Research On Numerical Prediction Of Non-Linear Slow Time-varying System Based On Neural Network

In this paper, numerical prediction of non-linear slow time-varying system is studied. There are many systems belonging to non-linear slow time-varying system in real world. The relationship between the input and output of this kind of system is non-linear and time-varying. In this paper, neural network is used as a tool for numerical prediction of non-linear slow time-varying system. Neural network has complex and non-linear mapping ability, and can adjust the weights adaptively by learning new samples. In this paper, numerical prediction of the nonlinear function with slow time-varying parameters, Mackey-Glass time series and the wind speed data collected from the wind farms, based on neural network are researched. The main work conducted is as follows:Firstly, BP neural network is applied for online prediction and batch prediction of the nonlinear function with slow time-varying parameters, Mackey-Glass time series and the wind speed data collected from the wind farms. For the the nonlinear function with slow time-varying parameters, online prediction performs unsatisfactory. For Mackey-Glass time series and the wind speed data collected from the wind farms, online prediction performs well. Batch prediction is relatively stable for all the three problems, while it takes more time to get the results. The momentum algorithm is used to improve the convergence speed of traditional BP algorithm.Then, the Cascade-Correlation network is applied for increment prediction of the nonlinear function with slow time-varying parameters, Mackey-Glass time series and the wind speed data collected from the wind farms. Cascade correlation network improves the convergence speed greatly because it solves the step-size problem and the moving target problem. Incremental prediction is stable and fast for all the three problems.After single value prediction, RBF neural network is applied for interval prediction of the wind speed data collected from the wind farms. The LUBE method and RBF neural network is combined to estimate the interval of output under a given confidence level. As the result of the LUBE method depends on the choice of initial value, a new interval prediction method is proposed:Create the RBF neural network to get the single value prediction, and calculate the residual error, then estimate the upper limit and lower limit of the output to initial the parameters of RBF neural network. The proposed method improves the result of interval prediction for the wind speed data collected from the wind farms.