Automatically discovering interpretable rainfall-runoff models
Date:
Short abstract. Hydrologic sensor networks are expanding rapidly and generating more data every day. As the volume and resolution of these data sets grow, manual analysis, model calibration and quality control become infeasible. Few existing approaches can transform raw data into interpretable results. Process-based models are not designed to consume large amounts of real-time data, while machine learning models are typically uninterpretable, limiting insight and trust. To address this gap of automation and interpretability, we present a scalable approach that discovers differential equations and latent state estimations in water systems using only rainfall and runoff measurements. This method generates approximations of watersheds as nonlinear, time invariant dynamical systems automatically from measurements. We capture rainfall-runoff relations for catchments and combined sewers of all scales using between five and thirty parameters. We also demonstrate the method’s potential for surrogate modeling by replicating the dynamics of a large process-based model at a small fraction of the computational complexity. This parsimonious representation of watershed dynamics provides theoretical insight and the computational efficiency to enable automated predictions across large sensor networks.
Keywords. model discovery in partially observable dynamical systems, rainfall-runoff modeling