Time-series weather prediction in the Red sea using ensemble transformers

Abstract

Accurately predicting Sea Surface Temperature (SST) and Wind Speed (WS) in the Red Sea is crucial for maritime safety, climate research, and ecosystem monitoring. However, modeling SST and WS is challenging due to the region’s complex and dynamic oceanic and atmospheric processes. This study introduces advanced predictive models based on Transformers to enhance the accuracy of SST and WS predictions by capturing time dependencies and modeling sequential data. A novel stacked architecture, named StackPred, improves the performance of individual Transformer models. Additionally, wavelet-based multiscale filtering is applied during preprocessing to separate relevant signals from noise, further enhancing prediction accuracy. The prediction performance is validated using publicly available SST and WS data from ten different locations in the Red Sea. Results consistently demonstrate the superiority of the proposed stacked deep learning models over existing techniques, including LSTM (Long Short-Term Memory), BiLSTM (Bidirectional LSTM), GRU (Gated Recurrent Unit), BIGRU (Bidirectional GRU), and single Transformers. Moreover, integrating wavelet denoising with StackPred significantly improves predictive performance compared to using data without denoising. The StackPred model achieves strong performance with an average R2 score of 99.83 for predicting SST and 99.96 for WS.