Using machine learning approaches for predicting suspended sediments in alpine catchments – uncertainties and limitations

Hydropower generation and the associated sediment management is one out of different water related services that are subjected to hydrological changes over time. Thus, the assessment and prediction of the sediment transported from catchments at varying temporal and spatial scales was and is an important task in hydraulic engineering. In this study, we focus on alpine catchments feeding a reservoir for hydropower production. Aim was to simulate and predict the suspended sediment input, which accounts for the vast majority of sediment loads.

The selected catchments, Pitzbach and Fagge, are part of the hydropower system Kaunertal Valley (Tyrol/Austria), operated by the TIWAG. The available measurements include discharges and turbidity/suspended solids contributing to the sedimentation of the Gepatsch reservoir. The discharge time series cover several decades, whereas turbidity was only measured during the recent years.

A combination of a process-based water balance modelling and a data driven approach to simulate sediment fluxes was combined to simulate extreme events and years as well as past periods where no material transport was measured.

For the two sub-catchments, different machine learning approaches were used to mimic suspended sediment transport, based on an available 11-year (2008-2018) long timeseries. Specifically, feed-forward neuronal networks (FFNN) and long short-term memory networks (LSTM), were tested and compared using different input combinations to identify the most suitable models for the respective catchment area.

For further validations the models were exanimated on a short “future” period (2019-2022), which was not part of the calibration. The model performance was evaluated for this time series, having a special focus on periods with exceptionally high transported sediment loads. For past periods (back until 1970), only discharge and reduced number of meteorological stations are available. Similarly, the models were applied to these periods in order to calculate sediment transport time series. On the one hand, a solely data driven approach using measured discharge and meteorological time series was tested. Beyond that, results from a process based hydrological model were used, aiming to cover also periods with gaps in the discharge data.

Overall, the simulations allowed to quantify the uncertainties associated to such modelling chains, when using them to describe sediment fluxes at different temporal scales.