bounds. The significance of this study lies in demonstrating that hybrid modelling combined with uncertainty quantification provides both accurate and trustworthy forecasts, offering a practical decision-support tool for smart grid operators managing volatile energy demand.