Scholarlink

<p>Accurate local rainfall prediction is vital for climate-vulnerable regions such as Sindh, Pakistan, where agriculture, water management, and flood preparedness depend on reliable forecasts under highly variable hydroclimatic regimes. This study proposes a compact physics-informed neural network that embeds an explicit relative-humidity constraint into the loss function to suppress unphysical precipitation under dry conditions, operationalized by penalizing predicted rainfall when humidity falls below 60%. The architecture is a lightweight feedforward network with 5,121 parameters, trained on multi-decadal daily observations from three contrasting climatic zones of Badin (coastal), Dadu (semi-arid), and Rohri (arid) by using engineered features that capture antecedent rainfall, humidity–temperature–wind interactions, and seasonal signals under strict, time-aware data splitting to prevent temporal leakage. A composite loss, couples data fidelity with physics regularization, and the model is evaluated by a station-wise train–validation–test blocks, expanding-window cross-validation, and leave-one-station-out experiments against optimized linear models, tree ensembles, recurrent networks, and a physics-guided deep baseline without tailored constraint design. The PINN attains consistently high skill across regimes, with representative station-level MSE of 0.79–1.19–0.49 and RMSE of 0.97–0.97–0.97 for Badin, Dadu, and Rohri, respectively, while effectively eliminating negative and moisture-inconsistent rainfall predictions. SHAP and LIME analyses identify rainfall ratio, cumulative antecedent rainfall, temporal encodings, and humidity–temperature interactions as dominant drivers, confirming that the humidity constraint sharpens physically meaningful attributions rather than spurious correlations. By unifying explicit atmospheric priors with dual explainability in a modular, low-cost architecture, the proposed framework delivers trustworthy, operationally relevant rainfall guidance for data-scarce environments and offers a transferable blueprint for physics-informed, interpretable AI in hydrometeorology.</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

A physics informed deep learning framework for rainfall forecasting in diverse climatic regions

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月10日

查看全文
  • Syed Azeem Inam,
  • Saddam Umer,
  • Haider Rajput

摘要

Accurate local rainfall prediction is vital for climate-vulnerable regions such as Sindh, Pakistan, where agriculture, water management, and flood preparedness depend on reliable forecasts under highly variable hydroclimatic regimes. This study proposes a compact physics-informed neural network that embeds an explicit relative-humidity constraint into the loss function to suppress unphysical precipitation under dry conditions, operationalized by penalizing predicted rainfall when humidity falls below 60%. The architecture is a lightweight feedforward network with 5,121 parameters, trained on multi-decadal daily observations from three contrasting climatic zones of Badin (coastal), Dadu (semi-arid), and Rohri (arid) by using engineered features that capture antecedent rainfall, humidity–temperature–wind interactions, and seasonal signals under strict, time-aware data splitting to prevent temporal leakage. A composite loss, couples data fidelity with physics regularization, and the model is evaluated by a station-wise train–validation–test blocks, expanding-window cross-validation, and leave-one-station-out experiments against optimized linear models, tree ensembles, recurrent networks, and a physics-guided deep baseline without tailored constraint design. The PINN attains consistently high skill across regimes, with representative station-level MSE of 0.79–1.19–0.49 and RMSE of 0.97–0.97–0.97 for Badin, Dadu, and Rohri, respectively, while effectively eliminating negative and moisture-inconsistent rainfall predictions. SHAP and LIME analyses identify rainfall ratio, cumulative antecedent rainfall, temporal encodings, and humidity–temperature interactions as dominant drivers, confirming that the humidity constraint sharpens physically meaningful attributions rather than spurious correlations. By unifying explicit atmospheric priors with dual explainability in a modular, low-cost architecture, the proposed framework delivers trustworthy, operationally relevant rainfall guidance for data-scarce environments and offers a transferable blueprint for physics-informed, interpretable AI in hydrometeorology.