Feature fusion temporal convolution: Wind power forecasting with light hyperparameter optimization

Abstract

Power forecasting is a critical component for maintaining the stability and efficiency of electrical systems, particularly in the context of applied energy. The integration of advanced deep learning (DL) techniques into power forecasting has seen a significant surge in interest and application. This focus is particularly relevant in the challenging domain of wind power forecasting, where the need for accurate predictions is amplified by the escalating variability of data and the growing prominence of wind energy within national power grids. This paper introduces a novel approach, integrating the strengths of Feature Fusion Networks (FFN) and Temporal Convolution Network (TCN) architectures for power forecasting tasks. Our primary contribution lies in harnessing these deep learning techniques, coupled with meticulous feature selection, to enhance forecasting precision and data variability. This research employed two diverse datasets, each presenting its unique set of challenges, to validate the robustness of the proposed FFN-TCN. The integrated FFN-TCN model outperformed state-of-the-art benchmarks, achieving a Mean Absolute Error of 19.0655, MSE of 19.5, RMSE of 14.6433, a correlation coefficient of 0.9737 and a coefficient of determination ( R 2 ) of 0.997. Our model achieves an average Nash-Sutcliffe Coefficient (NSC) of 0.9997 across case studies showing strong wind power forecasting capabilities. These metrics not only showcase the efficacy of the FFN-TCN framework but also highlight the benefits of rigorous feature selection. In comparison with existing models, the proposed FFN-TCN model, enhanced by feature selection, showcases potential as a gold standard in power forecasting. Its ability to deliver superior results across two datasets underscores its adaptability, reliability, and promising findings. This research paves the way for future work in the domain of power forecasting applications, emphasizing the synergy between advanced neural network architectures and wind energy supply and load demand. • We introduce an integrated approach of Feature Fusion Networks (FFN) and Temporal Convolution Network (TCN) architectures for power forecasting tasks. • Multiple studies are carried out to validate the robustness of the FFN-TCN. • FFN-TCN outperformed state-of-the-art benchmarks, achieving a MAE of 19.0655, MSE of 19.5, RMSE of 14.6433, and a coefficient of determination R 2 of 0.997. • The benefits of rigorous feature selection and the efficacy of the FFN-TCN framework is highlighted.

Topics & Concepts