Extreme heat prediction through deep learning and explainable AI.

Area of Science:

• Meteorology and Climate Science
• Artificial Intelligence and Machine Learning
• Environmental Science and Public Health

Background:

• Extreme heat waves pose significant ecological and societal risks due to rising global temperatures.
• Accurate heatwave forecasting is critical for proactive planning and public safety.
• Existing weather forecasting models have limitations in predicting extreme heat events.

Purpose of the Study:

• To investigate the efficacy of deep learning (DL) models for predicting extreme heat waves.
• To integrate Explainable AI (XAI) techniques for enhanced model interpretability.
• To address the gap in advanced computational modeling for extreme heat prediction.

Main Methods:

• Utilized five years of meteorological data from the Pakistan Meteorological Department (PMD).
• Developed and compared deep learning models: Artificial Neural Network (ANN), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM).
• Integrated Explainable AI (XAI) methods, specifically SHAP and LIME, for model interpretability.

Main Results:

• The Long Short-Term Memory (LSTM) model demonstrated superior performance with 96.2% accuracy for 1-3 day heatwave prediction.
• Explainable AI (XAI) methods identified humidity and maximum temperature as the most significant variables for predicting extreme heat.
• The study established a comprehensive framework integrating DL and XAI for improved heatwave forecasting.

Conclusions:

• Deep learning models, particularly LSTM, offer high accuracy in predicting extreme heat events.
• Explainable AI (XAI) is crucial for understanding model predictions and identifying key contributing factors.
• This research provides a foundation for enhanced heatwave forecasting and risk-reduction strategies.