A hybrid machine learning and optimization framework for energy forecasting and management

Accurate energy prediction and load optimization are crucial for improving grid efficiency and lowering operational costs in industrial and commercial energy systems. This study presents a hybrid framework that combines Fourier Transform (FT)-based transformers for high-resolution energy forecasting with an improved Covariance Matrix Adaptation Evolution Strategy (CMA-ES)-based genetic algorithm for optimal load scheduling. The novelty of this paper lies in the integration of FT-transformers with optimization algorithms to enhance forecasting accuracy and scheduling efficiency, offering a scalable solution for industrial-scale energy management. The FT-transformer model utilizes self-attention mechanisms and Fourier-based seasonality encoding to capture long-term dependencies, achieving a Mean Absolute Error (MAE) of 3.03 × 10 5 kWh and a Root Mean Square Error (RMSE) of 3.31 × 10 5 kWh, representing an improvement of 48% over traditional Recurrent Neural Networks (RNNs). The optimization component uses a multi-objective genetic algorithm CMA-ES to minimize peak energy demand fluctuations, reducing them by 27% while also minimizing cost deviations. Comparative analysis across various forecasting models, including RNNs, tree-based models, and CMA-ES, shows that the proposed method consistently outperforms existing techniques in both precision and computational efficiency. Scalability assessments indicate that, with their parallel processing capabilities, FT-transformers decrease the inference time by 38% compared to sequential models, making them suitable for real-time deployment in energy management systems. This study contributes to the field by integrating advanced machine learning with optimization for demand-side management, providing a scalable and efficient solution for industrial-scale energy forecasting. Future research will extend this framework with probabilistic forecasting and reinforcement learning for adaptive load control in dynamic energy environments. • Created a load forecasting model utilizing FT-transformers and genetic algorithms. • Achieved a 27% reduction in peak demand uncertainty during MATLAB simulations. • Enhanced predictive accuracy by 48% compared to Recurrent Neural Network utilizing self-attention. • Reduce the inference time by 38% to facilitate immediate energy management. • Validated the model utilizing MATLAB/Simulink and empirical mining data.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.