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.

Due to global warming, temperatures are expected to rise and, with it, the increase in moisture holding capacity of the atmosphere. Changes in precipitation extremes with temperature are governed by the Clausius-Clapeyron relationship (CC), which states that precipitation increases 7% per degree of warming. While global models and observations are consistent with CC, research at hourly observations have showed deviations in the CC relationship, showing steeper increase up to 14% per degree (super-CC) and <7% per degree for larger temperatures (sub-CC). Studies have shown that scaling rates can be influenced by several factors, such as the length of observations, rainfall percentiles or the timescale of rainfall, among others. Apart from methodological limitations, research on scaling rates is not readily available in all parts of the world, with scarce studies carried out in the African continent. In this context, there is a need for better understanding the relationship between precipitation and temperature as well as the implications for future extremes. To address these challenges, this research investigated 3 main topics. Firstly, existing methodologies were compared and analysed for 353 rain gauges in Germany. Using as extreme rainfall indicators wet percentiles (80th-99th quantile), 3 methods were tested and statistically compared to estimate scaling rates with dew point temperature (DPT): (i) binning methods, (ii) quantile regression and (iii) change-point models at an hourly timescale. Secondly, the presence of super-CC was evaluated for storm events instead of fixed hourly intervals. For that, data from Germany at 10-min resolution was used. The last step involved the analysis of scaling rates for a dense network of newly available rain stations in 4 countries in sub-Saharan Africa: Ghana,Kenya, Tanzania and Uganda. The results from measuring the scaling of extreme precipitation with DPT shows that two regimes arise for Germany: for DPT below 13ºC- 15ºC the scaling rates are close to the CC rate across all quantiles. For warmer DPT of 15ºC or more, scaling rates above the CC are observed, with increasing slope for the highest quantiles. The study of storm events shows that, for mean intensity and peak rainfall intensity, there is a strong sensitivity to DPT, with higher DPT associated with higher rainfall. The analysis of storm duration indicates that shorter events are more frequent for larger temperatures than for shorter ones. The study in sub-Saharan countries shows consistent negative scaling rates with dew point temperature. In addition, for large surface air temperatures (SAT), there is a decrease in relative humidity, suggesting that there is a general lack of moisture availability at the higher SAT ranges. To better understand the physical processes that influence these results, the use of alternative predictors that describe the atmospheric moisture are proposed. In a future warming world, current literature indicates that the greatest rainfall increases will be more visible for short-duration storms rather than increases in annual mean precipitation. As such, within this project, a novel analysis on scaling rates for storm properties shows the potential of these approach to obtain detailed information on the storm properties and their relationship with temperature, which could be used in storm simulations and projects characterising current or future climate.