Data-driven wave energy flux prediction with local decomposition

Abstract

As a clean and promising renewable energy source, ocean wave energy has attracted increasing attention, and its reliable prediction is essential for renewable energy assessment and related applications. In this study, a novel data-driven forecasting framework, termed ED-TimesNet, is developed for wave energy flux prediction by integrating a local decomposition strategy with an enhanced deep learning architecture to improve forecasting performance. Specifically, Local Empirical Wavelet Transform (Local-EWT) is employed to mitigate noise effects and avoid future information leakage, while TimesNet effectively captures the multi-scale periodic characteristics of wave-related parameters by learning both intra-period and inter-period dependencies. In addition, a DenseBlock module is incorporated to further strengthen feature representation. Model performance is evaluated for wave energy flux prediction with lead times of 3, 6, and 12 h, and the effects of training set size and regional strategy are analyzed. Results show that ED-TimesNet consistently outperforms comparison models across different datasets and lead times, and is least affected by increasing prediction horizons. Overall, ED-TimesNet is a promising tool for ocean wave energy prediction and offers valuable support for wave energy applications.