To bucket or not to bucket? Analyzing the performance and interpretability of hybrid hydrological models with dynamic parameterization

Abstract. Hydrological hybrid models have been proposed as an option to combine the enhanced performance of deep learning methods with the interpretability of process-based models. Among the various hybrid methods available, the dynamic parameterization of conceptual models using long short-term memory (LSTM) networks has shown high potential. We explored this method further to evaluate specifically if the flexibility given by the dynamic parameterization overwrites the physical interpretability of the process-based part. We conducted our study using a subset of the CAMELS-GB dataset. First, we show that the hybrid model can reach state-of-the-art performance, comparable with LSTM, and surpassing the performance of conceptual models in the same area. We then modified the conceptual model structure to assess if the dynamic parameterization can compensate for structural deficiencies of the model. Our results demonstrated that the deep learning method can effectively compensate for these deficiencies. A model selection technique based purely on the performance to predict streamflow, for this type of hybrid model, is hence not advisable. In a second experiment, we demonstrated that if a well-tested model architecture is combined with an LSTM, the deep learning model can learn to operate the process-based model in a consistent manner, and untrained variables can be recovered. In conclusion, for our case study, we show that hybrid models cannot surpass the performance of data-driven methods, and the remaining advantage of such models is the access to untrained variables.