Research On Time Series Prediction Model Based On SAX And Echo State Network

Time series is a set of random variables that are sorted by time and widely exist in many fields of life,including meteorology,finance,agriculture,biology and other fields.Time series prediction predicts the time series data of the future period by analyzing the historical time series data.It is one of the hot research issues of time series data mining.In the time series prediction,the traditional time series prediction model is limited by its own structure,resulting in unsatisfactory results for complex nonlinear data.Neural network and support vector machine based prediction model training algorithm is complex,easy to fall into local optimum and the network structure is unstable,which obviously reduces the efficiency of building a model..As a new type of recursive neural network,echo state network is widely used in the prediction of nonlinear time series.However,due to the high dimensionality and complexity of time series,it is often difficult to analyze and predict the time series both efficiently and accurately.Meanwhile,because of the unique structural characteristics of traditional ESN network,a large number of training samples are needed for network training,making training difficult and affecting prediction accuracy.In view of this,in order to be able to accuratelypredict different time series data,this paper improved the SAX representation of timing,combined with the echo state network prediction model,and carried out a time series prediction model based on SAX and echo state network.The study is as follows:(1)Time series correlation algorithm analysis.The basic theory and related algorithms in time series mining are summarized and explained,and the related research objectives and algorithm principles of time series dimensionality reduction representation and prediction methods are introduced and analyzed.(2)Time series SAX model of local mean decomposition and improved wavelet entropy.In order to solve the problem that the traditional SAX notation has the same number of segments,equal treatment in the partitioning interval and the characteristics of the mutation information in the non-stationary sequence,the local mean decomposition technique and the improved wavelet entropy segmentation algorithm are introduced,A time series SAX model of local mean decomposition and improved wavelet entropy is proposed.The model combines a local mean decomposition technique with an improved wavelet entropy algorithm can not only denoise the original time series,but also can extract feature information of non-stationary sequences better and fit the sequence to reduce the dimension,improve the classification accuracy of the KNN classification algorithm on the SAX notation rate.(3)Time series prediction method of grey wolf optimization echo state network.Echo state network uses the reserve pool instead of the hidden layer ofthe traditional neural network,the core of which is to train some connection weights,the algorithm is simple.Due to the unique structural features of ESN,and a large number of training samples are required to make the training difficult.At the same time,it is also necessary to improve the adaptability of the reserve pool and predict the accuracy of complex prediction tasks.In order to solve the problem of training difficulties in ESN prediction,the grey wolf algorithm is introduced,and the time series prediction method of grey wolf optimized echo state network is proposed.It can adaptively and quickly achieve global optimization according to its predation behavior,through grey Wolf algorithm to optimize output weight,solve the problem of training difficulties,and improve the accuracy of ESN network prediction.(4)Time series prediction model based on SAX and echo state network.The SAX representation was introduced into the temporal prediction algorithm of the gray wolf optimized echo state network to construct a new time series prediction model.This simplifies the training data in the prediction model,and improves the construction efficiency of the model on the premise of maintaining the prediction accuracy.