CN110991713B - Irregular area flow prediction method based on multi-graph convolution sum GRU - Google Patents

Irregular area flow prediction method based on multi-graph convolution sum GRU Download PDF

Info

Publication number
CN110991713B
CN110991713B CN201911148344.6A CN201911148344A CN110991713B CN 110991713 B CN110991713 B CN 110991713B CN 201911148344 A CN201911148344 A CN 201911148344A CN 110991713 B CN110991713 B CN 110991713B
Authority
CN
China
Prior art keywords
regions
area
flow
time
correlation
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201911148344.6A
Other languages
Chinese (zh)
Other versions
CN110991713A (en
Inventor
史晓颖
僧德文
吕凡顺
徐海涛
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Dianzi University
Original Assignee
Hangzhou Dianzi University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hangzhou Dianzi University filed Critical Hangzhou Dianzi University
Priority to CN201911148344.6A priority Critical patent/CN110991713B/en
Publication of CN110991713A publication Critical patent/CN110991713A/en
Application granted granted Critical
Publication of CN110991713B publication Critical patent/CN110991713B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
    • G06F16/2465—Query processing support for facilitating data mining operations in structured databases
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/29—Geographical information databases
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Tourism & Hospitality (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • Mathematical Physics (AREA)
  • Development Economics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Business, Economics & Management (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Marketing (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Molecular Biology (AREA)
  • Game Theory and Decision Science (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Computing Systems (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biophysics (AREA)
  • Educational Administration (AREA)
  • Primary Health Care (AREA)
  • Biomedical Technology (AREA)
  • Fuzzy Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Remote Sensing (AREA)

Abstract

The invention discloses an irregular area flow prediction method based on multi-graph convolution and GRU, which comprises the following steps: dividing a region into N unconnected irregular regions; step two, performing space-time simplification on the historical track data, and calculating to obtain the inflow and outflow of all the regions at each time step; establishing a plurality of correlation diagrams among the regions, constructing a corresponding adjacent matrix, and expressing diversified spatial correlation among irregular regions; designing a multi-graph convolution neural network based on the correlation diagram among the regions, fusing diversified spatial correlation characteristics among the regions, and obtaining a multi-graph convolution fusion result; step five, based on the multi-graph convolution fusion result, capturing time correlation by adopting a GRU neural network; and sixthly, selecting a proper loss function, training to obtain a prediction model, and predicting through the prediction model to obtain the inflow and outflow of each region.

