Optimization of Green Infrastructure Networks to Maximize Stormwater-Related Benefits and Minimize Life Cycle Costs Using a Noisy Genetic Algorithm and Machine Learning

Green infrastructure (GI) has become a common solution to mitigate stormwater-related problems such as water quality and flooding hazards. Despite  widespread acknowledgement of GI benefits, there is a lack of decision support methods that allow practitioners to identify optimal locations and evaluate the costs and benefits of numerous spatially distributed small GI practices at larger scales (subwatershed to entire watershed) under uncertainty. To address these needs, an online Cloud-based interactive tool coupling SWMM (Storm Water Management Model) and the Water Research Foundation LID life cycle model, , called Interactive DEsign and Assessment System for Green Infrastructure (IDEAS_GI), is optimized using a noisy genetic algorithm (GA) with life cycle costs and stormwater volume reduction as the primary objectives. To overcome the computational challenge of probabilistic sampling with the noisy GA and to identify significant features for preferable locations, the GA  is merged with an artificial neural network, which acts as a meta-model (surrogate) for the numerical simulation model (SWMM). Post-optimization, machine learning decision trees are also generated that classify the numerous potential solutions generated by the noisy GA into GI coverage classes based on sub-watershed parameters. This framework is applied to a watershed in Baltimore, Maryland, U.S., under multiple budgetary scenarios. The results suggest that the greatest GI investments under the highest and lowest budgetary scenarios should be allocated to subwatersheds closest to the watershed outlet. For the lowest scenario, GI practices should be installed only in subwatersheds closest to the watershed outlet. When the budgetary scenario is highest, GI is sited across the watershed but highest priority is still given to subwatersheds closest to the watershed outlet. On the other hand, the importance of total distance to the watershed outlet is lower for the medium budgetary scenario. In fact, the impacts of different features for preferable GI coverage for these solutions are more complex, don’t follow a consistent pattern, and require more depth to capture the patterns in their corresponding classifier decision trees. In addition to these GI findings, the results showed that the addition of meta-models decreases average computational time required to reach Pareto frontiers similar to the ones generated by the noisy GA by more than 95%.