Multi-output deep learning for high-frequency prediction of air and surface temperature in Kuwait

Area of Science:

• Environmental Science
• Climate Modeling
• Data Science

Background:

• Accurate air and surface temperature prediction is vital for urban planning and climate resilience, particularly in arid regions.
• High-frequency climate data is essential for understanding and predicting localized temperature variations.

Purpose of the Study:

• To evaluate multi-output regression models for predicting air and surface temperatures using high-frequency climate data.
• To benchmark deep learning and traditional machine learning approaches for 5-minute temperature forecasting in Kuwait.

Main Methods:

• Utilized 30 environmental variables collected every 5 minutes over four years in Kuwait.
• Benchmarked ten models, including FTTransformer and LSTM, using a leave-1-year-out validation strategy.
• Employed SHAP and permutation importance for model interpretation to identify key predictors.

Main Results:

• FTTransformer and LSTM achieved high predictive performance (R²=0.998, MSE=0.13, MAE=0.24) for six temperature variables at 5-min resolution.
• Deep learning models significantly outperformed traditional machine learning models.
• FTTransformer demonstrated stable accuracy across years, while LSTM's performance degraded with anomalous data.

Conclusions:

• Deep learning models, particularly FTTransformer, offer robust and accurate solutions for high-frequency climate prediction in arid regions.
• Key environmental variables significantly influence temperature prediction accuracy.
• FTTransformer shows superior stability compared to LSTM for long-term climate forecasting applications.