Research On Time Series Short-Term Prediction Based On Multigranularity Event Detection

Time series prediction refers to observing the tendency of time series data,estimating the conditional probability distribution and parameters of the data,and constructing statistical relationships between the data,to realize the prediction of the future development of the data.Time series prediction can help users to formulate work plans,avoid potential risks,and provide evidence for decisions.However,the non-stationary characteristics of the time series make it hard to obtain accurate prediction results,especially for short-time prediction,and the prediction accuracy need to be further improved.An important task of time series analysis is to detect events that have special meaning to users.However,time series with inaccurate sampling interval are common in different fields.In this thesis,to obtain latent events in time series,we present a multigranularity event detection method based on self-adaptive segmenting.Moreover,we build a short-term prediction model based on multigranularity event.The specific research of this paper is as follows.(1)A multigranularity event detection method based on self-adaptive segmenting.First,considering the trend information of the time series,a self-adaptive segmenting algorithm based on edge points is proposed,the algorithm is used to divide the time series into some unequal segments.Then,the clustering algorithm is used to discover the trend pattern of segments,and the segments are mapped into symbol based on the clustering results.Thus,the symbolic representation of the time series is obtained.Finally,based on the symbolic representation of the time series,a tree structure is used to detect the multigranularity events in the symbol sequence.In the validation session,the bus data set and ten public data sets are used to verify the feasibility of the proposed method.(2)Short-term prediction method for time series based on multigranularity events.First,a latest time match strategy is proposed to match events on real-time data.Then,to solve the problem of the unequal length of the event instances,a fixed number piecewise aggregate approximation method is proposed,mapping each event instance into the same low-dimensional space.Finally,the event instances after dimension reduction are used as state vectors to construct the XGBoost-based short-time prediction model.In the validation session,custom passenger traffic datasets and stock datasets are used to verify the effectiveness of the proposed method.The results show that the proposed method can obtain better accuracy than other short-term prediction methods.Finally,in order to verify the feasibility of the proposed method,the method is applied on the customized passenger route recommendation platform to predict the passenger flow,and the proposed method is verified.