Description

Irregular area flow prediction method based on multi-graph convolution sum GRU
Technical Field
The invention relates to the field of traffic flow prediction, in particular to an irregular area flow prediction method based on multi-graph volume sum GRU.
Background
Traffic flow prediction is an important component of intelligent transportation systems. The purpose of regional flow prediction is to predict future flow values in urban areas based on given historical data, and accurate prediction can help traffic managers to control and manage flow in advance.
Regional flow prediction methods typically utilize spatial and temporal correlations between regions. Conventional regional flow prediction uses time series prediction methods such as autoregressive moving average model (ARIMA), time-varying poisson model, vector autoregressive model. The prediction accuracy is low only by considering the time-dependent correlation. With the rise of deep learning, researchers use deep learning models to predict flow. Compared with the traditional method, the model of the long Short-Term memory network LSTM (Long Short Term memory) and the gated cyclic unit GRU (gated Recurrent Unit) has better effect on the Short-Term time sequence prediction. However, they still focus on temporal associations only. To better capture the spatio-temporal correlation, researchers have proposed predicting regional traffic using Convolutional Neural Network (CNN) and residual neural network based methods, first dividing cities into grids and then predicting traffic at the grid level. However, these methods can only predict the traffic in regular areas.
Cities can be divided into meaningful areas based on road network information or administrative boundaries, which are often irregular and have complex topologies that carry more semantic information than regular grids. The grid-based prediction model cannot predict the traffic demand of the irregular area, and the usability of the prediction result is reduced.
Disclosure of Invention
In order to overcome the defects of the prior art and improve the accuracy of flow prediction in irregular areas, the invention adopts the following technical scheme:
the irregular area flow prediction method based on the multi-graph convolution sum GRU comprises the following steps:
dividing a region into N unconnected irregular regions;
step two, performing space-time simplification on the historical track data, and calculating to obtain the inflow and outflow of all the regions at each time step;
establishing a plurality of correlation diagrams among the regions, constructing a corresponding adjacent matrix, and expressing diversified spatial correlation among irregular regions;
designing a Multi-Graph Convolutional neural Network (MGCN) based on the association diagram among the regions, and fusing diversified spatial association characteristics among the regions to obtain a Multi-Graph Convolutional fusion result;
step five, based on the multi-graph convolution fusion result, capturing time correlation by adopting a GRU neural network;
selecting a proper loss function, training to obtain a prediction model, and predicting through the prediction model to obtain the inflow and the outflow of each region;
and step one, dividing the region into N unconnected irregular regions by adopting an irregular region division method based on the road network structure data of the region.
The second step is that the whole time period is divided into a plurality of time steps according to unit time, and then the original historical track data is mapped into the area according to the unit time based on the area division result, so as to obtain a simplified track:
TRSimp＝(startRegion，startDate，startHour，endRegion，endDate，endHour)
wherein startRegion is the ID of the departure area, startDate is the departure date, startHour is the departure unit time, endRegion is the ID of the arrival area, endDate is the arrival date, endHour is the arrival unit time, and then the entering amount of each area is obtained by aggregation simplified track calculation
Figure BDA0002282852750000021
And outflow volume
Figure BDA0002282852750000022
Wherein said
Figure BDA0002282852750000023
Represents the amount of entry of zone i at the t-th time step, said
Figure BDA0002282852750000024
Represents the outflow of zone i at the t-th time step, based finally on the
Figure BDA0002282852750000025
And said
Figure BDA0002282852750000026
Calculating to obtain the entering amount of all the areas under the time step t
Figure BDA0002282852750000027
And outflow volume
Figure BDA0002282852750000028
Establishing different association graphs to represent diversified spatial associations among irregular areas, wherein the spatial associations comprise a distance graph, a flow interaction graph and a flow association graph, the different association graphs are represented by G (V, E), and a node V is represented by G (V, E)iE.g. V represents an irregular area, an edge (V)i，vj) E encodes the degree of association between irregular regions, represented by the adjacency matrix A E RN×NRepresents;
the weight of an edge of the distance map is the distance between two regions, and the adjacent matrix element A in the distance mapdThe values of (i, j) are calculated as follows:
Figure BDA0002282852750000029
the dist (i, j) represents the distance between the center points of the regions i and j, and the adjacency matrix AdNormalized to [0,1 ]]And based on a predefined distance threshold thresdThe A is addeddConvert to an 0/1 matrix if Ad(i，j)≤thresdIndicating that the distance between the regions i and j is very close, let A bed(i, j) ═ 1, otherwise the said Ad(i，j)＝0；
