Combining data-driven machine-learning and process-based models for streamflow simulation: Preliminary results from the ExaSheds project

Abstract:

The ExaSheds project was initiated in March 2019 to explore synergies between data-driven machine learning (ML) approaches and process-based hydro-biogeochemical simulation capability with a goal of improving predictive capability for watershed function. As part of that project, we are developing and testing hybrid approaches that combine process-resolving simulations with Long Short-Term Memory (LSTM) networks. LSTMs, a class of recurrent neural networks, have emerged as a promising machine learning approach for simulating streamflow and an alternative to traditional hydrological process-resolving simulations. The long-term vision for the ExaSheds project is to use high-resolution integrated surface/subsurface models as a process-based hydrological model to improve robustness of LSTM in a non-stationary climate. However, in initial testing activities we used well-established semi-distributed models. In our hybrid approach, the outputs of semi-distributed hydrological models were used along with precipitation inputs to train LSTMs. When applied to streamflow from 671 catchments from across the US, the hybrid approach outperformed both the standard LSTM and the process-based model (Konapala et al. 2020). Specifically, the hybrid model produced significantly better median Nash-Sutcliffe Efficiency (NSE) than LSTM and the process-based model. Moreover, the hybrid model produced fewer “catastrophic failures” (NSE<0). When applied to catchments with short data records, the hybrid model was significantly more robust than LSTM, producing acceptable NSE values in conditions outside the range of the training data, as opposed to LSTM, which suffered catastrophic failures (Lu et al. 2020). These results suggest that hybridization of ML and process-based models may provide a robust simulation capacity for projections in a non-stationary climate.