Research On Fusion Of Linear And Nonlinear Predictions For Multivariate Time Series

With the popularization of the Internet and the in-depth development of related information technologies such as data monitoring,how people obtain data have gradually become diversified,making the amount of data an unprecedented growth rate.Multivariate time series,also known as multivariate time series,usually describe time series data collected from different users or sensors in a stable network.In real life,fields such as transportation,finance,meteorology,and medical care are constantly generating massive amounts of time-series data every day.For example,the changing trend of the road occupancy rate in the traffic field over time,the daily electricity consumption of residents per hour in production and life,etc.At the same time,time series data is no longer limited to the characteristics of a single variable and small sample size.On the contrary,it is usually multivariate,large sample size,and more complex multivariate time series data.Multivariate time series forecasting predicts new trends or potential events by monitoring historical time series data to support future intelligent decision-making.However,it is difficult to learn the linear and nonlinear data features of time series at the same time,the periodic pattern is more complex,the time and space features are difficult to learn,and the interaction between different multivariate time series is difficult to learn.Aiming at the above problems,this paper designs a fusion forecasting method for multivariate time series.The main research contents and innovations of this paper are as follows:First,to make better use of the spatiotemporal characteristics of time series and the deep representation learning ability of deep learning to achieve efficient and accurate multivariate time series forecasting.In this paper,we propose linear and nonlinear forecasting modules for multivariate time series,respectively.Divide the historical time series data into long and short-term historical time series data as input,and go through a nonlinear prediction module based on neural network design.Among them,a feature extraction encoder is used to obtain spatiotemporal features in time series;interactive attention is used to maximize the interaction between spatiotemporal features from multivariate time series;finally,the fully connected layer outputs nonlinear prediction results.The short-term historical time series data is used as input,and the linear prediction result is generated through the linear prediction module based on the autoregressive model.Second,to make better use of the linear and nonlinear characteristics of time series,this paper proposes a linear and nonlinear fusion forecasting method for multivariate time series.The linear and nonlinear prediction results are used as input,and the prediction results of multivariate time series are obtained through the linear and nonlinear fusion prediction module.Extensive experiments and evaluations were conducted on four datasets,and the experimental results confirmed the effectiveness of the WFLNNet model.Different prediction algorithms,ideal prediction ranges,the number of convolutional neural network neurons,and different prediction model components were studied to predict the impact of the results.For example,on the Solar-Energy dataset of WFLNNet,as the ideal prediction range increases,the RRSE are 0.2186,0.2642,0.3113,and 0.3149,respectively;the CORR are0.9778,0.9671,0.9534,and 0.9508,respectively.