Reinforcement learning-based multi-model ensemble for ocean waves forecasting

This study addresses the challenges of uncertainty in wave simulations within complex and dynamic ocean environments by proposing a reinforcement learning-based model ensemble algorithm. The algorithm combines the predictions of multiple base models to achieve more accurate simulations of ocean variables. Utilizing the soft actor-critic reinforcement learning framework, the method dynamically adjusts the weights of each base model, enabling the model ensemble algorithm to effectively adapt to varying ocean conditions. The algorithm was validated using two SWAN models results for China’s coastal regions, with ERA5 reanalysis data serving as a reference. Results show that the ensemble model significantly outperforms the base models in terms of root mean square error, mean absolute error, and bias. Notable improvements were observed across different significant wave height ranges and in scenarios with large discrepancies between base model errors. The model ensemble algorithm effectively reduces systematic biases, improving both the stability and accuracy of wave predictions. These findings confirm the robustness and applicability of the proposed method for integrating multi-source data and handling complex ocean conditions, highlighting its potential for broader applications in ocean forecasting.