Assessing and optimizing green infrastructure designs for stormwater management using relative performance evaluation framework and data-driven approach

Abstract

Green infrastructures (GIs) are considered as environmentally-friendly alternatives to the conventional stormwater drainage infrastructures. Their hydro-environmental impacts and design optimization have been investigated in various previous studies. However, there exist a few common challenges confronting the modeling, assessment, and optimization of GIs.

The first challenge is that most of the current performance assessment frameworks of GIs do not explicitly account for the preferences of stakeholders and commonly involve a tedious parametrization process for valuing the intangible hydro-environmental impacts of GIs. This thesis thus proposes an integrated assessment framework that uses relative performance evaluation (RPE) methods to tackle this challenge. In this framework, the performance of a GI design alternative for each impact of interest is measured by its relative effectiveness as compared to the other considered alternatives. The relative effectiveness obtained for different impacts can be then aggregated to derive the overall effectiveness of an alternative. During aggregation, specific weights can be assigned to the different impacts to reflect the preferences of stakeholders. This assessment framework can be useful for comparing and selecting GI design alternatives when multiple performance indicators and various stormwater management interests are considered.

Second, in GI-related studies, the accuracy of the commonly-used process-based hydrological models is sometimes affected by their model structure and the availability of field measurements. Data-driven modeling approaches can potentially avoid these issues by directly modeling the correlation between the input (e.g., rainfall time series) and the response (e.g., outflow hydrograph) of a system. Deep learning is a specific type of data-driven modeling that is capable of modeling high-dimensional data. Deep learning is used in this thesis to solve various problems in several GI sites, e.g., high-resolution rainfall-runoff modeling in a complex urban catchment with multiple GIs, and overflow occurrence prediction in a bioretention cell with censored field observations. This thesis shows that deep learning models can achieve comparable or better prediction accuracies when compared to calibrated process-based models and conventional machine learning models. Deep learning models and calibrated process-based model are found to have similar responses when tested under various rainfall conditions. This thesis also shows that knowledge in runoff generation processes can improve the design of deep learning models, and a model’s hidden states can be useful for confirming its credibility and generating hydrological insights.

Third, researches on the optimal design of GIs for multi-objective stormwater management are limited. This thesis thus applies the RPE method to identify suitable implementation levels of GIs to concurrently achieve multiple stormwater quality and quantity management targets. This study also proposes a method to summarize and generalize simulation results for rapid assessment and selection of GI alternatives considering multiple targets.

In conclusion, this thesis evaluates the effectiveness of various potential solutions to several common challenges in GI-related studies to generate insights to inform decision making in stormwater management. The frameworks, methods, and tools developed for solving individual challenges, in the end, form a comprehensive simulation-optimization framework which improves the current practices of identifying and deriving optimal GI designs.