Study On Direct Optimal Dispatch Model For Large-scale Water Distribution Systems

Water distribution systems are the lifeline of a city which concern the townsman life and the city economy. Now the illogical operative control of water distribution systems causes huge energy waste, leakage or even explosion of pipe networks, so it is difficult to meet the demand of receivers with the appropriate amount and high quality of water in any period of time. This paper aims to build and solve the operative control model of water distribution systems in Hangzhou.Most macroscopic state models in China are developed based on BP neural network because of its strong nonlinear mapping ability. However neural network suffers from problems like the existence of many local minima and the choice of the number of hidden units, and few researchers solve these problems. In this paper, a binary-coded self-adaptive GA is employed to search in the topology space, and for every topology, a real-coded self-adaptive GA is used to search the optimal initial weight and threshold of BP network so that the generalization ability of BP neural network can be well evaluated after training. In this way, the optimal topology can be found by using binary-coded self-adaptive GA. In order to solve the problem of BP algorithm's local optima, chaos genetic algorithm(CGA) is proposed to optimize the weight and threshold of neural network. The case analysis in Hangzhou shows that the neural network-based state model optimized by CGA has higher predictive accuracy. Furthermore, by using the same methods, the relationship between pump station flow and system pressure demand is established based on BP neural network to form a base for the optimal operation model.As BP neural network based on the principle of empirical risk minimization may overfit the training data and its structure is usually difficult to be determined, an hourly water demand forecast model based on Bayesian least squares support vector machine(LSSVM) is proposed suiting to periodicity and trend of water demand series. LSSVM employs the idea of structural risk minimization which balances the empirical risk minimization and VC dimension of the learning machine to achieve high generalization performance. Bayesian least squares support vector machines can shorten the time of modeling greatly than traditional LSSVM whose parameters are determined by using cross-validation. The hourly water demand prediction for the working days and the playdays shows that the hourly water demand forecast model based on Bayesian LSSVM has better predictive performance and higher modeling speed than BP neural network and traditional LSSVM.Traditional optimization algorithm can't find good solutions to optimal operative model in high dimension space, thus orthogonal agent evolutionary algorithm is proposed to solve the direct optimal operation model in Hangzhou. Orthogonal operation is applied to initial population in order to find the optimal initial agents, and with operators of competition and self-learning, agent evolutionary algorithm has high global and local search performance. Compared with orthogonal self-adaptive genetic algorithm, orthogonal agent evolutionary algorithm has better global search performance and higher convergence speed. The direct optimal operation schedule obtained by using orthogonal agent evolutionary Algorithm greatly enhances the pump efficiency, and reduces electricity expanse by 2.96% compared with experience dispatch schedule.