Multivariate Time Series Forecasting Technologies And Applications

With the wide adoption of computing technology,collecting and storing time series data has become more convenient.The analysis of multivariate time series has found extensive applications in industrial production as well as in daily life,including stock forecasting,production line monitoring,digital marketing,and auxiliary medical care.However,the analysis of multivariate time series encounters various challenges such as:(1)Long-term dependencies.Multivariate time series data often display long-term dependencies that last for several months.However,current forecasting algorithms for multivariate time series lack the ability to model long-term dependencies explicitly.(2)Temporal dependencies and inter-variable dependencies.Statistical-based search methods for time series neglect complex inter-variable and time series dependencies in multivariate time series,yielding limited prediction accuracy.(3)Inconsistency between the distributions of training and inference data When applying multivariate time series forecasting to tasks of unsupervised anomaly detection,it is assumed that the data in the training set are all normal,and the forecasting model can learn the normal pattern of multivariate time series and detect anomalies through prediction errors.However,when confronted with abnormal data,this may result in unstable detection of anomalies.This thesis proposes two prediction methods to improve the accuracy of multivariate time series prediction.These prediction methods are based on statistical featurebased search and pre-trained representation-based search,which fully model long-term dependencies,inter-variable dependencies,and temporal dependencies.Moreover,this thesis proposes an unsupervised prediction-based method for detecting anomalies in multivariate time series using a dual masked mechanism.This mechanism reduces the inconsistency of training and inference data distribution,leading to a more robust detection result.The main contributions of this thesis is summarized as follows:· Statistical Feature-based Search for Multivariate Time Series Forecasting:The time series has long-term dependencies,such as trends,seasonality,and periodicity.The dependencies may span several months.Directly applying ex-isting methods is insufficient in modeling the long-term dependencies of the series explicitly.To address this issue,this thesis proposes a Statistical Featurebased Search for multivariate time series Forecasting(SFSF).Initially,statistical features which include smoothing,variance,and interval standardization are extracted from multivariate time series to enhance the perception of the time series’ trends and periodicity.Combine statistical features to search for similar sequences in historical sequences.The current and historical sequence information is then blended using attention mechanisms to produce accurate prediction results.Experimental results show that the SFSF method outperforms six state-of-the-art methods.· Pre-trained Representation-based Search for Multivariate Time Series Forecasting: The statistical feature-based search for multivariate time series forecasting method neglects modeling of dependencies between multivariate variables and temporal dependencies.This thesis proposes a Pre-trained Representationbased Search for multivariate time series Forecasting(PRSF).Initially,a contrastive learning-based method is introduced to acquire multigranular representations for multivariate time series.Subsequently,the acquired representations are used to identify similar sequences in the historical data,which are then combined with the present data to predict future sequences.Experimental results on four real datasets demonstrate the effectiveness of the proposed PRSF method.· Forecast-based Unsupervised Multivariate Time Series Anomaly Detection Method: Forecast-based unsupervised multivariate time-series anomaly detection methods are prominent.By training a forecast model using normal time series,the objective is to capture the normal patterns of a multivariate time series.It is assumed that the predicted sequence generated by the model is normal.If the prediction error exceeds a threshold,the predicted sequence would be considered abnormal.Inconsistency between the distributions of training and inference data make it difficult to accurately detect anomalies.The detection becomes challenging if the input window itself contains anomalies,leading to unstable results from forecast models.Moreover,the correlation between multivariate vari-ables can be complex,which can make noise and compromise the effectiveness of anomaly detection when attempting to directly fuse sequence information from uncorrelated variables.This thesis proposes a forecast-based anomaly detection method for multivariate time series called Dual Masked Self-Attention(DUMA)method.The method includes a block masked mechanism to improve the forecast model’s robustness.To address the noise interference issue between uncorrelated variables,this thesis presents the Max Mask Self-Attention algorithm.Experimental results on three real datasets corroborate the effectiveness of DUMA.