The flow interaction diagram indicates whether there is frequent bi-direction between the two areasFlow rate by applying to said TRSimpAggregating the data to obtain the flow value fn (i, j) from the area i to the area j and the flow value fn (j, i) from the area j to the area i in the whole analysis time period, wherein the flow interaction diagram is adjacent to the matrix element AinterThe values of (i, j) are calculated as follows:
Figure BDA00022828527500000210
the adjacency matrix AinterNormalized to [0,1 ]]And based on a predefined traffic interaction threshold thresinterThe A is addedinterConvert to an 0/1 matrix if Ainter(i，j)≥thresinterIndicating that the interaction between the regions i and j is very strong, let A beinter(i, j) ═ 1, otherwise the said Ainter(i，j)＝0；
The flow correlation diagram indicates the time correlation of flow among the areas, the historical flow value of each area in each time step is obtained, the time sequence of the inlet flow and the outlet flow of each area in the time period is set according to the time period needing to be analyzed, and for the area i, the time sequence is expressed as:
Figure BDA00022828527500000211
the T represents the length of a time slot, correlation between the region i and the region j is calculated by adopting a Pearson correlation coefficient, and the adjacent matrix element A in the flow correlation diagramcorrThe values of (i, j) are calculated as follows:
Figure BDA0002282852750000031
h isiAnd h is saidjSaid time sequence representing said regions i and j will listen to said adjacency matrix AcorrNormalized to [0,1 ]]And based on a predefined flow-related threshold threscorrWill hear the statement AcorrIs converted into one0/1 matrix, if A iscorr(i，j)≥threscorrIndicating that the regions i and j have similar time usage patterns, let Acorr(i, j) is 1, otherwise said Acorr(i，j)＝0。
Based on a plurality of association graphs among the regions, the multi-graph convolution neural network is proposed to fully mine useful information hidden in different graphs and capture complex spatial dependence association, and the method comprises the following steps:
(1) using Graph Convolutional neural Network (GCN) model f (X)t′A) processing each graph at time step t':
Figure BDA0002282852750000032
Figure BDA0002282852750000033
Figure BDA0002282852750000034
said Xt′＝(It′，Ot′) Is the input of the model at the t' time step, representing the ingress and egress of the region,
Figure BDA0002282852750000035
is an adjacency matrix with self-flow, said INIs an identity matrix, the
Figure BDA0002282852750000036
Is a diagonal matrix in which
Figure BDA0002282852750000037
The W isd、Winter、WcorrIs a trainable weight matrix, said tanh (-) represents an activation function;
(2) adding additional attribute attr to each time stept′Coding the flow influence factors at each time step;
(3) combining a plurality of the association graphs and the additional attributes by adopting a Full Connected Layer (FCL) to fuse various spatial association characteristics among the areas;
FCL(Xt′，Ad，Ainter，Acorr，attrt′)＝Wfcl[fd(Xt′，Ad)，finter(Xt′，Ainter)，fcorr(Xt′，Acorr)，attrt′]
the W isfclFor the weight matrix, the FCL (-) is the multi-graph convolution fusion result.
The additional attribute comprises a date attribute
Figure BDA0002282852750000038
An hour attribute
Figure BDA0002282852750000039
Weather Properties
Figure BDA00022828527500000310
And temperature property
Figure BDA00022828527500000311
The above-mentioned
Figure BDA00022828527500000312
The dimension is 7 dimensions, and represents the week number of one week; the above-mentioned
Figure BDA00022828527500000313
The dimension is 24 dimensions, which represents the first hour of the day; the above-mentioned
Figure BDA00022828527500000314
Is divided into 8 categories including sunny, cloudy, light rain, medium rain, heavy rain, light snow, medium snow and heavy snow; the above-mentioned
Figure BDA00022828527500000315
Is divided into 8 levels from 10 DEG F to 90 DEG F, each 10 DEG F corresponds to one level, all the additional attributes use one-hot coding, and vectors are obtained by connection
Figure BDA00022828527500000316
Based on the multi-graph convolution fusion result FCL (·), capturing time sequence correlation in history by adopting a GRU neural network, wherein the GRU takes a hidden state at a time step t '-1 and the multi-graph convolution result as input to obtain a flow state at a time step t', and a calculation process is as follows:
ut′＝σ(Wu[FCL(·)，ht′-1]+bu)
rt′＝σ(Wr[FCL(·)，ht′-1]+br)
ct'＝tanh(Wc[FCL(·),(rt'⊙ht'-1)]+bc)
ht′＝ut′ht′-1+(1-ut′)ct′
h ist'-1Represents the hidden state at said time step t' -1, said ut'To update the gate, the amount of which memory of the last moment is saved to the current time step is defined, rt'To reset the gate, it is decided how to combine the new input information with the previous memory, ct'For the stored content at the t' time step, the ht'For the output state at the t 'time step, the FCL (-) is the multi-graph convolution result at the t' time step, the Wu、Wr、WcIs a weight matrix, said bu、br、bcIs a deviation vector, the σ is a sigmoid function, and the |, is an element-by-element multiplication operation.
And step six, training by adopting a back propagation and Adam optimization algorithm based on a training set by using Smooth L1 as the loss functionTraining the weight matrix and the deviation vector to obtain the prediction model, wherein the target of the prediction model is based on input historical data [ (I)0，O0)，(I1，O1)，...，(It-1，Ot-1)]Learning a function f (-) and mapping the inlet amount and the outlet amount of each area in the historical data to obtain the inlet amount and the outlet amount of the next time step
