Multivariate Time Series Forecasting Based On Causality Analysis

In real life,multivariate time series data exists in various fields and plays an important role in analysis and prediction in various industries.These research areas involve weather forecasting to help in agricultural farming and aviation window forecasting,in finance to help in stock analysis and macro policy making,etc.In the past,many multivariate time series forecasting algorithms have been proposed.The initial methods applied to multivariate time series forecasting were based on statistics,and then with the rapid development of deep learning,many deep learning algorithms have been applied to this field.In the application of many deep learning algorithms in the field of multivariate time series forecasting,such as RNN,LSTM,GRU,and CNN,although good results have been achieved,these methods do not take into account the variables when making predictions.The relationship within the sequence and the relationship between variables directly treat all variables as input variables.Therefore,this paper introduces the concept of causal analysis and combines it with deep learning methods to perform prediction tasks.Meanwhile,causality can analyze the flow of information between different features,the size of the relationship between the already existing sequence and the predicted sequence,etc.The use of causality can also better constrain the information extraction in neural networks.Based on the above facts,this paper first studies the causal relationship between variables,and adds causal relationship analysis to the task of neural network prediction of time series.Then two different models were proposed for time series forecasting tasks,and both achieved good results.In this paper,we first propose a multivariate time series forecasting model based on causal analysis and a combination of multiple networks.This model uses the Lasso-Granger causality analysis method and simple convolution operation as the base structure of the model,which reduces the dimensionality of the input variables.And the sequences are encoded with two branches of bidirectional RNN branch,convolutional neural network and self-attention branch.Finally,LSTM is used for decoding and output.Taking the first experiment a step further,starting from the field of graph neural networks,this paper proposes a multivariate time series forecasting model using transfer entropy and graph neural networks.This chapter analyzes the shortcomings of Granger model which is difficult to deal with nonlinear problems,and proposes a solution to solve it by transfer entropy.This model uses transfer entropy to extract causal information and constructs multivariate time series into a directed graph structure;the original feature input is aggregated and feature extracted in the form of multiple convolution;and the final output and prediction of the sequence is done in the form of graph neural network.To summarize the full paper,this paper proposes two models for time series forecasting,and embeds causal analysis methods into our models from different perspectives.In the first experiment,the Lasso-Granger causal analysis method is used to analyze the input data After the causal relationship analysis,the variable dimension is reduced,and some variables that are not helpful for predicting the result are screened out;in the second experiment,the transfer entropy is used to calculate the causal relationship matrix between each feature node,forming a directed graph,build a time series encoder and decoder with a graph attention network as the core for time series prediction.This paper makes predictions on two data sets,and compares basic models such as VAR,CNN-AR,and RNN-GRU.The experimental results show that the prediction method based on causal analysis is effective and achieves better performance.