Interpretable Physics-Informed Neural Network for Reliable Humidity Forecasting in Contrasting Climates of Sindh.

Abstract

Reliable forecasting of humidity is fundamental to effective climate monitoring, agricultural scheduling, and public health planning. Traditional machine learning frameworks, however, struggle to mirror the largely nonlinear climatological processes governing moisture distribution and rarely satisfy governing physical laws. This work introduces a physics-informed neural network (PINN) explicitly tailored to forecast daily mean humidity, embedding thermodynamic principles and smaller-scale climatological knowledge to bolster generalization and interpretability. The analysis is based on a comprehensive dataset of meteorological measurements from three climatically contrasting districts in Sindh, encompassing temperature, wind velocity, cumulative rainfall, and derived interaction terms. The PINN topology augments the mean squared error loss with a physically grounded rainfall–humidity relationship, thereby constraining the output to physically realizable humidity profiles. To decode the model’s inner operation, we employed SHAP and LIME, providing affected stakeholders with a transparent rationale for predictions. Comparative benchmarks reveal that the physics-informed architecture exceeds the forecast skill of conventional regression, ensemble methods, and standard feed-forward neural networks, securing R² metrics greater than 0.99 and uniformly low mean squared error across all three locales. Sensitivity assessments based on SHAP and LIME indicate that seasonal modulation and the interaction between wind and humidity are the predominant drivers of the forecasts, whereas the contribution of rainfall-derived predictors remains negligible. The combination of PINN with contemporary interpretability methodologies yields a forecasting paradigm for meteorology that is both transparent and governed by established physical laws.

Topics & Concepts