Evaluation of Typhoon Forecasts Using the Pangu-Weather Neural Network Model: A Comparative Analysis with ECMWF and Other Numerical Prediction Models

In the realm of real-time weather prediction, notable institutions such as ECMWF and CMA Guangdong Meteorological Observatory have implemented the Pangu-Weather neural network model for real time forecasts. This integration has ushered in a valuable opportunity for the meteorological prediction community to evaluate the efficacy of AI-driven weather forecasts while also exploring avenues for enhancement. Throughout the year 2023, a sequence of typhoons swept across the southwest Pacific region, spanning a continuous two-month period. Leveraging the real-time predictions from the Pangu-Weather model alongside conventional numerical forecasting methods, we were able to conduct a comprehensive evaluation of their performance, focusing on both track predictions and intensity estimations. Remarkably, the track forecasts generated by the Pangu-Weather model demonstrated a striking alignment with those produced by the esteemed ECMWF model. This congruence was especially apparent when scrutinizing the intricate trajectories of powerful systems like Super Typhoon Khanun. However, it is worth noting that the strength forecasts derived from the Pangu-Weather model exhibited a consistent pattern of underestimation compared to other numerical predictions. This variance in intensity forecasts presents an evident avenue for refinement and advancement in the capabilities of the Pangu-Weather model.

Our comprehensive analysis encompassed a thorough examination of these forecasts, delving into their intricacies to unearth potential areas of improvement. By identifying the nuances responsible for the intensity underestimations, we aim to pave the way for future enhancements that will elevate the performance of AI-based weather predictions, bolstering their accuracy and reliability.