Genetic algorithms for control of wastewater conveyance systems
Abstract
Mathematically rigorous techniques of global real-time control for reduction of combined sewer overflows usually rely on simplified formulations of the problem. Consequently, these techniques may fail to operate efficiently in the field. Additionally, these techniques may be susceptible to computational problems. On the other hand, other existing approaches, including heuristic methods and fuzzy logic, trade some of the mathematical rigor for ease of application, and, if properly designed, achieve near-optimum control. However, these methods usually require prior development by other techniques. This thesis presents one of the first formulations of a genetic algorithm (GA) for wastewater conveyance control. Feasibility of the approach has been shown for a simple control problem with a known solution. The GA approximated the optimal solution derived by linear programming. When a more realistic formulation of the problem was used, the GA outperformed linear programming. Next, the GA developed was used in a more practical setting, i.e., to solve a problem of controlling a pumped diversion within a separate sewer area. Historical data, a verified mathematical model, and simulation results for several feedback control algorithms were available for the test site. Several versions of a GA were developed to solve the same control problem: the minimization of pumping with a downstream flow constraint. The GAs outperformed the feedback algorithms in all tests. Although the flexibility of the GA allows any flow routing method, the use of accurate models is likely to prove too time consuming for on-line applications. The GA's flexibility can only be fully exploited if the time constraints are relaxed to those of the off-line methods. A fast on-line method is needed to complement the GA. To provide the required on-line method, an artificial neural network (ANN) was combined with the GA, and a GA-ANN methodology resulted. Multi-layer feed-forward ANNs need to be trained before they can be used. In this case, the ANN training relied on a portion of the data produced by a selected GA. The behavior of the ANN in the simulations with the reference model matched that of the GA both for the training data and for the testing data.
This paper has been withdrawn.