Figure BDA0002282852750000041
So that
Figure BDA0002282852750000042
Said (I)t，Ot) And selecting a model with the minimum root mean square error as a final prediction model for the real flow value of the next time step t according to the verification set, predicting by adopting the final prediction model based on the test set, and carrying out reverse normalization on an output result to obtain a final prediction result.
The invention has the advantages and beneficial effects that:
the invention provides a novel deep learning model facing irregular area traffic prediction, which is suitable for predicting the inflow and outflow of irregular areas in cities, uses a plurality of association graphs to encode diversified spatial associations among the irregular areas, designs a multi-graph convolution neural network to fuse the associations among the irregular areas to capture spatial dependence, and then adopts a GRU neural network to capture dynamic time sequence association, thereby effectively capturing complex time-space association at the same time, improving the practicability of prediction results, and simultaneously improving the prediction accuracy of irregular area traffic flow, so that the prediction results better assist the management and control of pedestrian and vehicle flow in urban areas.
Drawings
FIG. 1 is a flow chart of irregular area traffic prediction in accordance with the present invention.
FIG. 2 is a diagram of an irregular area traffic prediction model based on multi-graph convolution and GRU in the present invention.
Detailed Description
The invention is described in detail below with reference to the figures and the embodiments.
As shown in fig. 1, the method for predicting irregular area traffic based on multi-graph convolution and GRU of the present invention includes the following steps:
dividing a city into N unconnected irregular areas;
step two, performing space-time simplification on the historical track data, and calculating to obtain the inflow and outflow of all the regions at each time step;
establishing a plurality of correlation diagrams among the regions, constructing a corresponding adjacent matrix, and expressing diversified spatial correlation among irregular regions;
designing a multi-graph convolution neural network based on the correlation diagram among the regions, fusing diversified spatial correlation characteristics among the regions, and obtaining a multi-graph convolution fusion result;
step five, based on the multi-graph convolution fusion result, capturing time correlation by adopting a GRU neural network;
selecting a proper loss function, training to obtain a prediction model, and predicting through the prediction model to obtain the inflow and the outflow of each region;
and step one, dividing the city into N unconnected irregular areas by adopting an irregular area dividing method based on the road network structure data of the city.
Step two, firstly, dividing the whole time period into a plurality of time steps according to hours, then mapping the original historical track data into the region according to hours based on the region division result to obtain a simplified track:
TRSimp＝(startRegion，startDate，startHour，endRegion，endDate，endHour)
wherein startRegion is the ID of the departure area, startDate is the departure date, startHour is the departure hour, endRegion is the ID of the arrival area, endDate is the arrival date, endHour is the arrival hour, and then the entering amount of each area is obtained by aggregating simplified track calculation
Figure BDA0002282852750000051
And outflow volume
Figure BDA0002282852750000052
Wherein said
Figure BDA0002282852750000053
Represents the amount of entry of zone i at the t-th time step, said
Figure BDA0002282852750000054
Represents the outflow of zone i at the t-th time step, based finally on the
Figure BDA0002282852750000055
And said
Figure BDA0002282852750000056
Calculating to obtain the entering amount of all the areas under the time step t
Figure BDA0002282852750000057
And outflow volume
Figure BDA0002282852750000058
Establishing different association graphs to represent diversified spatial associations among irregular areas, wherein the spatial associations comprise a distance graph, a flow interaction graph and a flow association graph, the different association graphs are represented by G (V, E), and a node V is represented by G (V, E)iE.g. V represents an irregular area, an edge (V)i，vj) E encodes the degree of association between irregular regions, represented by the adjacency matrix A E RN×NRepresents;
the distance map, in which the weight of an edge is the distance between two regions, such that adjacent regions are connected by a higher-weight edge, is bounded by a neighboring matrix element AdThe values of (i, j) are calculated as follows:
Figure BDA0002282852750000059
the dist (i, j) represents the distance between the center points of the regions i and j, and the adjacency matrix AdNormalized to [0,1 ]]And based on a predefined distance threshold thresdThe A is addeddConvert to an 0/1 matrix if Ad(i，j)≤thresdIndicating that the distance between the regions i and j is very close, let A bed(i, j) ═ 1, otherwise the said Ad(i，j)＝0；
The flow interaction diagram indicates whether there is frequent bidirectional flow between the two areas by aiming at the TRSimpAggregating the data to obtain the flow value fn (i, j) from the area i to the area j and the flow value fn (j, i) from the area j to the area i in the whole analysis time period, wherein the flow interaction diagram is adjacent to the matrix element AinterThe values of (i, j) are calculated as follows:
Figure BDA00022828527500000510
