Explainable deep reinforcement learning for climate forecasting with transfer learning.

Area of Science:

• Climate Science
• Artificial Intelligence
• Machine Learning

Background:

• Climate researchers use smart models to understand past climate records and predict future climate evolution.
• Climate prediction models simulate Earth's systems using high-performance computing and mathematical equations based on physical laws.
• Existing models often lack accuracy, flexibility, and explainability in climate forecasting.

Purpose of the Study:

• To propose a novel framework, the Explainable Reinforced Climate Modelling Framework (ERCMF), for accurate, flexible, and explainable climate predictions.
• To integrate explainable AI (XAI) and deep reinforcement learning (DRL) into a unified climate modeling approach.
• To enhance climate forecasting by providing valuable insights to scientists and policymakers.

Main Methods:

• Employs transfer learning (TL) for feature extraction from satellite and geospatial data via multi-variable climate representation learning (MV-CRL).
• Utilizes a deep reinforcement learning-based climatic policy network (CPN) for learning optimal forecasting strategies.
• Incorporates explainable AI (XAI) using Shapley Additive Explanations (SHAP) and attention mechanisms for prediction interpretability.

Main Results:

• Achieved a root mean square error (RMSE) of 1.84, outperforming existing methods (RMSE > 2.25).
• Demonstrated high performance with a temporal consistency score (TCS) of 0.91 and a policy convergence score (PCS) of 0.91.
• Excluding key modules significantly increased RMSE (up to 2.85) and reduced consistency, highlighting the framework's integrated approach.

Conclusions:

• ERCMF offers reliable, transparent, and adaptive climate prediction suitable for short-term and long-term analysis.
• The framework uniquely integrates TL, DRL, and XAI for real-time, interpretable climate forecasting.
• ERCMF provides valuable insights for climate scientists and policymakers, advancing climate modeling capabilities.