Typhoon track tracking and forecasting algorithm based on multi-source information

We propose a novel approach that combines regional ocean data prediction with typhoon trajectory inversion methods to improve the forecasting of typhoon trajectories. This model expands outward from the current path to create a grid within a fixed range of latitudes and longitudes, populated with future ocean environment information to effectively identify the next center point of the trajectory. The model's predictive performance is evaluated using 57 historical typhoons in the North Atlantic for 24- to 72-h lead times through comparative analysis. The results demonstrate a 6.3% reduction in error for 24-h lead-time forecasts and a significant 33.3% reduction for 48-hour lead-time forecasts compared to the popular gated recurrent unit convolutional neural network model, underscoring the enhanced accuracy of the new model, particularly for longer forecasting intervals. In addition, the gradient-weighted class activation mapping technique is employed to examine the relationship between specific inherent features of typhoons, such as wind speed, atmospheric pressure, turning points, and overall movement speed, and the model's prediction error, utilizing statistical relationships to identify conditions under which the model exhibits large or small errors so that we can evaluate its effectiveness in predicting mainstream typhoons while recognizing characteristics that complicate certain forecasts and necessitate increased caution. This provides new insights for artificial intelligence models exploring the physical mechanisms of typhoon path prediction.