Authors: Tarun Agrawal and Pin Shuai – Utah State University, Yanan Duan and Mukesh Kumar- University of Alabama
Title: A physics informed machine learning framework for baseflow separation
Presentation Type: Poster Presentation
Abstract: Baseflow is a fundamental component of streamflow, yet its estimation remains inherently uncertain due to the absence of direct observations and the reliance on empirical separation methods. Traditional digital filters and graphical separation methods impose fixed functional forms and fitting parameters that are often either time-invariant or, in a few studies, piecewise calibrated based on seasonality or flow regimes, limiting their ability to represent the dynamic and nonlinear nature of groundwater contributions.
We present a physics-informed machine learning (PIML) framework that learns parameters governing baseflow-streamflow partitioning using a long short-term memory (LSTM) network driven by predicted antecedent baseflow and observed current discharge, rather than prescribing them a priori. Hydrological consistency is enforced by incorporating a recession-based physical constraint into the training of the model. Recession periods are identified using a well-established irregular binning method, and an additional loss term ensures that predicted baseflow follows physically consistent recession behavior.
The framework has been tested in rain-dominated watersheds in the US and compared against the Eckhardt digital filter estimates, with a focus on recession periods where streamflow approximates baseflow. Results show that the proposed model produces physically consistent baseflow estimates and captures event-scale variability, while performing comparably to the Eckhardt method. The proposed PIML framework further demonstrates improved robustness to data gaps in streamflow records, suggesting its potential for flexible and physically consistent baseflow estimation in data-limited settings such as cold regions, where stream gages are not operational during winter.