Leveraging Hydrological Models in Conjunction with Multi-Objective Optimization Based Methods to Design Streamflow Monitoring Networks

Abstract

Hydrometric data provides forcing data inputs to run hydrological models and observed output time-series to facilitate the calibration and validation process. Hydrometric monitoring networks are often designed without considering the innate relationship between data collection, model set-up, and model application. This research compares the relative effectiveness of a previously established model-based network design strategy to a newly proposed method. The traditional design method identifies a set of Pareto-optimal networks using intermediate entropy-based design objectives, facilitated by the dual entropy multi-objective optimization (DEMO) tool, and then applies models as a post-processing mechanism. Streamflow time-series from networks initially identified by DEMO are used to calibrate two semi-distributed rainfall-runoff models. The calibration process enables a reassessment for non-dominance based on the primary network design objectives, which are maximizing model performance at manually defined flood sensitive catchment outlets and minimizing network size. The newly proposed alternative method embeds the hydrological models and their calibration process into the optimization algorithm, resulting in direct optimization based on the primary design objectives. Both techniques were applied to design networks in two large western Canadian watersheds. Bubble maps are presented to illustrate variations in the spatial distribution of optimal solution sets, with respect to both model performance at flood sensitive catchments and individual station selection frequency, for all design scenarios. Results indicate the newly proposed method provides superior results regardless of network size and that trends in the spatial distribution of optimal networks are highly case-specific. The proposed methodology can be readily adapted to address a wide variety of design applications by varying the models and model performance criteria used in the design process. The findings from this research can be used to guide future network design projects when the proposed network is intended to support one or more model-based applications.