A HYBRID POWER DEMAND FORECASTING MODEL WITH UNCERTAINTY ESTIMATION UNDER INPUT PERTURBATIONS

Abstract

Accurate power demand forecasting is crucial for effective electricity planning, resource allocation, and maintaining power grid stability, particularly in the context of dynamic consumption patterns and increasing system complexity. Traditional forecasting models often fall short when exposed to uncertainties caused by input perturbations such as fluctuations in voltage, current, and peak load, resulting in suboptimal forecasts and unreliable planning decisions. To overcome these limitations, this study developed a Hybrid Power Demand Forecasting Model with Uncertainty Estimation under Input Perturbations (HPDEF-MUIP). The study objectives were to assess the weaknesses of existing models, to design, implement, and evaluate the HPDEF-MUIP. The model was designed to enhance robustness and forecasting accuracy by fusing three top-performing base models: XGBoost, CatBoost, and Random Forest, each offering complementary strengths in computational scalability, handling of categorical features, and precision in regression tasks. The research employed a quantitative methodology using half-hourly historical power demand data from a Kenya Power and Lighting Company substation, supplemented with meteorological data from the Nakuru Meteorological Station, and critical input features prone to input perturbation. After comprehensive data preprocessing, including cleaning, normalization, and feature engineering, the hybrid model was rigorously trained and validated on real-world data from regional power substations to ensure operational relevance. The findings show that the model achieved an R² of 0.9539, an RMSE of 1.7128, and an MAPE of 3.12%, along with a robust F1-score of 0.9112, demonstrating superior performance in both regression and classification metrics. A notable contribution of this study is the incorporation of a structured uncertainty estimation mechanism through input perturbation analysis, which provides critical insights into model reliability under non-ideal operating conditions—a crucial aspect often overlooked in conventional approaches. The HPDEF-MUIP model not only improved accuracy over individual base models but also enhanced interpretability and resilience, making it suitable for deployment in edge computing environments and adaptive smart grid infrastructures. The study concluded that the HPDEF-MUIP represented a significant advancement in forecasting methodologies by effectively bridging the gap between forecasting accuracy, uncertainty estimation, and operational feasibility in a real-world context. Based on these findings, the study recommends the adoption of hybrid uncertainty-aware models within power forecasting systems to enhance the planning and stability of the smart grid. Formulation of supportive policies is also recommended, and further research is suggested to expand the scalability and interpretability of Hybrid models across diverse Power systems. This research contributes a scalable and replicable framework that supports Sustainable Development Goals (SDGs) through more informed decision-making in load balancing, infrastructure planning, and sustainable energy management.