Time Series Prediction Based On Extreme Learning Machine

Time series contains a lot of valuable information.Time series prediction refers to the estimation of future development trends by analyzing known data,and accurately predicting time series can effectively guide people's production and life.Extreme Learning Machine(ELM)is widely used as a simple and fast learning algorithm for time series prediction.Compared with the traditional feedforward neural network,the extreme learning machine only needs to randomly set the input weight and the hidden layer threshold.The output weight can be obtained by solving a linear equation group.The training process does not need to be iterated and complets in one step.It can improve the generalization and learning speed of the network.This thesis is based on the extreme learning machine to predict the time series,the main research is as follows:(1)For the basic extreme learning machine,the values of the old and new training samples should be equal,and the weight of the new training samples cannot be obtained by the old training samples.This thesis proposes an improved sample learning of the extreme learning machine for the extreme learning machine.The mechanism improves the regression prediction ability of the extreme learning machine.(2)The Rajda standard is explained that eliminates unreasonable data in the sample.The cubic spline interpolation method that supplements the missing sample data,and the method of sample normalization are introduced.The time series contains different frequency components,so it is necessary to decompose the time series into different frequency components.By empirical modal decomposition and wavelet decomposition simulation of a typical time series,the two decomposition methods are compared.It is found that the empirical mode decomposition method has its unique advantages in dealing with nonlinear and non-stationary time series,and finally determines the empirical mode.State decomposition is the decomposition method of the prediction method in this thesis.(3)For the nonlinear and non-stationary characteristics of time series,this thesis proposes a prediction method combining with improved empirical mode decomposition and extreme learning machine.The realization process is as follows: the time series is decomposed into several components of different frequencies by the empirical mode decomposition method,thereby reducing the non-stationarity of the series,and the components obtained after the decomposition exhibit a short correlation,and the series of the short correlation is complicated.Low degree is more conducive to the establishment of predictive models.For each component after decomposition,the prediction is performed by an improved extreme learning machine,and the prediction results of the respective components are obtained,and the prediction results of the models are obtained by superimposing the prediction results.(4)Using two different types of time series for simulation prediction,one is Lorenz chaotic time series,one is the actual measured network traffic time series,Compared with other time series prediction methods,the simulation results show that the prediction method is more accurate than others.