Font Size: a A A

Gaussian Processes Mixture With Its Application On Time Series Prediction

Posted on:2018-12-12Degree:MasterType:Thesis
Country:ChinaCandidate:H L ChangFull Text:PDF
Abstract/Summary:PDF Full Text Request
Gaussian Process Mixture(GPM)model has been paid much attention to machine learning in recent years.GPM is composed by multiple Gaussian process(GP)models which based on the set threshold function combination.GPM model is superior to GP model in two aspects: First,it can describe the multimodal data precisely and the other is to shorten the model parameter estimation time.This paper focuses on the time series multimodal prediction by Gaussian Process Mixture and its improved model.The main work is as follows:(1)Multimodal prediction of short high-speed traffic flow based on GPM modelTraffic flow sequence has several characteristics such as dynamic,temporal similarity and nonstationary randomness.It is the core guarantee for the precise operation of intelligent traffic management system(ITS).In the past,the prediction model has not improved the prediction accuracy by using the time similarity,that is,the multimode feature as the entry point.In view of the above problem,this paper proposes that applying GPM model to traffic flow forecasting.The model possesses divide and conquer strategy,it could divide samples into different groups according to the multimodal characteristics,and then each group is trained by GP model alone.The results of experiments show that GPM model not only obtain the training parameters by fine fitting learning samples but also improve the accuracy of multimode prediction by reducing the matrix inverse calculation time.The model training algorithm adopts the maximum expectation algorithm(EM),we use it to compare with other two mainstream learning algorithms Variational and LooCV of GPM model.It is found that the EM learning algorithm could get better results in parameter training time,prediction accuracy and anti-noise performance.(2)Multimodal prediction of electric load and traffic flow based on Sparse Gaussian Process Mixture model.Although GPM model can reduce the training time by cut down the matrix dimensions by divide and conquer strategy,the training time is still long for the time series which have too large samples.Aiming at this problem,Sparse Gaussian Process Mixture(Sparse-GPM)model is proposed for electric load and traffic flow prediction.The key improvement is that after grouping the learning samples by the set threshold function,training each group independently by fewer pseudo-input samples which are approaching the original input samples to further reduce the matrix inversion time.Experiments show that Sparse-GPM has a better result for multimodal prediction.In addition,comparing the Sparse-GPM three algorithms including EM,Loocv,and Variatonal algorithms from training time,prediction accuracy and anti-noise performance,the results show that the EM algorithm of Sparse-GPM can achieve better results.(3)Prediction for remaining useful life of lithium-ion battery based on GPM modelRemaining useful life(RUL)prediction of lithium-ion battery is closely related to its capacity degradation trajectory.Since the capacity regenerative phenomenon occurs the self-charging and discharging processes,the capacity degradation trajectory exhibits monotonous and multi-peak characteristics.Applying GPM model predict the monotonic multimodal time series.First,the learning samples are grouped according to the multi-peak characteristic,and then each group is trained individually by the GP model,whereby the multimodal state can be more accurately fitted.The model is applied on the commercial Type 1850 Lithium-ion battery from the Prognostics Center of Excellence(PCoE),NASA.Comparing GPM model with GP model and SVM on the accuracy of the prediction capacity and RUL,The results show that the GPM model has higher computational accuracy than the GP model;the prediction accuracy is comparable to that of the SVM,however,the GPM can output the prediction confidence interval,so it can provide more prediction information.
Keywords/Search Tags:Gaussian Process Mixture Model, Time Series Prediction, Short High-speed Traffic Flow Prediction, Short-time Electric Load Prediction, Lithium-ion Battery Remaining Useful Life Prediction
PDF Full Text Request
Related items