A multi-metric evaluation and dynamic correction framework for local wind field prediction during super typhoons

Super typhoons can pose severe threats to coastal cities. For instance, Typhoon Yagi caused hundreds of fatalities and extensive property damages while sweeping across southern China and Southeast Asia in 2024. Accurately predicting the dynamic motion and intensity of super typhoons in high resolution is critical for effective disaster prevention and mitigation. Over the past few decades, the accuracy of typhoon track predictions has notably improved due to the advancements in global numerical weather prediction models. However, their capability to assess typhoon intensity and dynamic structure is still limited by coarse spatiotemporal resolution. The application of physics-based regional models, such as the Weather Research and Forecasting (WRF) model, presents a promising solution to this challenge.

To simulate high-resolution wind fields during super typhoons through WRF, it is essential to determine optimal and robust physical parameterization schemes. In previous studies, sensitivity analysis is often carried out solely based on the error criteria related to typhoon track and intensity, which are inadequate for the performance evaluation of local wind simulation. Additionally, there is a lack of consistent physical parameterization settings for different super typhoons. Furthermore, due to the inherent biases and model errors, a dynamic bias correction strategy is required for local wind forecasting. To this end, we aim to develop an integrated framework in this study that combines typhoon simulation, multi-metric evaluation, and dynamic bias correction.

The super typhoons that have significantly impacted Hong Kong over the past two decades have been chosen as study cases, i.e. Hato, Mangkhut, and Saola. A series of numerical experiments were designed to assess the impact of various physical models. By comparing simulation results with best track data and field observations from the Hong Kong Observatory and the Shenzhen Meteorological Gradient Mast, the multi-metric evaluation method provides a comprehensive understanding of both global and local wind field simulation performance. The best-performing physical models were thereby identified, achieving consistent typhoon tracks (MAE < 30 km), relatively accurate typhoon intensity predictions (RMSE < 5 m/s), and highly correlated wind fields (r > 0.9) between simulation and observation results. To further reduce the effects of systematic biases, a dynamic linear bias correction strategy was introduced to adjust local wind predictions dynamically based on real-time observations. Given the time-evolving local wind data, the linear bias correction factor can reach convergence and provide reliable forecast corrections with a lead time of approximately 15 hours. The proposed framework shows great potential to enhance disaster warning systems and improve local wind prediction accuracy in typhoon-prone regions.