the adjacency matrix AinterNormalized to [0,1 ]]And based on a predefined traffic interaction threshold thresinterThe A is addedinterConvert to an 0/1 matrix if Ainter(i，j)≥thresinterIndicating that the interaction between the regions i and j is very strong, let A beinter(i, j) ═ 1, otherwise the said Ainter(i，j)＝0；
The flow correlation diagram indicates the time correlation of the flow among the zones, the historical flow value of each zone at each time step is obtained, assuming that the analysis time period is 1 year, each zone has a time sequence with the length of 17520(365 × 24 × 2), the inlet flow and the outlet flow of each hour in one year are recorded, and for the zone i, the time sequence is expressed as:
Figure BDA0002282852750000061
the T represents the time period length, and the Pearson correlation coefficient is adopted to calculate the area i and the listening areaj, the adjacency matrix element A in the traffic correlation diagramcorrThe values of (i, j) are calculated as follows:
Figure BDA0002282852750000062
h isiAnd h is saidjRepresenting said time series of said regions i and j, said adjacency matrix AcorrNormalized to [0,1 ]]And based on a predefined flow-related threshold threscorrWill hear the statement AcorrConvert to an 0/1 matrix if Acorr(i，j)≥threscorrIndicating that the regions i and j have similar time usage patterns, let Acorr(i, j) is 1, otherwise said Acorr(i，j)＝0。
As shown in fig. 2, the fourth step is to propose the multi-map convolutional neural network to fully mine useful information hidden in different maps and capture complex spatially dependent relationships based on a plurality of relationship maps among regions, and includes the following steps:
(1) using a graph convolution neural network model f (X)t′A) processing each graph at time step t':
Figure BDA0002282852750000063
Figure BDA0002282852750000064
Figure BDA0002282852750000065
said Xt′＝(It′，Ot′) Is the input of the model at the t' time step, representing the ingress and egress of the region,
Figure BDA0002282852750000066
is an adjacency matrix with self-flow, said INIs an identity matrix, the
Figure BDA0002282852750000067
Is a diagonal matrix in which
Figure BDA0002282852750000068
The W isd、Winter、WcorrIs a trainable weight matrix, said tanh (-) represents an activation function;
(2) adding additional attribute attr to each time stept′Coding the flow influence factors at each time step;
(3) combining a plurality of association graphs and the additional attributes by adopting a full connection layer to fuse various spatial association characteristics among the regions;
FCL(Xt′，Ad，Ainter，Acorr，attrt′)＝Wfcl[fd(Xt′，Ad)，finter(Xt′，Ainter)，fcorr(Xt′，Acorr)，attrt′]
the W isfclFor the weight matrix, the FCL (-) is the multi-graph convolution fusion result.
The additional attribute comprises a date attribute
Figure BDA0002282852750000069
An hour attribute
Figure BDA00022828527500000610
Weather Properties
Figure BDA00022828527500000611
And temperature property
Figure BDA00022828527500000612
The above-mentioned
Figure BDA00022828527500000613
The dimension is 7 dimensions, and represents the week number of one week; the above-mentioned
Figure BDA00022828527500000614
The dimension is 24 dimensions, which represents the first hour of the day; the above-mentioned
Figure BDA00022828527500000615
Is divided into 8 categories including sunny, cloudy, light rain, medium rain, heavy rain, light snow, medium snow and heavy snow; the above-mentioned
Figure BDA00022828527500000616
Is divided into 8 levels from 10 DEG F to 90 DEG F, each 10 DEG F corresponds to one level, all the additional attributes use one-hot coding, and vectors are obtained by connection
Figure BDA00022828527500000617
The fifth step, based on the multi-graph convolution fusion result FCL (-), capturing the time sequence correlation in the traffic flow history by using a GRU neural network, taking the hidden state at the t '-1 time step and the multi-graph convolution result as input by the GRU, obtaining the flow state at the t' time step, and calculating the flow as follows:
ut′＝σ(Wu[FCL(·)，ht′-1]+bu)
rt′＝σ(Wr[FCL(·)，ht′-1]+br)
ct'＝tanh(Wc[FCL(·),(rt'⊙ht'-1)]+bc)
ht′＝ut′ht′-1+(1-ut′)ct′
h ist'-1Represents the hidden state at said time step t' -1, said ut'To update the gate, the amount of which memory of the last moment is saved to the current time step is defined, rt'To reset the gate, it is decided how to memorize the new input information with the previous oneIn combination with ct'For the stored content at the t' time step, the ht'Is said t'Output state at time step, FCL (-) being the multi-graph convolution result at time step t', Wu、Wr、WcIs a weight matrix, said bu、br、bcIs a deviation vector, the σ is a sigmoid function, and the |, is an element-by-element multiplication operation.
Step six, using Smooth L1 as the loss function, training the weight matrix and the deviation vector by adopting a back propagation and Adam optimization algorithm based on a training set to obtain the prediction model, wherein the target of the prediction model is based on input historical data [ (I)0，O0)，(I1，O1)，...，(It-1，Ot-1)]Learning a function f (-) and mapping the inlet amount and the outlet amount of each area in the historical data to obtain the inlet amount and the outlet amount of the next time step
Figure BDA0002282852750000071
So that
Figure BDA0002282852750000072
Said (I)t，Ot) And selecting a model with the minimum Root Mean Square Error (RMSE) as a final prediction model for the real flow value of the next time step t according to the verification set, predicting by adopting the final prediction model based on the test set, and performing reverse normalization on an output result to obtain a final prediction result.

