Time Series Forecasting Based On Causal Network Learning Method

Time series widely exists in all fields of life,especially in recent decades,there has been an explosive growth of time series data.At the same time,time series forecasting research based on time series data has also received more and more attention.This paper takes multivariate time series as the research object,explores the relationship between variables through the causal analysis method,and selects the most relevant feature set of the target variable,so as to complete the analysis and prediction of the complex system.The research contents of this paper are as follows:Aiming at the problem of the interference of irrelevant variables and redundant variables in the actual time series,this paper proposes a two-stage causal network learning method based on feature selection and causal analysis to explore the causal relationship between variables and realize the time series prediction of complex systems.In the first stage,the method adopts a feature selection strategy based on global redundancy minimization(GRM)to remove irrelevant and redundant variables in the original time series.The GRM strategy can consider the correlation and redundancy between variables from a global perspective,and can provide suitable feature sets for subsequent models.In the second stage,the method uses the momentary conditional independence(MCI)method to calculate the causal relationship between variables,and accurately determines the relevant feature set of the target variable.Based on the results of the proposed two-stage causal network learning method,an accurate prediction model can be constructed.Finally,the causal analysis ability of the proposed method is verified based on standard nonlinear data set,and the prediction performance is verified on the real-world Ei nino southern oscillation(ENSO)time series and Beijing air quality index(AQI)and meteorological time series.Aiming at the problem that the general Granger causality analysis model does not consider the dynamic behavior information between variables and is only suitable for linear and bivariate causal analysis,this paper proposes a two-stage causal network learning method based on the Peter-Clark(PC)algorithm and local Granger causality(LGC).In the first stage,the method uses the PC algorithm to determine the set of relevant features for each variable in the system.The PC algorithm can effectively deal with the correlation between variables in a complex system through an iterative search strategy.In the second stage,the method improves the condition set of the LGC model based on the results of the PC algorithm,so that it can handle the causal analysis problem among nonlinear,high-dimensional system variables.In addition,this method can not only obtain the quantified causal effect size between variables based on the modified LGC,but also further obtain the dynamic behavior information between variables.Finally,the validity of the model is verified by one standard linear data set and two real-world data sets,ENSO time series and Shanghai AQI and meteorological time series.