Under future warmer climates, drought events are projected to occur more frequently with increasing impacts in many regions and river basins. This study focuses on exploring the potential of the LSTM deep learning (DL) approach for operational streamflow drought forecasting for the Rhine River at Lobith with a lead time (LT) of up to 46 days.
The research investigates optimal spatial resolution, input and target variables, and loss functions. Four LSTM-based model architectures are developed and tested, incorporating both historical observation and forecast data to generate 46-step forecasts simultaneously. The robustness and stability of the models are assessed through cross-validation, and their performances are compared. Subsequently, the performance of the LSTM-based model is compared to the physically-based models, namely Wflow-Rhine and FEWS-Rhine, in forecasting streamflow drought.
The results suggest that utilizing a subbasin spatial resolution, including historical discharge as input, and training the model on time-differenced data enhance the forecast skill. Among the evaluated models, the model architecture with two LSTMs in cascade exhibits stable and robust performance across the forecast horizon and is considered for operational use in this study. Comparisons between the DL model and physically-based models indicate that: 1) When using observed meteorology forcing from ERA5, the DL model demonstrates a notable performance compared to Wflow-Rhine simulation using the same forcing data. 2) When utilizing SEAS5 for forecasting, the DL model demonstrates skill over Wflow-Rhine in predicting discharge levels during the dry season up to 10 days ahead, as well as for discharges between 950 and 2200 m3/s across the entire forecast horizon. However, for discharges between 700 and 950 m3/s with longer LTs beyond 20 days, Wflow-Rhine shows skill over the DL model. 3) While FEWS-Rhine successfully forecasts drought events in 2018 throughout the forecast horizon, it tends to produce more Type I errors (false positives). The DL model, forecasting with SEAS5, accurately predicts drought events in 2018 for LTs up to 30 days and generally has higher precision values. Despite using different forcing datasets, the DL model can predict the timing and trend of past drought events, indicating its potential in capturing streamflow patterns.
This study contributes to operational water management in the Netherlands by employing the LSTM deep learning approach in an operational framework for drought forecasting. By leveraging historical observation data and forecasted meteorology forcing data, these models achieve skillful performances for streamflow drought forecasts. Future research could focus on further enhancing model performance, exploring the applicability of the LSTM-based models in other river basins, and validating the results in real operational settings.

This study focuses on exploring the potential of using Long Short-Term Memory networks (LSTMs) for low-flow forecasting for the Rhine River at Lobith on a daily scale with lead times up to 46 days ahead. A novel LSTM-based model architecture is designed to leverage both historical observation and forecasted meteorological data to carry out multi-step discharge time series forecasting. The feature and target selection for this deep learning (DL) model involves evaluating the use of different spatial resolutions for meteorological forcing (basin-averaged or subbasin-averaged), the impact of incorporating past discharge observations, and the use of different target variables (discharge Q or time-differenced discharge dQ). Then, the model is trained using the ERA5 dataset as meteorological forcing, and employed for operational forecast with ECMWF seasonal forecast (SEAS5) data. The forecast results are compared to a benchmark process-based model, wflow_sbm. This study also explores the flexibility of the DL model by fine-tuning the pretrained model with limited SEAS5 dataset. Key findings from feature and target selection include: (1) opting for subbasin-averaged meteorological variables significantly improves model performance compared to a basin-averaged approach. (2) Utilizing dQ as the target variable greatly boosts short-term forecast accuracy compared to using Q, with a mean absolute error (MAE) of 25 m³ s⁻¹ and mean absolute percentage error (MAPE) of 0.02 for the first lead time, ensuring reliability and accuracy at the onset of the forecast horizon. (3) While incorporating historical discharge improves the forecasting of Q, its impact on predicting dQ is less pronounced for short lead times. In the operational forecast with SEAS5, compared to the wflow_sbm model, the DL model exhibits skill in forecasting low flows as evidenced by Continuous Ranked Probability Skill Score (CRPSS) median values of all lead times above zero, and better accuracy in forecasting drought events within short lead times. The wflow_sbm model shows higher accuracy for longer lead times. In the exploration of fine-tuning approach, the fine-tuned model generates marginal short-term enhancements in forecasting low-flow events over a non-fine-tuned model. Overall, this study contributes to advancing the field of low-flow forecasting using deep learning approach.

@en