Application Of Gaussian Process Mixture Model In Multi-modal Prediction Of Information Stream And Energy Stream

Stream is a phenomenon that is common in nature.Both information stream and energy stream are concrete manifestations of stream.Common information streams include network traffic,traffic flow,and the like.Common energy streams include solar radiation,tides,sunspots and the like.Network traffic prediction can ensure network fluency;solar radiation and ocean tides prediction are the basis of photovoltaic power generation and hydropower generation;sunspot prediction is the premise to ensure the safety of spacecraft and communication equipment.Therefore,it has important practical significance for information stream and energy stram prediction.The Gaussian Process Mixture(GPM)model is an intelligent model developed in the field of machine learning in recent years.It is composed of multiple Gaussian processes(GPs)through threshold functions,which can describe the multi-modal characteristics of the data stream in detail.In this paper,the Gaussian Process Mixture model is used to predict the multi-modality of information stream and energy stream data,which not only accurately analyze the characteristics of two data streams,but also provide reference for the change trend of subsequent data streams.The main research work is as follows:(1)Application of Gaussian Process Mixture model in information stream(network traffic)multimodal predictionBased on the multimodal characteristics of network traffic,this paper uses the GPM model for network traffic prediction.Firstly,the network traffic sequences in two different regions are characterized,and then they are reconstructed by normalization and phase space to generate a sample set and input into the GPM model.Finally,the hard classification iterative learning algorithm is used,the posterior probability is maximized,the optimal grouping of the sample set is realized by iteration,and the model parameter learning is completed by the likelihood function.The GPM is compared with support vector machine(SVM),kernel regression(KR),minimum and maximum probability machine regression(MPMR),and Gaussian process(GP).By comparing the root mean square error(RMSE)and the decision coefficient(R~2),the prediction accuracy of the GPM model is better than the other four models.The GPM model can be well applied to network traffic prediction and can provide reference for network administrators to allocate network resources.(2)Application of dual-kernels Gaussian Process Mixture model in energy stream(solar radiation and tidal power generation data)multimodal predictionBecause the GPM model uses a single kernel function,the prediction accuracy sometimes is not optimal.Therefore,this paper uses a variety of kernel function GPM model for multi-modal prediction of power generation data.First,the characteristics of two sets of power generation data are analyzed.Then the three kernel functions of square exponent(SE),rational quadratic(RQ)and Matern are combined to a new kernel function and used for prediction.Comparison of GPM model prediction results with seven different kernel functions,it was found that the dual-kernels GPM model combining SE and RQ is the best.Then,based on this kernel function,the GPM model is compared with the traditional learning model.By comparing the RMSE and R~2,the GPM model using the combined kernel function of SE and RQ is superior to all the traditional models.The dual-kernels Gaussian Process Mixture model can be well applied to power generation data prediction,which can provide reference and research for grid operation.(3)Application of Sparse Gaussian Process Mixture model in energy stream(sunspot)multimodal predictionTo further reduce the training time complexity,this paper uses the Sparse Gaussian Process Mixture(SGPM)model for multimodal prediction of sunspots.The core is to replace the original samples with a small number of pseudo input samples,thereby shortening the matrix inverse operation time and improving the training speed.After characterizing the sunspot sequence,the SGPM model is compared with the GPM model using Loo CV,Variational and hard classification iterations.By comparing RMSE,R~2 and training time(Time),the SGPM model and the GPM model using the hard classification iterative algorithm are basically consistent in prediction accuracy,but the SGPM model is obviously better than the other two GPM models.The training time is more than half of the GPM model using the hard classification iterative algorithm.It shows that the SGPM model can be applied to sunspot prediction well,which can provide learning and reference for outer space environment research.