LSTM Networks for Catchment Response Simulation

Catchments convert precipitation into water flows for rivers and other surface water bodies. Transforming rainfall to runoff is nonlinear and complex. Initially, the use of deep neural networks in runoff modeling was limited, but this changed with the successful use of long short-term memory (LSTM) networks in runoff preditiction. This research delves into LSTM-based runoff models, examining them concerning hydrological concepts like catchment timing and regionalization, specifically within the French context. Both time-variant data and time-invariant attributes, spanning 376 French catchments, are utilized. A regime classification based on three hydroclimatic variables, derived from the analysis of interannual monthly regimes of runoff, total precipitation, and temperature, is used and validated. Three LSTM training levels are conducted resulting in three primary model types: SINGLEs (each catchment gets its LSTM model), REGIMEs (an LSTM model trained per hydrologic regime using data from respective catchments), and NATIONAL (trained using data from all catchments at the national level). The study finds that: (a) model performance often improves with longer sequence lengths up to a certain point, after which it plateaus or declines, varying by hydrologic regime; (b) the Uniform and Nival regimes demonstrate the most sensitivity to LSTM sequence length, indicative of their long-term hydrological dynamics; (c) regional LSTMs don’t always offer superior performance, and in some cases, local LSTMs can achieve minimal error with adequate high-quality data; (d) training at both the regime and national levels produces similar results, but the HYBRID NATIONAL LSTMs, which combine national training with localized hyperparameter tuning, yield the most optimal outcomes.