Research On Time Series Prediction Based On BP Neural Network

The analysis of the collected historical observation data is conducive to the discovery of new laws and new knowledge.The time series is a set of observed data arranged by time.People use the results of the time series to predict the development and change of the object for a period of time in the future,so as to make better decisions.Time series forecasting is widely used,such as in the financial field,meteorological field,business,and other fields.The time series has the characteristic of large amount of data,noise interference,non-linear and rapid updating.The previous time series research mainly relies on the traditional statistical method,but the complexity of the time series data makes these methods cannot meet the requirements gradually.Artificial neural networks have the advantages of selforganization and non-linear ability in solving complex nonlinear problems,and are more effective in time series prediction.In this paper,we introduce several prediction models that are mainly used in time series prediction,and analyze the advantages and disadvantages of current methods and models for nonlinear system prediction.And then introduces the artificial neural network,especially the advantages of BP neural networks in nonlinear system prediction.Some shortcomings of the BP neural network prediction model are summarized,mainly due to the improper selection of the initial weights,which makes the neural network easy to fall into the local minimum.In this paper,a BP neural network model optimized by genetic algorithm is proposed to improve the prediction performance of neural network.With the development of information technology such as Internet,mobile communication and Internet of things,the data show the trend of exponential growth.The convergence rate of the neural network is very slow when the network has more hidden layer nodes and large data samples.In order to solve the above problems,based on the open source distributed cloud platform Hadoop,this paper proposes a parallel prediction model based on Map Reduce.The distributed parallel processing is carried out at different stages of the prediction model,which improves the computational efficiency.The main work of this paper is as follows:1.The network structure selection of neural network lacks effective theoretical guidance,and the initial value of network has great influence on the prediction qualityof neural network.Improper value may cause the network to fall into the local optimal solution.In this paper,the initial weights of the neural network are trained according to the characteristics of the global optimization of the genetic algorithm.The training results are taken as the initial values of the neural network to improve the prediction quality of the BP network.2.Genetic algorithm optimization neural network algorithm in the sample set is very large,the training is very slow,or even unable to converge.Therefore,this paper designs a neural network parallel method based on MapReduce genetic algorithm optimization.The parallelization of the optimization method is divided into two stages:the parallelization of the genetic algorithm and the parallelization of the BP neural network.Parallelization of genetic algorithms using multiple population parallelization methods,multiple populations are assigned to different node operations.Select the highest degree of fitness in all nodes,the weight obtained at this time as the initial weight of the BP neural network.Based on Map Reduce BP network parallelization stage,the training data set is distributed in each node to achieve parallelization,the reduction phase of BP network to calculate the cumulative error,batch adjustment network connection weight threshold,multiple iterations to complete the model processing.The results show that the parallel method based on MapReduce is obviously shortened compared with single node in the large scale of data sample,and good acceleration effect is achieved.