Physical knowledge on historical data matters more than enforcing physical constraints on the forecast

Physics Informed Recurrent Neural Networks (PIRNN) improve time series forecasting by predicting unobservable intermediate physical variables alongside target values. This approach prioritizes integrating physical knowledge from historical data over enforcing constraints during forecasting, enhancing model robustness and interpretability for domain experts in applications like groundwater level prediction.