Research Of Hybrid Model Based On Relevant Vector Machine In Time Series Forecasting

Time series is a sequence of values recorded in chronological order.With the rapid development of society and the coming of Big Data era,a large number of time se-ries data are generated in various fields,such as economy,meteorology,transportation,medicine and energy.Time series prediction refers to model the historical data using prediction algorithms,and employ the model to analogize and estimate the future de-velopment trends.Time series forecasting can be applied in all-around fields,which have significant impact on the integral development of the economy,the harmony and stability of society,and the safety of citizens.The time series has nonlinearity and periodicity generally.Moreover,the single prediction algorithm also has some limitation.Consequently,these factors make the time series prediction hard.Considering that,researchers have proposed various methods to improve the accuracy of prediction.Data pre-processing technology,for example,would be employed to remove noise.It is a good method to adjust the input variable and dimensionality of model with the characteristic of time series.Optimization algorithms are used to determine the optimal parameter value of model.Based on above problems,this thesis has proposed two hybrid model,which depend on all the virtues of excellent learning ability of RVM,date preprocessing technology and optimization algorithm.(1)A combined prediction model WSFRVM is proposed based on Relevant Vector Machine.The proposed model first utilizes WT to decompose the original electrical load series into two components,and then uses SSA to remove the noise of detail sub?series.RVM shows potential application in nonlinear time series prediction,but its performance is seriously affected by parameters.Consequently,FOA is used to auto-matically determine the optimal parameter value for RVM.To verify the effectiveness of the proposed hybrid model,the proposed model is compared with six benchmark mod-els on three open datasets using the single-step-ahead and multi-step-ahead forecasting strategies.Furthermore,the evaluation module,which includes the DM test and eval-uation criteria,is introduced to assess the comprehensive performance of the proposed hybrid model.The experimental results demonstrate that the proposed model has higher forecasting accuracy and better practicability than benchmark models.(2)Combined prediction model SMFRVM is proposed based on multivariate aux-iliary prediction.Since the time series data is affected by the vertical information and the horizontal information,the model considering auxiliary variables comprehensively can improve the prediction accuracy by correcting the output of the prediction variables.Firstly,the correlation coefficient between the relevant variables and the predictor vari-ables is calculated.Based on the correlation coefficient value and the complexity of the model,the appropriate number of auxiliary variables are selected.Then the sin-gular spectrum analysis is employed to extract the principal features in each variable.To overcome the difficulty of determining the RVM kernel parameters,we utilize the Fruit Fly Optimization Algorithm to optimize the selection of appropriate parameters,and then use this model for each variable prediction.Finally,the results predicted by each variable are calculated and output using a linear regression model optimized by Genetic Algorithm.The model is experimentally verified on two real data sets.The experimental results show that the SMFRVM model has better prediction accuracy than the benchmark model.