The Study Of Neural Network Time Series Prediction Based On Particle Swarm Optimization

With the society stepping into the information age,data with various types are generated in large numbers.And one of data obtained according to the time are called time series,such as,the price of stock,the observation of the sunspot,power load,meteorological change,number of website visits and so on.The most typical feature of time series is the interdependence between data.Finding the internal laws of time series from data and used to predict the future value is called time series prediction.The traditional time series forecasting method is to establish the forecasting model after using the pure mathematic tools to analyze the data.However,it is difficult to establish the ideal forecasting model for the system because that the time series has nonlinear characteristics.Artificial neural network has the characteristics of self-organization,self-learning and strong fault-tolerance,meanwhile it has a good approximation ability to deal with nonlinear problems.So,it can gain a forecasting model with good performance by using the recurrent neural network(RNN)with dynamic characteristics to predict time series.However,there are some defects for RNN,such as,local optimum,calculation load and gradient disappeared and so on.To overcome the defects of RNN's learning algorithm,particle swarm optimization(PSO)is used to optimize the RNN in this paper.Furthermore,a recurrent neural network based on PSO is proposed to solve time series forecasting problem in this paper,and the main contributions of this dissertation can be summarized as follows:1.There are some shortcomings for PSO,such as,falling in the local minimal and algorithm's diversity is poor etc.To get over these shortcomings,a adaptive chaos particle swarm optimization(ACPSO)is proposed in this paper.In the proposed algorithm,the learning parameters are dynamically adjusted according to individual's fitness value.Meanwhile,chaos interference factor and variation factor are introduced to increase the diversity of the algorithm.The exploration and mining capacity of algorithm are balanced by these improvements.The weights and biases of RNN are optimized by using the ACPSO,and the optimized network is used to predict the oil price series.The forecast results demonstrated that the optimized neural network can obtain the prediction result with higher accuracy at a small time cost,and the network has a good generalization ability.2.To solve the complex time series prediction problem with high dimension,a dynamic multi-group particle swarm optimization(DMPSO)is proposed in this paper.In the proposed algorithm,the main swarm and assistant swarm are used to search the global optimum,and the main swarm and assistant swarm are updated by the ACPSO and quantum particle swarm optimization respectively.What's more,the DMPSO are used to optimize RNN,and the optimized RNN is applied to predict the sunspot dataset.The prediction results indicated that the optimized neural network has higher prediction precision when it is used to deal with the complex time series prediction problem,and the time cost is small.3.In order to verify the effectiveness of PSO in RNN prediction of time series,time series datasets of different domains are selected to be as test sample.These prediction errors are produced when RNNs with different learning algorithm are used to predict time series.The prediction errors shown that the RNN based on PSO algorithm has a good prediction performance in solving time series forecasting problem.