Multivariate Time Series Missing Data Imputation Based On TSnet

Multivariate time series are widely present in daily life and scientific study,such as air quality data,stock data,climate change record and data used in other studies.However,due to sensor failure,data transfer failure and damage in storage units,these time series may contain plenty of missing data.Therefore,how to accurately impute such missing data is a practical issue.Existing methods do not take into account the different effects of different features on missing values imputation,and do not effectively fuse temporal and spatial information.In order to solve such problems,a TSnet(Temporal-Spatial net)deep learning model is proposed to improve the accuracy of missing value filling,which combines temporal information and spatial feature information.The TSnet model consists of two parts,the first part is the feature attention CNN neural network,and the second part is the temporal information extraction network.The main contributions and work of this thesis are as follows:(1)We propose a feature-attention CNN network guided by temporal information,which can help the CNN neural network to learn spatial features better according to the different weights of information provided by different features.(2)We propose an improved LSTM-based temporal information extraction module Imputation-LSTM and an improved BERT-based temporal information extraction module I-BERT.Both modules can learn temporal information from multivariate time series data with missing values.(3)We construct the TSnet-L model based on Imputation-LSTM and the TSnet-T model based on I-BERT.This thesis conducts experiments on three datasets from different fields,and compares the experimental results of TSnet with other commonly used models.The results show that TSnet-L performs better than other models on the missing value imputation task of multivariate time series.TSnet-T has an advantage in training speed.At the same time,the results of ablation experiments also show that after adding the feature attention module,the prediction accuracy of the model for missing values is improved.