What should we actually regionalize? - The benefits of temporal aggregation for low-flow prediction.

Predicting low-flow characteristics in ungauged basins is crucial for effective water management. Regionalization of low flow can either directly focus on lumped characteristics, such as mean annual minimum (MAM), or on seasonal or monthly characteristics (e.g., mean winter minima, mean summer minima, monthly mean minima). Alternatively, regionalization can focus on the time series (e.g., annual, monthly, or daily time series), as in rainfall-runoff models, which are subsequently used to predict the characteristics of interest. Most studies to date have regionalized runoff characteristics separately, leading to inconsistencies for each catchment. We propose regionalizing a full time series for each site to derive all low-flow characteristics from this single time series.

We regionalize daily streamflow and monthly, seasonal, and lumped low-flow characteristics using the US-CAMELS dataset. Low-flow characteristics are derived from the 7-day average streamflow, allowing us to compare annual, seasonal, and overall minima across different regionalization methods. Our approach leverages state-of-the-art machine learning models, such as tree-based models, support vector regression, and deep-learning architectures. For rainfall-runoff modeling of daily streamflow, we use an LSTM model tailored to low-flow prediction with an expectile loss function. Model validation is performed using 10-fold cross-validation. We evaluate our approach not only with common error metrics - such as RMSE and MAE - but also by quantifying the error in estimating the extreme value distribution of annual minima from the predicted time series.

Our results indicate that higher temporal resolution yields higher prediction accuracy compared to lumped characteristics. However, tailoring daily streamflow predictions to the lower quantile of the data is essential for more accurate results.