Spatio-Temporal Wind Speed Prediction Model Based on Multi-Scale Fusion MambaFormer

In recent years, spatiotemporal wind speed prediction using deep learning methods has become a hot research topic. However, these predictions still face challenges as existing models primarily focus on spatial features and exhibit weaker capabilities in capturing temporal characteristics, failing to meet the efficiency and accuracy requirements of practical applications. To address these challenges, this paper introduces a novel method, the Multiscale Fusion MambaFormer spatiotemporal wind speed prediction model (MF-MambaFormer). This model integrates the Transformer with Mamba, leveraging the Transformer's ability to model long-distance dependencies and Mamba's capability to extract temporal features. By combining the strengths of these technologies, the model significantly enhances prediction accuracy. Additionally, through the Spatial Channel Gating Aggregation Module (SCGAM), which utilizes dilated convolutions to extract spatial features across different receptive fields and a channel attention module to capture channel weights, the model improves its ability to apprehend spatial and channel features. Moreover, by incorporating the Mam-baFormer module into the multiscale fusion, we propose a novel Multiscale Spatiotemporal Fusion Module (MSTFM) that integrates and effectively combines spatiotemporal features at various scales. We comprehensively evaluated our model's performance on the WeatherBench dataset, and extensive experimental results demonstrate significant improvements in spatiotemporal wind speed prediction.