Temprerature Predition Model Based On Improved PSO-RBF Neural Network

To advance the accuracy of weather forecasting plays a significant role for reducing the effect of meteorological disasters. So far, the commonly used numerical forecast products including ECMWF are seeking to find the nonlinear relationship among meteorological elements, thus to improve the forecasting results. However, numerical forecasting products also exists some shortcuts, such as, in the light of different weather factor, numerical prediction will have a large deviation. With the development of artificial neural network technology, researchers carried out a lot of study work in the field of neural network combined with meteorological forecast.RBF neural network has a great advantage in ability of classification, generation and approximation, appropriate for meteorological prediction. But RBF also exists some problems and obstacles in practical operation and application. How to confirm the number of nodes in hidden layer, hidden layer centers, widths and the connection weights, is the problems we faced in establishing the RBF neural network.Compared with other intelligent algorithms, particle swarm optimization possess the characters of few parameters and strong ability of global optimization, is constantly applied to the optimization of neural network. Nevertheless, particle swarm optimization is easy fall into the some disadvantages of local optimization.This paper around radial basis function(RBF) hidden layer center of neural network, width and the problem of connection weights are not easy to identify, as well as the issue of particle swarm optimization is easy to fall into local optimization, so that we can establish a effective temperature prediction model. In this paper,the major research goal is temperature forecast mode which based on the improvement of PSO optimized RBF neural network.The main works can be showed as follow:(1) In view of the lack of frequently used method of time series analysis, we adopt RBF neural network to predict time series. Meanwhile, summarize the process of RBF neural network prediction model of time series of the creation. On the other hand, analyse the structure characteristics of RBF neural network the main learning algorithm, the generation of RBF neural network mathematical model and the function of related parameters deeply, and we also analyzed the primary study algorithm to determine RBF neural network parameters, points out the shortage of these algorithm. Furthermore,it is pointed out that the focus of this paper is using particle swarm optimization to optimize the RBF neural network.(2)For radial basis function(RBF) hidden layer center of neural network, width and the problem of connection weights are not easy to identify, we apply particle swarm optimization to research optimize parameter. Moreover, we analyzed the fundamental idea and algorithm flow of particle swarm optimization in depth. According to particle algorithm easy to fall into local optimal value, the slow convergence rate and low convergence precision problem, summarized the present situation of improved particle swarm optimization algorithm, focus on the analysis of the existing deficiencies of particle swarm optimization algorithm based on parameters, includes linear decreasing weight, adaptive weight, random weights and increases contraction factor method. Besides, we analyzed the main position mutation particle swarm optimization algorithm at present, therefore put forward a kind of position mutation particle swarm optimization.(3)As to the analysis of the problem before and after the effect of variation, respectively, use more than 5 kinds of particle swarm optimization to do fusion variation, on the basis of this, we build 10 kinds of RBF neural network temperature prediction model.(4)Then use various of model for temperature prediction, compares the predicted results variation before and after. Through one year’s temperature forecast of Beijing area, we acquired the forecasting data which showed the prediction model of variation after has a higher accuracy and a low degree of deviation. At the end of this paper, by comparing the results obtained, we concluded that fusion variant of the PSO optimized RBF neural network prediction model is more accurate.