Scholarlink

<p>To understand previous climate records and predict how future climates may evolve under different natural and anthropogenic forces, climate researchers are increasingly employing smart models. By implementing complex connections between the atmosphere, oceans, land surface, and biosphere, climate prediction models simulate Earth’s systems to project future trends. High-performance computing techniques are used to perform simulations, which are governed by mathematical equations based on physical laws. This work proposes the Explainable Reinforced Climate Modelling Framework (ERCMF), a novel framework for climate prediction that incorporates explainable AI (XAI) and deep reinforcement learning (DRL) to produce accurate, flexible, and explainable climate predictions. The suggested ERCMF employs transfer learning (TL) methods to derive useful features from satellite-based climatic records and geospatial data through multi-variable climate representation learning (MV-CRL). A deep reinforcement learning-based climatic policy network (CPN) processes these characteristics through episodic interactions with the climatic environment and progressively learns optimal forecasting methods. In addition, an explainable AI component employs Shapley Additive Explanations (SHAP) and attention mechanisms to expose the rationale behind every prediction, providing climate scientists and policymakers with valuable insights. ERCMF offers reliable and transparent climate prediction that is suitable for both short-term anomaly detection and long-term trend analysis by incorporating a real-time feedback loop that constantly adapts to new data. Experimental evaluation indicates that the proposed framework achieves a root mean square error (RMSE) of 1.84, with a temporal consistency score (TCS) of 0.91 and a policy convergence score (PCS) of 0.91, outperforming existing climate prediction approaches whose RMSE values exceed 2.25. Further analysis shows that excluding key modules results in an increase in RMSE of up to 2.85 and reduced consistency measures, highlighting the importance of comprehensive feature learning, sequential decision strategies, and interpretability of results for reliable climate forecasting. ERCMF uniquely integrates transfer learning, deep reinforcement learning, and post-hoc explainability into a unified framework, enabling real-time, interpretable, and adaptive climate forecasting—unlike existing models that typically lack one or more of these capabilities.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable deep reinforcement learning for climate forecasting with transfer learning

  • 期刊论文

  • 发表日期：

    2025年10月23日

查看全文
  • Thulasi Bikku,
  • Ramadevi Chappala,
  • Angotu Nageswara Rao,
  • Venu Gopal Gaddam,
  • Selva Joseph,
  • Srinivasarao Thota

摘要

To understand previous climate records and predict how future climates may evolve under different natural and anthropogenic forces, climate researchers are increasingly employing smart models. By implementing complex connections between the atmosphere, oceans, land surface, and biosphere, climate prediction models simulate Earth’s systems to project future trends. High-performance computing techniques are used to perform simulations, which are governed by mathematical equations based on physical laws. This work proposes the Explainable Reinforced Climate Modelling Framework (ERCMF), a novel framework for climate prediction that incorporates explainable AI (XAI) and deep reinforcement learning (DRL) to produce accurate, flexible, and explainable climate predictions. The suggested ERCMF employs transfer learning (TL) methods to derive useful features from satellite-based climatic records and geospatial data through multi-variable climate representation learning (MV-CRL). A deep reinforcement learning-based climatic policy network (CPN) processes these characteristics through episodic interactions with the climatic environment and progressively learns optimal forecasting methods. In addition, an explainable AI component employs Shapley Additive Explanations (SHAP) and attention mechanisms to expose the rationale behind every prediction, providing climate scientists and policymakers with valuable insights. ERCMF offers reliable and transparent climate prediction that is suitable for both short-term anomaly detection and long-term trend analysis by incorporating a real-time feedback loop that constantly adapts to new data. Experimental evaluation indicates that the proposed framework achieves a root mean square error (RMSE) of 1.84, with a temporal consistency score (TCS) of 0.91 and a policy convergence score (PCS) of 0.91, outperforming existing climate prediction approaches whose RMSE values exceed 2.25. Further analysis shows that excluding key modules results in an increase in RMSE of up to 2.85 and reduced consistency measures, highlighting the importance of comprehensive feature learning, sequential decision strategies, and interpretability of results for reliable climate forecasting. ERCMF uniquely integrates transfer learning, deep reinforcement learning, and post-hoc explainability into a unified framework, enabling real-time, interpretable, and adaptive climate forecasting—unlike existing models that typically lack one or more of these capabilities.