Reinforcement learning based typhoon-wave model ensemble for multi-physical parameterization schemes

Engineering notes

Key topics: autonomous driving, reinforcement learning, control. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Introduction This study proposes a dynamic weight optimization strategy based on the Soft Actor-Critic (SAC) algorithm for multi-source data-driven typhoon wind field fusion. The proposed method can adaptively adjust weights over time and achieve dynamic weight allocation. This not only better captures the complexity of the marine environment but also improves the simulation of typhoon wind speed evolution and enhances the reliability and accuracy of the fused results. Methods We adopt the SAC algorithm to realize time-varying weight allocation for multi-source typhoon wind data. WRF parameterization schemes, the typhoon model, and the sliding-window optimal weight averaging (SW-OWA) are set as baseline models for comparison. The fused wind field is further used to drive the SWAN model for wave simulation under typhoon conditions. Results The error results show that the SAC model performs better in wind-related metrics, with an RMSE of 1.745 m/s for 10-m wind speed and 5.321 m/s for maximum wind speed. The correlation coefficients of the evaluated variables range from 0.905 to 0.921, indicating that the SAC-based fusion method can reasonably capture the temporal variation characteristics of the typhoon wind and pressure fields. However, the RMSE of sea level pressure produced by the SAC model is higher than those of the WRF parameterization schemes, the typhoon model, and the sliding-window optimal weight averaging (SW-OWA) baseline model, suggesting that the current fusion framework still has limitations in simulating pressure-related variables. The simulated significant wave height (HS; CC = 0.947, RMSE = 0.434 m) and mean wave period (TM; CC = 0.858, RMSE = 0.438 s) both show good simulation performance. The results indicate that the proposed method can accurately reproduce the variation characteristics of typhoon-induced significant wave height. Although the simulated TM exhibits certain deviations in long-period wave regions, its overall variation trend remains reasonable. Discussion Overall, the SAC-based dynamic weight optimization strategy proposed in this study improves the stability and accuracy of typhoon wind field fusion, especially showing advantages in wind speed trend variation and peak error control, and verifies the application potential of reinforcement learning-based dynamic weight optimization in typhoon wind field fusion and wave simulation.

Links and sources

• PDF from original source

Comments