An ensemble of Deep Learning, Machine Learning, and statistical methods stacked with meta-learning for forecasting net energy consumption in Multi-Carrier Energy Systems: Economic impact assessment

Accurate forecasting of net energy consumption in Multi-Carrier Energy Systems (MCES) is essential for effective resource management and economic planning. Accordingly, Machine Learning (ML) and Deep Learning (DL) methods are applied for energy forecasting, but often rely on individual methods, especially with complex datasets. This frequently results in reduced generalizability and robustness, making accurate forecasting with complex datasets more difficult. This paper presents an ensemble model of ML, DL, and statistical methods that are stacked with a meta-learner to forecast net energy consumption in an MCES. In the proposed model, advanced preprocessing techniques, such as cyclical feature encoding and feature ranking using eXtreme Gradient Boosting, are employed to enhance the model’s robustness against complex features. The results show that the proposed model achieves the lowest error metrics with a Mean Absolute Percentage Error of 9.8% and 17.6%, compared to individual methods at the hourly and daily horizons, respectively. By reducing the energy forecasting error, the model lowers operating costs, with an economic evaluation demonstrating a reduction of approximately 8% to 23% in energy costs during peak hours. Overall, the proposed model outperforms individual methods and provides more accurate and robust net energy forecasting for MCES.