Forecasting Research Of Subway Passenger Flow Based On Internet Search Index

With the rapid growth of urban population,the problem of road traffic congestion is be-coming more and more serious.Intelligent traffic system(ITS)is an indispensable part of the future smart city system.Rail transit has become an important part of its because of its green,convenient and large traffic volume.Major cities have vigorously developed rail transit sys-tem,which has become an important way for people to travel.The explosive growth of pas-sengers has brought great challenges to the effective management and daily operation of com-plex subway system.With the development of information technology,subway passenger flow(SPF)travel behavior is affected by the interweaving of multi-source information,show-ing the characteristics of volatility,randomness,heterogeneity,multi-source,mixing and so on.Accurate and reasonable passenger flow forecast is one of the key preconditions to deal with sudden passenger flow in time,formulate reasonable operation organization plan,bal-ance traffic demand and capacity supply,regulate daily production and transportation,im-prove railway service quality,increase operation income of railway department,and then im-prove urban service and management level.Recently,the technical framework based on a kind of data signal decomposition and then integration has attracted much attention for a single complex data source in various fields.Facing the nonlinear SPF data,the SPF is decomposed into multiple modal components with obvious single spectrum characteristics,and then the purpose of accurate prediction is realized by integrating the prediction results of each mode.For the prediction of multi-source data SPF,it is necessary to propose a multivariable feature fusion model.At present,the automatic fair collection of urban rail transit can automatically obtain a large amount of historical data in-formation of SPF.At the same time,the increasing popularity of the Internet has created a fa-vorable environment for us to study the search engine index.This information has brought unprecedented auxiliary opportunities to the SPF forecast and helped us to achieve a more robust SPF prediction.With the rapid development of Internet technology and artificial intelligence technology,big data mining,machine learning,deep learning and other methods show many advantages that traditional models do not have,and they are better at capturing the dynamic changes of complex data structures.To sum up,based on the feature analysis of multi-source data and information,and with the advantages of different intelligent models such as big data mining,machine learning,deep learning and statistical analysis,this paper puts forward a reasonable theoretical framework,and constructs a comprehensive and integrated subway short-term pas-senger flow demand prediction model driven by multi-source data and information through the optimal combination of various technical modules.This paper innovatively proposes a SPF point prediction and interval prediction model integrating Internet search engine index.The point prediction model framework includes the following aspects:(1)Multi-source data characteristics:the collection,dimensionality reduc-tion and statistical analysis of Baidu search index(BSI)of keywords related to SPF,so as to screen out Baidu keywords with strong causality.(2)Multi feature data fusion:Based on the multi-modal decomposition technology optimized by multi-objective algorithm,the SPF and related BSI are decomposed into a cluster of intrinsic mode function with single feature;Fur-thermore,in order to eliminate the statistical pseudo causality,the quadratic feature extraction of each cluster intrinsic mode function is carried out to form the optimal feature combination,which is used as the feature input of each cluster.(3)Model matching strategy:according to the structure and volatility of each optimal feature combination,a prediction model suitable for the distribution of model matching mechanism is established.(4)Establishment of com-prehensive integration framework:Based on the idea of comprehensive integration,consider-ing the advantages of each module and the balance of system error,an integration method is proposed and a comprehensive prediction model is established.Furthermore,for high complexity and irregular data,as a supplement to the insufficient accuracy of point prediction,we propose a new interval prediction method.By constructing the upper and lower bounds at a certain confidence level,the model can not only provide more uncertain information than point prediction,but also solve the error accumulation phe-nomenon caused by constructing interval prediction based on the estimated point prediction results.The method based on upper and lower bound estimation proposed in this paper inte-grates the technical advantages of multi-source data characteristics,multi-objective optimiza-tion and deep learning.The prediction framework mainly integrates the following modules:(1)According to the market share of China’s three major search engines in 2018,we first weighted Baidu Index,Sogou index and 360 index to obtain the search engine index(SEI).In order to obtain the auxiliary SEI as a strong predictor,SEI with causal relationship with SPF was screened by statistical test.(2)In order to effectively extract the auxiliary information hidden in the screened predictors,accurate fusion matching strategy is particularly important.Firstly,the multi-modal decomposition technique of multi-objective optimization decomposes the multi-channel input SPF and SEI into an intrinsic mode function matrix with obvious sin-gle characteristics.The modal functions with similar periods and frequencies in the matrix are regarded as an information cluster,so as to form a multi-cluster structure.Then,the intrinsic mode function within each cluster is extracted twice to eliminate the pseudo causality caused by statistical test,and the intrinsic mode function of the optimal input combination of multiple clusters is obtained.Finally,the optimal combination is input to the multi-objective optimiza-tion deep learning model with appropriate structure,and the upper and lower bounds of each cluster are directly predicted and output.(3)The decomposition technology reduces the com-plexity of time series,integrates the matching strategy to improve the prediction accuracy of the model,and the integration technology strengthens the generalization ability of the model plate.Based on this,the upper and lower bounds of each cluster prediction are superimposed to obtain the final interval prediction result.Finally,Beijing,Shanghai and Guangzhou SPF data are used to verify the achievements of the latest technical level of the proposed model.The experimental results of our proposed model and the designed benchmark models show that both point prediction and interval pre-diction have achieved accurate prediction results.Taking Beijing as an example,the one-step point prediction error MAE=5.3117,MAPE=0.7319%,RMSE=79.7419,R~2=0.9991;In-terval prediction error PICP=0.9041,PINDE=0.0019,PINAW=0.0768,CWC=0.0768,W_s=-20.5061.Therefore,the SPF prediction method combining multi-source Internet search engine index and multimodal analysis proposed in this paper can provide excellent prediction performance and help to improve rail transit operation management and operation efficiency.The proposed technical framework also provides a technical reference for dealing with multi-source information and multivariable fusion.