Assessing robustness and uncertainty in rainfall–runoff modelling using LSTM ensembles

Long Short-Term Memory (LSTM) networks are widely used for rainfall–runoff modelling and have demonstrated strong performance in regional applications. A key advantage of LSTMs is their ability to learn from large samples of catchments, in contrast to traditional approaches that rely on individual catchment-by-catchment calibration. The objective of this abstract is to assess the robustness of regional LSTM models with respect to their behaviour at the individual catchment scale. To this end, ensemble training and simulation experiments were conducted using the CAMELS-GB and CAMELS-US datasets. An identical LSTM architecture was trained 100 times with different random weight initializations, and model performance was evaluated separately for each catchment. For a substantial number of basins, model performance varied strongly across realizations, with considerably larger variability observed for the CAMELS-US dataset. Excluding catchments with known data quality issues or highly nonlinear responses led only to minor improvements and a modest reduction in performance spread. Furthermore, large differences between validation and test performance were frequently observed, indicating that model skill is often not stable across evaluation periods for individual catchments. The results indicate that uncertainty estimates derived from ensembles of random initializations appear overconfident and do not reflect the full epistemic uncertainty.