Research On Deep Learning Based Multi-Variate Time Series Forecasting Algorithms And Application

Time series forecasting(TSF)is a typical time series analysis task,which is of great significance for assisting decision-making,resource allocation,and loss prevention in advance.It is widely used in the fields including power,weather,transportation,and business.TSF algorithms based on deep learning learn sequence characteristics in a data-driven manner and model the mapping between historical data and future data.In the recent years,TSF has become a hot research topic,and many research focused on univariate time series prediction,multivariate time series prediction,and time series multi-step prediction,etc.While modern real-time information systems generate large amounts of time series data for the purpose of system performance monitoring.Making prediction of future system performance provides helpful information for the daily operation and maintenance of modern information systems.With the widespread deployment of information systems,TSF algorithms increasingly used to predict their performance and make intelligent operation decisions.This paper focuses on the following two problems:(1)the local variable prediction accuracy problem of multivariate prediction algorithms,that is,the multivariate prediction algorithm cannot guarantee the prediction accuracy of local variables while optimizing the overall prediction accuracy;(2)system performance prediction problem of database management system(DBMS),including trend prediction of performance scores and mining key performance indicators with top-K influence(top-K KPI mining).To solve the local variable prediction accuracy problem of multivariate prediction algorithms,we design and implement SEPNets,a multivariate time series predictionalgorithm based on self-evolutionary pre-training.Inspired by the pre-trained model,SEPNets first constructs and trains a univariate time series model as a basis for downstream modeling;then it captures the complex temporal dependence between the timeseries by using a convolutional network(1-dimensional convolution)and a long and short memory(LSTM)unit model;using the pre-trained model for fusion and retraining,SEPNets solves the problem of local variable prediction accuracy for multivariate time series prediction.The experimental results show that the SEPNets algorithm ensures the prediction accuracy of local variables while obtaining the relatively highest overall prediction accuracy.To solve the system performance prediction and KPI mining,we design a system performance prediction algorithm called MTSL based on multi-task neural network.Starting from task relevance,MTSL uses wavelet analysis to capture multiresolution information of performance score time series,and combines with multiple LSTM networks and stacked long and short memory networks(Stacked LSTM)to capture complex time series information of high-dimensional KPI time series.By carrying out multi-task training through sharing the hidden layer vectors of two task networks,MTSL can realize multi-step prediction of system performance score sequence in parallel,and mine key performance indicators with top K influence.The experimental results show that the proposed MTSL algorithm achieves good accuracy in system performance score trend prediction and top-K KPI mining.