Identification of the controlling factors for hydrological responses by artificial neural networks

Identifying the controlling factors for hydrological responses is of great importance for artificial neural network-based flood forecasting models, which are often hindered by the lack of physical mechanisms. To explore the first-order controlling factors of hydrograph patterns, a hybrid neural network was designed to analyse the impacts of potential driving variables with different temporal and spatial resolutions on hydrograph patterns. The Jinhua River Basin in Southeast China was used as an example in this study. Flood events with different hydrograph patterns and six external factors denoting potential controlling factors were individually classified into specific clusters using self-organizing maps (SOMs). Based on the back-propagation neural network (BPNN) and leave-one-out cross-validation methods, the controlling factors of different flood patterns were identified by comparing the performances of flood simulation models trained with datasets before and after the potential controlling factor classification. The results showed that (i) the classification of controlling factors indicating various runoff regimes significantly improved the performance of data-driven models in flood simulation in terms of correlation coefficient, Nash-Sutcliffe coefficient, and normalized root mean square error; (ii) the spatial distribution of antecedent soil moisture and vegetation conditions as well as the temporal distribution of rainfall dominated different hydrograph patterns; and (iii) the transition of dominant rainfall-runoff processes could be identified in an individual flood event using the hybrid SOM–BPNN model, indicating the varying influence of potential controlling factors on streamflow. Overall, the hybrid neural network models trained with datasets classified by controlling factors provide a general analytical framework to identify the governing dynamics for different flood patterns and improve the accuracy of flood simulations. Additionally, more attention should be devoted to improving the time to peak error of hydrological models, which cannot be settled by data-driven models trained with different data-splitting strategies.