Research On Multivariate Time Series Prediction With Data Preprocessing

As a means of predicting the future based on historical information,time series prediction provides the possibility of reference for decision-making and adoption.Therefore,it has become a research hotspot in transportation,energy,environment,finance,logistics and other industries.With the advent of the Internet of Things and the era of big data in recent years,the collection and acquisition of time series has become more and more convenient,but at the same time,it has exposed the problems of uneven data quality and increasing volume of data,which brings the following challenges to time series prediction: one is that abnormal data generated by random factors tends to destroy the regular pattern of the time series;the other is that the noise component that may be generated during data collection and transportation reduces the proportion of useful information in the time series.Third,the progress of data acquisition hardware not only makes the total length of time series larger,but also makes the data face the trend of dimension expansion.How to detect the anomalies in the time series,eliminate the noise components in them,and fully extract the feature information from the complicated multivariate time series data is the key issue in the study of time series prediction.To solve these problems,this paper aims to establish a accurate multivariate time series prediction method with data preprocessing including time series anomaly detection and noise removal.The specific work content and research results are as follows:(1)Aiming at the problem of time series anomaly detection,an unsupervised anomaly detection method(FE-EBi LSTM-AE)based on feature enhancement and ensemble bidirectional long-and short-term memory autoencoder is proposed.This method enriches the original data features through feature engineering,and then trains EBi LSTM-AE to obtain a time series anomaly detection model.Finally,experiments on multiple public datasets show the effectiveness of the anomaly detection method.(2)Aiming at the problem of time series noise removal,a time series noise reduction method(PAVMD-WPE-WTD)with parameter adaptive variational mode decomposition and wavelet threshold denoising is proposed.This method adaptively determines the number of modes for VMD according to the characteristics of the time series,and introduces wavelet threshold denoising algorithm with heuristic threshold and adaptive threshold function to remove noise.Finally,experiments on simulated artificial time series prove the effectiveness of the method.(3)Aiming at the problem of multivariate time series prediction,a prediction model(SRGRUs-MLAttn)based on stacked residual gated recurrent units and multi-level attention is proposed,and the proposed anomaly detection and noise reduction algorithms are combined.The spatiotemporal features of time series are extracted by stacked residual GRUs,and a multilevel attention mechanism is proposed to make full use of the extracted features.Finally,prediction experiments on five datasets covering the four fields of environment,transportation,energy and cloud computing prove the effectiveness of the proposed anomaly detection,noise reduction and prediction methods.