Claims (6)

1. The irregular area flow prediction method based on the multi-graph convolution sum GRU is characterized by comprising the following steps of:
dividing a region into N unconnected irregular regions;
step two, performing space-time simplification on the historical track data, and calculating to obtain the inflow and outflow of all the regions at each time step;
establishing a plurality of correlation diagrams among the regions, constructing a corresponding adjacent matrix, and expressing diversified spatial correlation among irregular regions; establishing different association graphs to represent diversified spatial associations among irregular areas, wherein the spatial associations comprise a distance graph, a flow interaction graph and a flow association graph;
designing a multi-graph convolution neural network based on the correlation diagram among the regions, fusing diversified spatial correlation characteristics among the regions, and obtaining a multi-graph convolution fusion result; based on a plurality of correlation graphs among the regions, the multi-graph convolution neural network is proposed to fully mine useful information hidden in different graphs and capture complex spatial dependence correlation, and the method comprises the following steps:
(1) using a graph convolution neural network model f (X)t'A) processing each graph at time step t':
Figure FDA0003484952040000011
Figure FDA0003484952040000012
Figure FDA0003484952040000013
said Xt'＝(It',Ot') Is the input of the model at the t' time step, representing the ingress and egress of the region,
Figure FDA0003484952040000014
is an adjacency matrix with self-flow, said INIs an identity matrix, the
Figure FDA0003484952040000015
Is a diagonal matrix in which
Figure FDA0003484952040000016
The W isd、Winter、WcorrIs a trainable weight matrix, said tanh (-) represents an activation function; a is describeddRepresenting the adjacency matrix in the distance map, said AinterRepresenting the adjacency matrix in the traffic interaction graph, said AcorrRepresenting an adjacency matrix in the traffic correlation diagram;
(2) adding additional attribute attr to each time stept'Coding the flow influence factors at each time step;
(3) combining a plurality of association graphs and the additional attributes by adopting a full connection layer to fuse various spatial association characteristics among the regions;
FCL(Xt',Ad,Ainter,Acorr,attrt')＝Wfcl[fd(Xt',Ad),finter(Xt',Ainter),fcorr(Xt',Acorr),attrt']
the W isfclThe FCL (-) is a multi-graph convolution fusion result as a weight matrix;
step five, based on the multi-graph convolution fusion result, capturing time correlation by adopting a GRU neural network; based on the multi-graph convolution fusion result FCL (·), capturing time sequence correlation in history by adopting a GRU neural network, taking a hidden state at a t '-1 time step and the multi-graph convolution result as input by the GRU, and obtaining a flow state at a t' time step, wherein the calculation process is as follows:
ut'＝σ(Wu[FCL(·),ht'-1]+bu)
rt'＝σ(Wr[FCL(·),ht'-1]+br)
ct'＝tanh(Wc[FCL(·),(rt'⊙ht'-1)]+bc)
ht'＝ut'ht'-1+(1-ut')ct'
h ist'-1Represents the hidden state at said time step t' -1, said ut'For updating the door, it is defined that the last moment will beWhich memory is saved to the amount of the current time step, rt'To reset the gate, it is decided how to combine the new input information with the previous memory, ct'For the stored content at the t' time step, the ht'For the output state at the t 'time step, the FCL (-) is the multi-graph convolution result at the t' time step, the Wu、Wr、WcIs a weight matrix, said bu、br、bcIs a deviation vector, the σ is a sigmoid function, the | _ is an element-by-element multiplication operation;
and step six, selecting a loss function, training to obtain a prediction model, and predicting through the prediction model to obtain the inflow and outflow of each region.
2. The irregular area traffic prediction method based on multi-graph convolution and GRU as claimed in claim 1, wherein in the first step, the region is divided into N unconnected irregular areas by an irregular area division method based on road network structure data of the region.
3. The irregular area traffic prediction method based on multi-map convolution and GRU as claimed in claim 1, wherein in step two, the whole time period of processing is firstly divided into a plurality of time steps according to unit time, and then based on the area division result, the original historical track data is mapped into the area according to unit time, so as to obtain a simplified track:
TRSimp＝(startRegion,startDate,startHour,endRegion,endDate,endHour)
wherein startRegion is the ID of the departure area, startDate is the departure date, startHour is the departure unit time, endRegion is the ID of the arrival area, endDate is the arrival date, endHour is the arrival unit time, and then the entering amount of each area is obtained by aggregating simplified track calculation
Figure FDA0003484952040000021
And outflow volume
Figure FDA0003484952040000022
Wherein said
Figure FDA0003484952040000023
Represents the amount of entry of zone i at the t-th time step, said
Figure FDA0003484952040000024
Represents the outflow of zone i at the t-th time step, based finally on the
Figure FDA0003484952040000025
And said
Figure FDA0003484952040000026
Calculating to obtain the entering amount of all the areas under the time step t
Figure FDA0003484952040000027
And outflow volume
Figure FDA0003484952040000028
4. The irregular area traffic prediction method based on multi-graph convolution and GRU (generalized regression analysis) of claim 3, wherein in the third step, different association graphs are established to represent diversified spatial associations between irregular areas, including distance graphs, traffic interaction graphs and traffic association graphs, the different association graphs are all represented by G ═ V (E), and node V is represented by G ═ EiE.g. V represents an irregular region, an edge (V)i,vj) E encodes the degree of association between irregular regions, from the adjacency matrix
Figure FDA0003484952040000029
Represents;
the weight of an edge of the distance map is the distance between two regions, and the adjacent matrix element A in the distance mapdThe values of (i, j) are calculated as follows:
Figure FDA00034849520400000210
the dist (i, j) represents the distance between the center points of the regions i and j, and the adjacency matrix AdNormalized to [0,1 ]]And based on a predefined distance threshold thresdThe A is addeddConvert to an 0/1 matrix if Ad(i,j)≤thresdIndicating that the distance between the regions i and j is very close, let A bed(i, j) ═ 1, otherwise the said Ad(i,j)＝0；
The flow interaction diagram indicates whether there is frequent bidirectional flow between the two areas, by the pair TRSimpAggregating the data to obtain the flow value fn (i, j) from the area i to the area j and the flow value fn (j, i) from the area j to the area i in the whole analysis time period, wherein the flow interaction diagram is adjacent to the matrix element AinterThe values of (i, j) are calculated as follows:
Figure FDA00034849520400000211
the adjacency matrix AinterNormalized to [0,1 ]]And based on a predefined traffic interaction threshold thresinterThe A is addedinterConvert to an 0/1 matrix if Ainter(i,j)≥thresinterIndicating that the interaction between the regions i and j is very strong, let A beinter(i, j) ═ 1, otherwise the said Ainter(i,j)＝0；
The flow correlation diagram indicates the time correlation of flow among the areas, the historical flow value of each area in each time step is obtained, the time sequence of the inlet flow and the outlet flow of each area in the time period is set according to the time period needing to be analyzed, and for the area i, the time sequence is expressed as:
Figure FDA0003484952040000031
the T represents the time period length, the correlation between the area i and the area j is calculated by adopting a Pearson correlation coefficient, and the adjacent matrix element A in the flow correlation diagramcorrThe values of (i, j) are calculated as follows:
Figure FDA0003484952040000032
h isiAnd h is saidjRepresenting said time series of said regions i and j, said adjacency matrix AcorrNormalized to [0,1 ]]And based on a predefined flow-related threshold threscorrThe A is addedcorConvert to an 0/1 matrix if Acorr（ i ， j ）≥ threscorrIndicating that the regions i and j have similar temporal usage patterns, let Acorr(i, j) ═ 1, otherwise the said Acorr(i,j)＝0。
5. The method of claim 1, wherein the additional attributes comprise a date attribute
Figure FDA0003484952040000034
An hour attribute
Figure FDA0003484952040000035
Weather Properties
Figure FDA0003484952040000036
And temperature property
Figure FDA0003484952040000037
The above-mentioned
Figure FDA0003484952040000038
Dimension is 7, which represents week of weekA table; the above-mentioned
Figure FDA0003484952040000039
The dimension is 24 dimensions, which represents the hours of the day; the above-mentioned
Figure FDA00034849520400000310
Is divided into 8 categories including sunny, cloudy, light rain, medium rain, heavy rain, light snow, medium snow and heavy snow; the above-mentioned
Figure FDA00034849520400000311
Is divided into 8 levels from 10 DEG F to 90 DEG F, each 10 DEG F corresponds to one level, all the additional attributes use one-hot coding, and vectors are obtained by connection
Figure FDA00034849520400000312
6. The irregular area traffic prediction method based on multi-graph convolution and GRU (generalized regression analysis) as claimed in claim 1, wherein in the sixth step, the weighting matrix and the deviation vector are trained by using a back propagation and Adam optimization algorithm based on a training set by using Smooth L1 as the loss function to obtain the prediction model, and the goal of the prediction model is based on input historical data [ (I) that0,O0),(I1,O1),...,(It-1,Ot-1)]Learning a function f (-) and mapping the inlet amount and the outlet amount of each area in the historical data to obtain the inlet amount and the outlet amount of the next time step
Figure FDA00034849520400000313
So that
Figure FDA00034849520400000314
Said (I)t,Ot) Selecting a model with the minimum root mean square error as a final prediction model according to the verification set for the real flow value of the next time step t, and based on the test setAnd predicting by adopting the final prediction model, and performing reverse normalization on an output result to obtain a final prediction result.
CN201911148344.6A 2019-11-21 2019-11-21 Irregular area flow prediction method based on multi-graph convolution sum GRU Active CN110991713B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911148344.6A CN110991713B (en) 2019-11-21 2019-11-21 Irregular area flow prediction method based on multi-graph convolution sum GRU

