Research And Application Of Time Series Prediction Algorithm Based On Gaussian Process Regression

Time series analysis has been widely used in engineering,meteorology,economics,finance and other fields,and its research on modeling and prediction methods has been a hot topic in various fields.The linear regression model represented by the autoregressive model has been the main method for time series prediction in the past decades.In recent years,the nonlinear regression method represented by Gaussian Process Regression(GPR)has received widespread attention,but at the same time,the time series prediction based on GPR There are still some key problems to be solved in the method:such as the selection of kernel functions,automatic selection of autoregressive orders,and overfitting of the model.This thesis focuses on the above-mentioned related issues of GPR timing modeling.The research in this thesis was supported by the Natural Science Foundation of Zhejiang Province.The main research contents and results are summarized as follows:(1)A method for automatic selection of autoregressive terms based on the maximum information coefficient(MIC)is proposed.In the traditional method,the autoregressive order is mainly obtained by filtering through the autocorrelation function and partial autocorrelation function method,or by the wrapped feature selection method.The former can only evaluate the linear relationship,and it is not suitable for non-linear models such as Gaussian process regression Although the latter can achieve good results,it requires multiple tests and verifications,which is inefficient.This thesis proposes a MIC-based autoregressive item selection method.This method can not only measure complex nonlinear correlations,but also quickly complete filtering.Test verification results also show that the proposed method can efficiently complete filtering and screening,and can help Gaussian process regression models to achieve higher prediction accuracy.(2)An integrated Gaussian process regression model based on guided aggregation algorithm is proposed.The proposed integrated model can alleviate the problem that the MIC-based feature selection method easily converges to the local optimal solution,on the other hand,it can introduce random factors to reduce the risk of overfitting.In this thesis,a Gaussian process regression model is used as the base model and a column sampling method is introduced.After filtering the regression term for each sampled sample subset,it is used to train the base model and obtain multiple sets of prediction results.At the same time,the Gaussian process regression model can The predicted posterior variance can be used as a natural model evaluation criterion,so this thesis weights the prediction results of each model based on the posterior variance of each base model,which can enrich the search space,reduce the risk of overfitting,and improve the prediction effect of the model.(3)A data analysis platform was constructed to implement data storage,management,and visualization,as well as the embedding of time series analysis algorithms.Firstly,the communication channel between the software platform and the OneNET cloud was built to realize data transmission and protection.Secondly,a local database was developed for data storage and analysis.By inductively storing the data sources,a geographic location visualization of the embedded Baidu map was established.Interface and data query interface;finally,for the time series data stored in the local database,the prediction algorithm proposed in this thesis is used to analyze and predict,and the prediction results are visualized in the form of charts.