Use of Genetic Algorithms for Collection System Rehabilitation Planning

The Town of West Seneca (Town), a suburb of Buffalo New York, owns and operates a sewer system with pipes ranging in size from 8 to 48 inches in diameter that has been the subject an Order on Consent (Order), which the Town has been administering since 2004 (Figure 1). The goal of this Order is to eliminate sanitary sewer overflows (SSOs) through a program of sewer rehabilitation projects that mitigates infiltration and inflow (I/I) in the sewer system. 20 years later, the program is currently in its eighth phase of a ten-phase mitigation plan. Significant tasks planned in this phase include lining the main 36-inch trunkline in Flowmetering Basin (FM) 19 and FM 7, which follow Cazenovia Creek to the interconnection with the Buffalo Sewer Authority. This program to date has improved sewer performance and reduced I/I volumes by approximately 40% but has not eliminated overflows within the system. The Town has noticed diminishing returns on the amount of infiltration removed through lining in recent phases of the program. As a result, they were interested in determining whether there are changes to the work plan that can improve outcomes and prevent the activation of their remaining active SSOs. The Town chose to take advantage of a planned 'break year' while large diameter sewers are being lined to re-evaluate the current program and determine if changes to the approach could improve outcomes. The current approach, called the 'status quo plan' is to complete the final 2 phases of the sewer rehabilitation program by lining approximately 187,000 linear feet of sewer main. This has been a standard approach to system rehabilitation since the early 2000's and the analysis took the opportunity to consider additional options like replacing existing sewers with larger sewers and newer technologies, lateral lining, micrometering, and permanent flow metering among other items. Identification of an optimal plan required evaluating combinations of improvements throughout the Town. To do this, the Town's sewer model was converted to run in SWMM5 and updated to better simulate current flows using flow data collected during post-construction evaluations. This model was then imported into a genetic algorithm optimization software to perform a multi-objective analysis on system improvements and identify a plan that would improve on the existing approach at a lower cost to rate payers. The genetic algorithm is an engineer's decision-making tool that works within existing models like SWMM and EPANET that performs hundreds to millions of parallel simulations on combinations of system improvements. The algorithm tracks costs associated with implemented projects and penalties that can be associated with unwanted system behavior like overflows and surcharging. Each simulation produces a tabulation of costs and penalties that can be used to identify one or many local optimal solutions. The collection of these local optimums correlates with the shape of a pareto distribution. The analysis was performed by preparing detailed probable unit costs for a range of improvements to the sewer system. These costs were assigned to various sewer assets so that the optimization algorithm could select and simulate improvements, while balancing the tradeoffs against the associated cost. Unit costs attempted to include both pipe-specific costs and associated costs like surface restoration and new manholes that's associated with larger pipe to minimize unrealistic biases in results produced by the algorithm. Improvements that were simulated included sanitary sewer/manhole lining, lateral lining, increased conveyance capacity, and off-line storage. The existing conditions model assumed that interceptor lining as part of year 8 was complete and that operations staff would proactively remove blockages when discovered because the newly lined sewer should be less likely to become obstructed. The goal was to determine if wet-weather overflows could be mitigated more effectively at a lower cost compared to the status quo plan using these assumptions. The multi-objective optimization analysis examined more than 150,000 plan simulations, testing out different combinations of the intervention options. This optimization used an evolutionary algorithm, which is an AI process based on the concepts of natural selection. The optimization identified a range of solution plans that best balanced the tradeoff between cost, hydraulic performance, and overflow volume. This optimal set of cost-effective solutions is represented as a Pareto Front shown in Figure 2, with the set of solutions that were considered 'sub-optimal' represented in the background. Plans were selected from the Pareto Front and evaluated to determine which improvements had a high probability of being selected by the model. To simplify the analysis, an evaluation of 'standard' capital improvements was performed in Phase 1 (Figure 3), and lateral lining was added to a secondary evaluation in Phase 2 (Figure 4). The model tended to group similar plans together in similar parts of the Pareto Front and were named to help with differentiating model results between each other. Results from the optimization analysis suggested that the addition of lateral lining to the FM 6 sewershed has the potential to significantly reduce flows to some of the most active overflows because this result showed up in nearly all plan groups. Similarly, lateral lining and aggressive I/I removal within the FM 15 and FM 19 sewershed could reduce backwater effects on the FM 6 sewershed and further reduce overflows. The benefits are dependent on keeping sewers flowing freely so permanent flowmetering at key locations was also recommended to help monitor sewer levels so that blockages could be removed before overflows occur. Model outputs were subject to potential biases that needed to be considered by the modeler when evaluating alternatives. Some examples include: - Accuracy of model calibration could bias output towards preferring rehabilitation in different sewersheds if the calibration is not within generally accepted tolerances. - Assumptions for how much flow can be removed by sewer and lateral lining can also make the algorithm prefer lining to new sewers. This was controlled by analyzing how sewer lining affects changed in unit inflow hydrographs. - Biases for costs were controlled by attempting to make unit prices account for most inputs for a particular improvement and limit the model to evaluate only construction costs based on recent bid prices. The recommended program is anticipated to significantly mitigate overflows within the system at a probable construction cost of approximately $4.4 million (2024 Dollars). The status quo program is anticipated to have a probable construction cost of $8.2 million. The proposed program is also anticipated to reduce overflows by about 900,000 gallons during the design event when compared to the status quo model. The predicted overflow volume is low enough that overflows can be effectively controlled under most rain events at a significantly lower cost to residents.