From Events to Insights: Event-Process Based Diagnostics of Hydrological Model Performance

Hydrological models play a crucial role in understanding and managing water resources. However, accurately representing complex streamflow generation processes remains a significant challenge. We introduce an innovative diagnostic framework designed to evaluate process limitations in hydrological models, emphasizing event-based and multi-dimensional assessments. The framework first evaluates model error variability by classifying streamflow events into distinct types (e.g., Snow-or-Ice, Rain-on-Dry, Rain-on-Wet) and leveraging multi-dimensional metrics (i.e., timing and relative magnitude errors). It then assesses the importance of error drivers (e.g., hydrographic properties, model fluxes and states, model inputs, and pre-event errors) using explainable machine learning (XAI). A case study involving 340 German catchments demonstrates the framework's applicability. The results reveal that the majority of model-simulated streamflow events exhibited time delays and magnitude underestimations. Specifically, Rain-on-Dry events showed higher timing errors, while Snow-or-Ice events had larger relative magnitude errors. Furthermore, errors varied across different hydrograph components (pre-event, rising limbs, peaks, and recession limbs) for each event type. Simulated streamflow at all components, especially peaks, was predominantly delayed in timing and underestimated in magnitude in more than 50% of events. Using Random Forest regression with Accumulated Local Effects, the analysis found that pre-event errors are the dominant driver for both timing and relative magnitude errors across all event types. The relative magnitude errors were also strongly affected by hydrograph-related event properties and model fluxes and states for land surface and groundwater dynamics, with these drivers having greater importance for Snow-or-Ice events. This framework enhances diagnostic capabilities, providing a robust tool for advancing hydrological model evaluation and understanding under diverse hydrometeorological conditions.