Optimal weights and Heuristic Meta-Models: Learning-based multi-model strategies for energy forecasting

This paper introduces three novel forecasting frameworks designed to enhance prediction accuracy through an intelligent combination of machine learning models. The first two frameworks, Heuristic Meta-Model Frameworks (HMMF 1 & HMMF 2), employ distinct selection criteria: HMMF 1 prioritizes models based on R 2 scores, maximizing explained variance, while HMMF 2 selects models using RMSE, focusing on minimizing overall prediction error, particularly penalizing large deviations. The third framework, Optimally Weighted Bias-Variance Forecasting (OWBVF), optimally combines multiple forecasts by leveraging bias estimation and the variance–covariance matrix of forecast errors, ensuring a balanced trade-off between bias reduction and variance minimization. Each framework integrates predictions from four individual machine learning models—XGBoost, K-Nearest Neighbors, Random Forest, and Decision Tree—along with a combined model merging XGBoost and Random Forest predictions. The models are dynamically weighted based on their predictive performance, ensuring an adaptive and robust forecasting strategy. Extensive evaluations were conducted on multiple datasets, including PV production, building energy consumption, and regional power consumption. Results consistently demonstrate that the proposed frameworks outperform both individual and traditional ensemble models, highlighting their effectiveness as advanced forecasting solutions for energy systems and beyond. • Introduce three novel frameworks for enhanced forecasting accuracy. • HMMF 1 and HMMF 2 select models using R 2 and RMSE criteria. • OWBVF balances bias reduction and variance minimization. • Combine XGBoost, KNN, Random Forest, and Decision Tree adaptively. • Outperform individual and traditional ensemble models in energy tasks.