Sabitlenmiş Tweet

We conducted benchmarking for different sea surface temperature (SST) prediction models for a square domain in the Arabian Sea, covering prediction periods of 1 week, 15 days, and 30 days. For short-term predictions (1 week), all models (CNN, LSTM, Transformer, PINN-Transformer) showed good agreement with actual sea surface temperatures, with minor deviations and overall high accuracy. The CNN model initially performed well for shorter lead times, demonstrating high accuracy and good alignment with actual observations. As the prediction period increased to 15 days, the models began to show noticeable deviations from actual observations. However, the CNN and PINN-Transformer maintained relatively better consistency compared to the others. By the 30-day mark, significant deviations from actual observations were observed in all models, but the PINN-Transformer demonstrated better consistency compared to the other models, despite some variability.
When evaluating model performance for 30-day ahead predictions using metrics such as Anomaly Correlation Coefficient (ACC), Nash-Sutcliffe Efficiency (NSE), Normalized Root Mean Square Error (NRMSE), and Mean Absolute Error (MAE), both the PINN-Transformer and Transformer models consistently outperformed the other models. The PINN-Transformer and Transformer models exhibited higher ACC and NSE values, indicating better agreement with observed data and predictive performance. Additionally, they showed lower NRMSE and MAE values, reflecting higher accuracy and lower prediction errors. The superior performance of the PINN-Transformer is likely due to the preservation of physical laws within its framework, while the Transformer model benefits from its advanced architecture, which helps maintain accuracy and stability in predictions over longer periods. In contrast, the CNN model tends to lose information over longer lead times, resulting in reduced performance. This highlights the importance of model choice, with physics-informed and Transformer models being more effective for long-term SST predictions.
@ClimateIITB
GIF
English






















