Multiobjective evolutionary computation for sanitary sewer overflow reduction

Over 10 billion gallons of raw sewage is discharge into the environment in the US annually as sanitary sewer overflows (SSOs). These events are public health treats and violations of the Clean Water Act. The raw sewage introduces dangerous pathogens to the environment and deteriorates the quality of receiving water bodies. Because the cost of controlling SSOs are substantial, water utilities generally spread the rehabilitation program over time. Typically, the network is segmented into multiple drainage areas and sewer rehabilitation is prioritized based on observed deterioration levels or previous SSO occurrences. This approach is costly, inefficient and time consuming. Additionally, rehabilitation in one part of the network could transfer the overflows to other parts of the system resulting in wasted expenditures. This dissertation presents three studies that utilized innovative evolutionary computation approach and procedure to solve SSO problems in a complex real world sanitary sewer system. The evolutionary algorithms (EA) implemented in the dissertation threats the sewer system as holistic units, therefore the rehabilitation plans generated by the EAs takes into account the effect of the solutions on other part of the system.

The first study tested the efficacy of evolutionary computation to identify solutions for SSO rehabilitation problems. A simulation--optimization platform, the Sanitary Sewer Overflow-Reduction Optimization (S2O2), was developed that linked the hydrodynamic model, USEPA SWMM to evolutionary algorithm engine. In the study, single and multiobjective genetic algorithms (GAs) were utilized to identify solutions that reduce or completely eliminate SSO in the network. For the multiobjective problem formulation, the Nondominated Sorting Genetic Algorithm II was used to generate nondominated sets of solutions that characterizes the tradeoffs between reduction in number of SSOs and cost (Case I), and the tradeoff between of volume of SSOs and cost (Case II). The results show that, when maximizing the reduction of number SSOs, the algorithm target first regions of the network with higher density of SSOs. When maximizing the reduction of volume of SSOs, the solutions prioritize the nodes with the largest overflow volumes. The tested approach provides a range of options to decision makers that seek to reduce or eliminate SSOs in an existing sanitary sewer system.

The second study address the high computation demands that may discourage the acceptability of EA by decision makers. The study introduced a novel MOEA, the Enhanced Nondominated Sorting Evolution Strategy eNSES that utilizes a specialized operator to guide the algorithm towards known SSOs locations in a sewer system. In the study, the eNSES was compared to NSGA-II and the NSES based on hypervolume and the overall nondominated vector generation ratio (ONVGR). The results show that the eNSES improves convergence rate by approximately 70% over the tested alternative algorithms, performing as well as NSGA-II and outperforming NSES in terms of the hypervolume by nearly 10%. In terms of the ONVGR, which reflect the ability of an algorithm to generate nondominated solutions, eNSES performed similar to the NSES but outperforms NSGA-II by 42%.

The first two studies focused on the optimization of the model objectives: (1) maximize SSO percent reduction, (2) minimize rehabilitation cost. Decision-makers however, often have to consider other socio-cultural objectives that are difficult to model quantitatively. Additionally, other objectives may arise later that may not present during problem formulation. To address these limitations, a subpopulation algorithm is needed that generate quantifiable good solutions that are similar in the objective space but very different in the decision space. The alternative solution will potentially address to various degrees the model and unmodel objectives. In the third study, a subpopulation MOEA, the Multiobjective Niching Co-Evolutionary Algorithm (MNCA) is utilize to identify distinct set of alternative solutions to the tradeoffs generated from the Nondominated Sorting Genetic Algorithm (NSGA-II). The results show that the MNCA produced alternatives that are very dissimilar to the modeled objectives in the decision space, thereby providing decision makers with needed flexibilities to select from diverse options while meeting similar objectives.