Using sequential and non-sequential LSTM neural networks with global climate model forcing for daily streamflow reconstruction and flood projection: Applications to two contrasting flood regimes in Norway

Climate change is altering floods in northern, high latitude catchments. This study assesses the ability of Long Short-Term Memory (LSTM) neural networks to translate changes in large-scale atmospheric variables modelled by a global climate model (GCM) to long-term trends in streamflow (daily discharge and higher flood quantiles), considering two unregulated catchments representing differing seasonal flood regimes in Norway (rainfall and snowmelt) in the present climate. Two sequential (LSTM90 and BI-LSTM) and one non-sequential (RF-LSTM) architecture were trained on global reanalysis data and benchmarked against a hydrological model. The non-sequential RF-LSTM employing bias-corrected atmospheric inputs of predefined lagged features projects changes in the 200-year flood of similar magnitudes to estimates from the hydrological model in the rainfall-dominated (+46%) and the snowmelt-dominated (-25%) catchment. The LSTMs collectively project the expected intensification of autumn and winter floods in the rainfall-dominated catchment and the smaller and earlier spring floods in the snowmelt-dominated catchment. Moreover, statistically significant increasing (decreasing) trends in annual maximum flow are detected for the rainfall-dominated (snowmelt-dominated) catchment. Yet, modelling the expected shift from snowmelt-driven to rainfall-driven floods remains a challenge. This work demonstrates the utility and limitations of the proposed LSTMs for hydrological climate-impact modelling.