Spatiotemporal prediction of offshore wind fields based on a hybrid deep learning model for maritime navigation

Abstract

Accurate wind speed prediction is crucial for offshore wind power generation, ship navigation, and the use of renewable energy, as it optimises energy production, enhance maritime safety. This study introduces GswinLSTM, a novel hyrid model that integrates Long Short-Term Memory (LSTM) neural networks with the Group Swin Transformer (Gswin Transformer) to address the limitations of existing prediction models and improve wind speed prediction accuracy. Compared to conventional approaches, GswinLSTM simultaneously captures temporal dependencies and spatial correlations in wind speed data, significantly improving forecasting accuracy and robustness. The model is validated using ERA5 reanalysis data, which accurately represents offshore climate conditions. Experimental results demonstrate that GswinLSTM outperforms state-of-the-art models, including Transformer, Residual U-Net (ResUnet), and Convolutional LSTM (ConvLSTM), across four evaluation metrics, particularly in long-term forecasting where conventional methods struggle with error accumulation. By effectively capturing spatiotemporal dependencies, GswinLSTM enhances both prediction stability and precision in extended forecasting horizons. With strong theoretical contributions and practical applicability, this model offers valuable insights for wind field operations, maritime navigation, and climate monitoring. The findings underscore GswinLSTM's potential to drive advancements in renewable energy forecasting and environmental risk assessment, making it a promising tool for future atmospheric and meteorological studies. Additionally, the model's strong predictive capability supports maritime navigation safety, offshore wind energy optimization, and provides actionable insights for coastal management policies, such as maritime spatial planning and carbon reduction strategies.