Benchmarking physics-based, machine learning, and hybrid hydrology models at multiple catchments in Southern Norway.

Accurate inflow forecasts play a crucial role in the daily operations of hydropower reservoirs. Practitioners in the hydropower industry typically combine physically based models coupled with weather forecasts to produce inflow forecasts to the reservoirs. Despite the emergence of physically-based hydrological models since the early 1960s, their growing complexity has posed challenges in usage and calibration. Recent work by Kratzert et al. (2018) suggests that non-linear regression models such as LSTM neural networks (Hochreiter & Schmidhuber, 1997) may outperform traditional physically based models. Given the plethora of hydrology models, it is crucial to identify the most effective configurations within a diverse range of catchments using objective quantitative performance criteria.

This research aims to evaluate various model configurations across multiple catchments, determining the optimal hydrological model for streamflow prediction. Two physics-based models, the Distributed Regression hydrological Model (DRM) by Matheussen at Å Energi and the Statkraft Hydrology Forecasting Toolbox (SHyFT) from Statkraft, were applied alongside two versions of LSTM models tested as standalone and hybrid models with different input and model configurations. Thirteen model configurations underwent testing in sixty-five catchments in southern Norway. The models, including LSTM networks, were trained on either one catchment (Local) or all catchments (Regional) and tested using two train/test periods and two objective criteria: Nash-Sutcliffe Efficiency (NSE) and Kling Gupta Efficiency (KGE). The validation scores for NSE and KGE during the two train-test periods were used for benchmarking. 

Daily observed climate and streamflow data stems from The Norwegian Water Resources and Energy Directorate (NVE), The Norwegian Meteorological Institute, Å Energi's internal databases and ECMWF (ERA5). We extracted digital elevation data, land cover types, and vegetation information from 
www.hoydedata.no and the CORINE Land Cover inventory. The key findings of the study shows that the data-driven model outperformed physically based models (SHyFT, DRM). Hybrid models, incorporating output from physics models and meteorological data, surpassed purely data-driven models. The most successful configuration involved two hybrid models, utilizing an LSTM network forced with outputs from physically based models and climate  forcings. The results clearly demonstrates that information about the physical hydrological processes enhance the LSTM model performance.