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911148344.6A CN110991713B (en) 2019-11-21 2019-11-21 Irregular area flow prediction method based on multi-graph convolution sum GRU

Publications (2)

Publication Number Publication Date
CN110991713A CN110991713A (en) 2020-04-10
CN110991713B true CN110991713B (en) 2022-04-01

Family

ID=70085720

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911148344.6A Active CN110991713B (en) 2019-11-21 2019-11-21 Irregular area flow prediction method based on multi-graph convolution sum GRU

Country Status (1)

Country Link
CN (1) CN110991713B (en)

Families Citing this family (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111639787B (en) * 2020-04-28 2024-03-15 北京工商大学 Spatio-temporal data prediction method based on graph convolution network
WO2021221563A1 (en) * 2020-04-30 2021-11-04 Grabtaxi Holdings Pte. Ltd. Method for predicting the destination location of a vehicle
CN111583263B (en) * 2020-04-30 2022-09-23 北京工业大学 Point cloud segmentation method based on joint dynamic graph convolution
CN111651504B (en) * 2020-06-03 2021-10-08 湖南大学 Multi-element time sequence multilayer space-time dependence modeling method based on deep learning
CN111695679B (en) * 2020-06-09 2023-12-29 北京百度网讯科技有限公司 Method and device for predicting access amount, electronic equipment and storage medium
CN111882925A (en) * 2020-07-27 2020-11-03 交通运输部水运科学研究所 Shipping traffic flow prediction system based on information propagation diagram and recurrent neural network
CN112200351A (en) * 2020-09-24 2021-01-08 深圳市综合交通运行指挥中心 Urban area passenger flow volume prediction method based on mobile phone signaling data
CN112562312B (en) * 2020-10-21 2022-10-28 浙江工业大学 GraphSAGE traffic network data prediction method based on fusion features
CN112419710B (en) * 2020-10-22 2022-07-26 深圳云天励飞技术股份有限公司 Traffic congestion data prediction method, traffic congestion data prediction device, computer equipment and storage medium
CN112561118B (en) * 2020-10-29 2022-09-02 北京水慧智能科技有限责任公司 Municipal pipe network water flow prediction method based on GRU neural network
CN112383516A (en) * 2020-10-29 2021-02-19 博雅正链(北京)科技有限公司 Graph neural network construction method and abnormal flow detection method based on graph neural network
CN112801355B (en) * 2021-01-20 2022-05-24 南京航空航天大学 Data prediction method based on multi-graph fusion space-time attention of long-short-term space-time data
CN112819213B (en) * 2021-01-22 2022-06-14 华南理工大学 Highway freight volume prediction method and system based on deep learning network
CN112911626B (en) * 2021-02-01 2022-05-31 福州大学 Wireless network flow prediction method based on multi-graph convolution
CN114358375B (en) * 2021-11-29 2024-05-24 重庆邮电大学 Crowd density prediction method and system based on big data
CN114358213B (en) * 2022-03-08 2022-06-10 湖南大学 Error ablation processing method, system and medium for nonlinear time series data prediction
CN114973653B (en) * 2022-04-27 2023-12-19 中国计量大学 Traffic flow prediction method based on space-time diagram convolutional network
CN115018553B (en) * 2022-06-30 2024-05-07 东南大学 Regional logistics single quantity prediction system and method based on deep learning
CN116405976B (en) * 2023-06-06 2023-09-22 中国民用航空飞行学院 ADS-B-based data bidirectional communication optimization method and system

