A neural - optimal control algorithm for real-time operation of combined sewer systems

Most urban populations are growing rapidly worldwide, at unpredictable rates, which increases the quantity and reduces the quality of the storm water discharge. During the flood period, urban drainage areas that have combined sewer networks may directly release part of the drainage water into the receiving water. Therefore, two objectives of real-time control for urban drainage system are to reduce the amount and the frequency of damage, and to reduce the environmental impact caused by pollutant discharges. In the last decade, there have been many attempts to improve and to maintain the urban living environment by improving the efficiency of the combined sewer system. Researchers are still searching for an efficient optimization or a control technique toward simplification of analysis and reducing the requirement of computer time and memory.; The rainfall-runoff module, the flow routing module, the optimization module, and the neural control module are four modules that are required for the real-time control model of the dissertation. The main purpose of the dissertation is to explore and to demonstrate an effective dynamic neural network model for the real-time control module. The rainfall-runoff module uses kinematics wave routing techniques to simulate the rainfall hyetograph become an inflow hydrograph. The Preissman Four-Point implicit scheme is the technique to solve the St. Venant equation in the hydraulic or flow routing module. The neural control module is Jordan's neural network architecture. The training process for the neural control module uses optimal policies that were produced by the optimization module as the reference of the desired output. The optimization module uses the Fletcher-Reeves conjugate gradient method to solve the direct optimization algorithm of a discrete form for the Potryagin's Maximum Principle (Optimum Control Theory).; A typical three-layer of an artificial neural network with Jordan architecture gives a good result of dynamic urban drainage real-time control. The training or learning process is a step to estimate parameters or weight of the neural networks. The back propagation learning technique is a simple and a common supervised learning technique for determining the model weights. The testing process is a validation of the weights that were produced in the training process. The neural control module is considerably faster to evaluate and to determine optimal gate openings. Therefore, an artificial neural network technique is an effective and a viable tool for a real-time control module. Many possibilities for further research in the area of learning technique and application of the model are still open.