Attention ResNet-GRU model for Typhoon Prediction

Multimodal data analysis refers to the integration of different types of data (such as images and numerical features) to leverage the complementarity and interactive information between various data sources, aiming to enhance the accuracy and effectiveness of analytical tasks. The advantage of multi- modal methods lies in their ability to comprehensively utilize diverse data sources, thereby providing more comprehensive and accurate analysis results. Consequently, they have significant advantages in various prediction tasks. In this study, we investi- gate multimodal data analysis methods using a typhoon dataset obtained from the Chinese Academy of Sciences’ Ocean Science Data Center. Through preprocessing, intensity mapping, and data augmentation techniques, we construct a dataset containing features such as longitude, typhoon category, wind speed, and typhoon intensity. To improve prediction accuracy, we utilize the data from the past five days to predict the current typhoon longitude and combine every five samples into a complete ty- phoon record. Additionally, we transform numerical features into images and normalize them to enhance data representation. In terms of model design, we employ a multimodal neural network, where ResNet and a GRU network with multi-head attention mechanism handle image and data features, respectively. The typhoon longitude prediction is achieved through feature fusion and fully connected layers. Experimental results demonstrate that the multimodal approach outperforms single-modal methods in regression tasks. This research holds significance for typhoon path prediction and disaster prevention.