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107967532A (en) * 2017-10-30 2018-04-27 厦门大学 The Forecast of Urban Traffic Flow Forecasting Methodology of integration region vigor
WO2018155397A1 (en) * 2017-02-24 2018-08-30 株式会社日立製作所 Congestion forecasting system and pedestrian simulation device
CN109389244A (en) * 2018-09-06 2019-02-26 浙江鸿程计算机***有限公司 Tourist's number prediction technique in a kind of short-term scenic spot of multifactor perception based on GRU
CN110119482A (en) * 2019-05-13 2019-08-13 杭州电子科技大学 Based on the crowd of POI and multi-source mobile data collection trip mode visible analysis method

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018155397A1 (en) * 2017-02-24 2018-08-30 株式会社日立製作所 Congestion forecasting system and pedestrian simulation device
CN107967532A (en) * 2017-10-30 2018-04-27 厦门大学 The Forecast of Urban Traffic Flow Forecasting Methodology of integration region vigor
CN109389244A (en) * 2018-09-06 2019-02-26 浙江鸿程计算机***有限公司 Tourist's number prediction technique in a kind of short-term scenic spot of multifactor perception based on GRU
CN110119482A (en) * 2019-05-13 2019-08-13 杭州电子科技大学 Based on the crowd of POI and multi-source mobile data collection trip mode visible analysis method

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
基于卷积循环神经网络的城市区域车流量预测模型;薛佳瑶等;《信息工程大学学报》;20190415(第02期);全文 *
基于多图卷积网络和门控循环单元的不规则区域交通流量预测（英文）;僧德文 等;《Frontiers of Information Technology & Electronic Engineering》;20210729;全文 *
基于门控循环单元的多因素感知短期游客人数预测模型;王敬昌 等;《浙江大学学报（工学版）》;20190906;全文 *

Also Published As

Publication number Publication date
CN110991713A (en) 2020-04-10

Similar Documents

Publication Publication Date Title
CN110991713B (en) Irregular area flow prediction method based on multi-graph convolution sum GRU
CN109658695B (en) Multi-factor short-term traffic flow prediction method
CN110223517B (en) Short-term traffic flow prediction method based on space-time correlation
Lin et al. Quantifying uncertainty in short-term traffic prediction and its application to optimal staffing plan development
CN115240425B (en) Traffic prediction method based on multi-scale space-time fusion graph network
CN110827543A (en) Short-term traffic flow control method based on deep learning and spatio-temporal data fusion
CN112489426B (en) Urban traffic flow space-time prediction scheme based on graph convolution neural network
CN111612243A (en) Traffic speed prediction method, system and storage medium
CN112071062B (en) Driving time estimation method based on graph convolution network and graph attention network
Minglei et al. Classified real-time flood forecasting by coupling fuzzy clustering and neural network
CN112863182B (en) Cross-modal data prediction method based on transfer learning
CN111242395B (en) Method and device for constructing prediction model for OD (origin-destination) data
CN111242292A (en) OD data prediction method and system based on deep space-time network
US20240143999A1 (en) Multi-modal data prediction method based on causal markov model
CN115204477A (en) Bicycle flow prediction method of context awareness graph recursive network
CN113327417A (en) Traffic flow prediction method based on 3D dynamic space-time residual convolution associated network
CN114572229A (en) Vehicle speed prediction method, device, medium and equipment based on graph neural network
CN116913098B (en) Short-time traffic flow prediction method integrating air quality and vehicle flow data
KR100540009B1 (en) Water need forecasting method using neural network
CN116227738B (en) Method and system for predicting traffic interval of power grid customer service
CN114566048B (en) Traffic control method based on multi-view self-adaptive space-time diagram network
CN115589369A (en) Base station flow prediction method and system based on double-attention machine system
CN115796030A (en) Traffic flow prediction method based on graph convolution
CN117037461A (en) Road network traffic jam prediction method based on multi-weight graph three-dimensional convolution
CN113869990A (en) Shop lease pricing method based on space-time representation learning model

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant