Predicting estuarine salt intrusion with a long short-term memory model

Abstract

Estuarine salt intrusion causes problems with freshwater availability in many deltas. For water managers to mitigate and adapt to salt intrusion, they require timely and accurate forecasts. Data-driven models derived with machine learning can help with this, as they can mimic complex non-linear systems and are computationally very efficient. We set up such a model for salt intrusion in the Rhine-Meuse delta. The model predicts chloride concentrations at Krimpen aan den IJssel, an important location for freshwater provision. As input features, we selected observations of water level, discharge, chloride concentration and wind speed. We then used the Boruta algorithm to select a subset of relevant features. We set up a Long Short-Term Memory network (LSTM) to make predictions of chloride concentrations one day ahead and ran the resulting model multiple times to simulate a multi-day forecast. This model predicts baseline concentrations and peak timing well, but peak height is underestimated, a problem that gets worse with increasing lead time. Because this model is reasonably successful, we aim to extend it to other locations in the delta. We also expect a similar setup can work in other deltas, especially those with a similar or simpler geometry. A more complete version of this model should finally be made suitable for use in an operational forecasting system.