Bridging The Gap Between Physical-Conceptual Modeling And Machine Learning For Catchment-Scale Rainfall-Runoff Modeling
Presenter: Yuan-Heng Wang1
Co-Author(s): -
Advisor(s): Dr. Hoshin V. Gupta
1Department of Hydrology and Atmospheric Sciences, University of Arizona
Panapto Presentation Video
Oral Session 2: Data-Driven and Physically-Based Modeling
Long Short-Term Memory networks (LSTMs) are currently the most accurate and extrapolatable predictive streamflow models available. However, their lack of physical explainability impedes widespread acceptance of these ML-based models by the hydrological science community. To address this issue, we propose a Physical-Conceptual LSTM (PC-LSTM) structure that is suitable for representing mass/energy-conserving processes. The goal is to demonstrate how the inherent isomorphism between LSTM structures and dynamical systems enables the representation of important hydrological process in a meaningful manner, thereby bridging the gap between current process-based modeling and machine learning. We develop a specific PC-LSTM structural node called the “Finite Capacity Gated non-linear Reservoir with Loss” (FCGRL) and use it to simulate the 40-year daily time-step rainfall-runoff behavior of the Leaf River in Mississippi. A single FCGRL cell/node of the PC-LSTM architecture, represents a generic bucket model, where the “input gate” controls how much of the water input enters the node and how much bypasses the node on its way to the output, a newly introduced “loss gate” accounts for any external losses of water from the system, and a constraint is imposed on the “output”, “loss” and “forget” gating functions to ensure conservation of mass. Particular attention is given to exploiting machine-learning technologies that 1) ensure robust training to achieve good generalization performance, 2) enable simulating both the system output and its prediction uncertainty, and 3) enable various plausible physically-meaningful serial and parallel network architectures to be explored. Overall, this study paves the way for the development of flexible system architectures that can facilitate the development of synergistic physics-AI modeling approaches.