Study Of The Operation Function Of Step Reservoirs Based On Improved Back-propagation Neural Network

In the reservoir operation, the implicit stochastic optimization method with its advantage of overcoming the difficulties of calculation has been widely used.There are two parts included in this method:(1)For the given historical runoff,use the deterministic optimization theory to make the calculation;(2)Based on the results of the calculation,approximate the function to get the rules for operation function of reservoir.For the first part,the POA algorithm,which can overcome the problem of dimension disasters and can be carried out easily by programming,has been used widely.For the second part,the method of linear regression is commonly used to approximate the operation function.Howerver,the operation function is nonlinear usually,and it is difficult to confirm which kind of nonlinear function it belongs to.So the method of linear regression will inevitably cause some errors.Using the strong function of the neural network's nonlinear mapping, the nonlinear operation function will be approximated and then applied to the dispatching of multi-reservoir system.The instance shows good results this method takes.The key points and some relevant achievements are as follows:(1)The basic theories of the neural network,especially the back-propagation neural network and the improved methods are introduced.The improved methods can converge from ten to one hundred times faster than the typical algorithms when training the neural network.For medium-size networks that contain up to a few hundred weights, the Levenberg-Marquardt algorithm will have the fastest convergence.(2)In the case of not enough runoff provided,the technique of Bayesian regularization,which uses Levenberg-Marquardt algorithm for training and chooses Bayesian framework to determine the regularization parameters in an automated fashion, is taken to prevent the overfit.Using this method,a credible operation function is achieved,and also,the network has high approximation precision.(3)Water level of every end period and outflow are used respectively in the operation function,and then are combined to verify and complement each other.Through the simulative dispatching process for six-reservoir system consisting of downstream Jinsha river and the three gorges plus gezhouba,it can be seen that the operation function which has two decision-making variables and is approximated by neural network has more effective results than normal dispatching and linear operation function do.