A novel ensemble algorithm based on hydrological event diversity for urban rainfall–runoff model calibration and validation

Abstract

Listen

Translate article

Hydrological models such as the stormwater management model (SWMM) are essential tools for managing water resources. Model performance typically relies on calibration to historic rainfall–runoff events. However, traditional calibration techniques suffer many issues, such as ad-hoc event selection, ignoring model parameter uncertainty, and poor validation techniques. We introduce a novel ensemble framework that relies on data-driven event selection of calibration and validation events to achieve robust model generalisation.The proposed framework is applied to SWMM models for two urbanised catchments in Southern Ontario, Canada. The framework uses unsupervised machine learning to cluster rainfall–runoff events. Events in different clusters have diverse characteristics; sampling events from each cluster ensures a diverse set of events for calibration. Validation events can be sampled in the same manner, which ensures similarity between calibration and validation data. We study the effects that the calibration event cluster has on model sensitivity and generalisation. Only minor differences in model parameter sensitivity were observed between different clusters. However, the cluster(s) used for calibration were found to be very important for model generalisation. Next, we generate ensemble predictions, which use multiple SWMM models, each calibrated to different events. Ensemble performance is compared to the single-model baseline and across increasing numbers of ensemble members. Results demonstrate improvements in generalisation and decrease in prediction variance across increasing ensemble size. Finally, we compare several different techniques for ensemble combination, which increase in complexity from simple averaging to ANN-based stacking augmented with exogenous input features. The ANN-based stacking performs the best according to two out of three evaluation criteria. Simple averaging, while inferior in performance to stacking, consistently outperforms the single model baseline. The novel framework proposed in this work provides a data-driven, adaptable alternative to conventional calibration frameworks that have been proven to improve model performance.