Daily discharge prediction is one of the key challenges in the field of hydrology. Difficulties are encountered for unraveling the complicated non-linear physical mechanisms behind runoff generation. Limitations of hydrological models to predict discharge are identified as a result of inconsistencies found between the hydrological model derived discharge and observed discharge. It is the objective of this study to bring both physical insight and data perspectives to improve discharge predictions.
As a newly developed framework, KGML (knowledge guided machine learning) is applied to achieve this purpose. KGML leverages machine learning models for knowledge discovery through physically-based models. Deep learning techniques are applied by coupling them with the SWAT (Soil and Water Assessment Tool) model, which provides physical insights. A small catchment located in the South Branch of the Root River Watershed in southeast Minnesota was selected for a case study. Currently, based on a 200-year synthetic dataset generated by SWAT, the KGML framework was implemented to predict discharge from the same meteorological data used as input for the SWAT simulation. In the early phases of this study the ability of the ML models to emulate the synthesized data has been moderate because the models are not yet complex enough. Complex interactions in the physics model (e.g., snowmelt, soil freezing, evapotranspiration) require ML models with more advanced structure to capture these complexities.
Guided by this investigation and the interpretation of machine learning models, progress is being made to incorporate the features of physical processes underlying runoff generation and other state variables, such as soil moisture, into machine learning models. As more complexity is introduced into the present implementation of KGML, the research result of this case study will be generalized to more sophisticated cases where spatial heterogeneities are evolved.
Speaker(s)
Xiang Li, John Nieber, Vipin Kumar, Ankush Khandelwal, Jared Willard, Xiaowei Jia, Michael Steinbach, Shaoming Xu, University of Minnesota Twin Cities; Christopher Duffy, Penn State University