AI-driven Physics Informed Neural Network for Daily Temperature Forecasting with Constraint Aware Learning and Explainable Feature Attribution

Abstract

The daily mean temperature prediction is essential to implement agricultural adaptation and disaster risk reduction strategies in countries with varied climatic regions like Pakistan. Traditional machine learning methods often had difficulty complying with thermodynamic constraints, limiting their practicality for temperature estimation. The study proposed a Physics-Informed Neural Network (PINN), which integrates data and governing thermal principles to overcome these limitations. Integrating thermal equilibrium conditions within the loss function inherently consolidates the thermodynamic coherence of the construction. The high-resolution meteorological data from the Badin, Dadu and Rohri observation networks are used to build domain-specific features. These include diurnal excursions of temperature and humidity, a cyclic year encoding and wind-humidity ratios that capture the nonlinear mesoscale thermodynamics of the system. The PINN shows strong predictive ability as compared to a benchmark linear regression, some en-semble algorithms, and feedforward networks (R2 = 0.9975 Badin, 0.9974 Dadu, 0.9949 Rohri). SHAP and LIME, used in feature importance quantification, help to identify temperature drivers. In Badin, wind regimes have the most influence, whereas in Dadu and Rohri, lingering time trends have the most impact. With a focus on physical plausibility and explainable AI, the proposed methodology combines the probabilistic advantages of statistical learning with the constraint-based approach of atmospheric physics. This leads to resilient and spatially flexible predictions of temperature in data-scarce regions. As the study shows, PINNs could become a game-changing operational meteorological forecast technology when observational networks are weak or lacking altogether.