Towards Enhancing Robust Energy Forecasting: A Hybrid Model with Input Perturbation to Overcome Uncertainty in Power Demand Prediction

Abstract

The purpose of this article is to address the persistent challenge of reliable power demand forecasting in modern energy systems, where dynamic fluctuations and noisy signals often reduce model accuracy and credibility. Traditional forecasting methods, although widely applied, struggle to adapt to stochastic variations, limiting their usefulness for grid stability and long-term planning. This study proposed a Hybrid Power Demand Forecasting Model with Uncertainty Estimation under Input Perturbations (HPDEF-MUIP). The model combined three machine learning algorithms, Extreme Gradient Boosting, Categorical Boosting, and RandomForest, into a hybridised model designed to enhance robustness. Data preprocessing included Empirical Mode Decomposition for signal refinement, Kalman Filtering for noise reduction, and normalisation for balanced scaling. To simulate and evaluate resilience against noisy environments, adversarial perturbation strategies such as the Fast Gradient Sign Method were introduced. The model was trained and validated on a large smart meter dataset spanning 2022–2025, using ten-fold cross-validation and hyperparameter optimisation with Genetic Algorithms. Performance was assessed through standard accuracy metrics, including Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), the coefficient of determination and Mean Absolute Percentage Error (MAPE). Findings showed that HPDEF-MUIP achieved an R² of 0.9539 and a MAPE of (3.12%), significantly outperforming baseline models. Under perturbed conditions, adversarial training reduced error variance by (17%), confirming improved resilience. The study concludes that hybrid model learning with uncertainty estimation offers a reliable and interpretable tool for supporting smart grid operations, demand-response planning, and sustainable energy management.