Physical Knowledge on Historical Data Matters More Than Enforcing Physical Constraints on the Forecast

This arXiv paper addresses the challenge of forecasting time series in applications involving physical processes. While recent deep learning models have advanced time series forecasting, they struggle with physical process prediction where intermediate physical variables are unobservable. The authors propose Physics Informed Recurrent Neural Network (PIRNN), which incorporates physical knowledge into training data to help domain experts better understand target behavior, rather than merely enforcing physical constraints during forecasting.