CN112365708B - Scenic spot traffic volume prediction model establishing and predicting method based on multi-graph convolution network - Google Patents

Scenic spot traffic volume prediction model establishing and predicting method based on multi-graph convolution network Download PDF

Info

Publication number
CN112365708B
CN112365708B CN202011052408.5A CN202011052408A CN112365708B CN 112365708 B CN112365708 B CN 112365708B CN 202011052408 A CN202011052408 A CN 202011052408A CN 112365708 B CN112365708 B CN 112365708B
Authority
CN
China
Prior art keywords
scenic spot
scenic
traffic flow
map
target
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
CN202011052408.5A
Other languages
Chinese (zh)
Other versions
CN112365708A (en
Inventor
张蕾
施元磊
高原
张小溪
王洁
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.)
Northwest University
Original Assignee
Northwest 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 Northwest University filed Critical Northwest University
Priority to CN202011052408.5A priority Critical patent/CN112365708B/en
Publication of CN112365708A publication Critical patent/CN112365708A/en
Application granted granted Critical
Publication of CN112365708B publication Critical patent/CN112365708B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/01—Detecting movement of traffic to be counted or controlled
    • G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
    • G08G1/0125—Traffic data processing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/901—Indexing; Data structures therefor; Storage structures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • 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
    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/01—Detecting movement of traffic to be counted or controlled
    • G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
    • G08G1/0137—Measuring and analyzing of parameters relative to traffic conditions for specific applications
    • G08G1/0145—Measuring and analyzing of parameters relative to traffic conditions for specific applications for active traffic flow control
    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/065—Traffic control systems for road vehicles by counting the vehicles in a section of the road or in a parking area, i.e. comparing incoming count with outgoing count

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Chemical & Material Sciences (AREA)
  • Analytical Chemistry (AREA)
  • Databases & Information Systems (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Traffic Control Systems (AREA)

Abstract

The invention belongs to the field of data mining and urban traffic data analysis, and discloses a scenic spot traffic volume prediction model building and predicting method based on a multi-graph convolution network. The method comprises the following steps: 1. constructing a plurality of characteristic graphs representing the relationship between scenic spots; 2. constructing a convolution cyclic neural network model based on a plurality of characteristic graphs; 3. training a convolution cyclic neural network model; 4. and predicting the future traffic flow value of the scenic spot by using the trained convolutional recurrent neural network model. The novel point of the invention is that an urban scenic spot network is constructed from the angle of the map, the relation of different scenic spots is reflected by constructing a plurality of feature maps, then the future traffic flow of the scenic spots is predicted by fusing the plurality of feature maps and capturing the space-time features at the same time, the hourly traffic flow value is obtained, the accuracy of the medium-long term prediction is improved, and the convergence speed of the training and the robustness of the algorithm are improved.

Description

Scenic spot traffic volume prediction model establishing and predicting method based on multi-graph convolution network
Technical Field
The invention belongs to the field of data mining and urban traffic data analysis, and particularly relates to a scenic spot traffic volume prediction model building and predicting method based on a multi-graph convolution network.
Background
With the popularity of portable GPS smart devices and the maturity of social media platforms, a large amount of heterogeneous data containing both temporal and spatial information is generated. Such as taxi GPS track data published by the duet corporation geya program, tourist travel data of a hornet tourism platform, Foursquare point of interest (POI) check-in data, and the like, these multi-source heterogeneous data provide powerful support for regional traffic prediction.
The cultural tourist attraction is an important interest point in the city and is an important component of the city. Accurate and timely scenic spot traffic flow prediction is important content of urban intelligent traffic system research, and the research is helpful for traffic departments to take measures in time to relieve congestion and improve the utilization rate of a road network. Meanwhile, the method is beneficial to reducing the travel time and the cost of the traveler and improving the experience of the traveler. In addition, the method can be widely applied to position-based applications such as point of interest recommendation, path planning and city planning.
The existing research techniques mainly include: statistical-based methods, machine-learning-based methods, and deep-learning-based methods. The traffic flow prediction method based on statistics only focuses on the regularity of time sequence data, cannot analyze the spatial characteristics and dynamic changes of a traffic system, and is easily interfered by abnormal values; the method based on machine learning performs well in short-term prediction, but the robustness and the effect of long-term prediction need to be improved; the existing traffic flow prediction method based on deep learning only considers the characteristics of a traffic network, such as network connectivity, during spatial dependency modeling, but does not fully apply semantic information contained in a prediction object, such as popularity and functionality of a prediction region, and the factors have considerable influence on regional traffic flow.
Disclosure of Invention
The invention aims to provide a scenic spot traffic volume prediction model establishing and predicting method based on a multi-graph convolution network, which is used for solving the problems in the prior art and the like.
In order to realize the task, the invention adopts the following technical scheme:
a scenic spot traffic prediction model building method based on a multi-graph convolution network comprises the following steps:
step 1: acquiring multi-source heterogeneous data of a scenic spot, extracting features of the multi-source heterogeneous data of the scenic spot, and acquiring a scenic spot popularity feature map, a scenic spot function similarity feature map, a scenic spot distance feature map and a scenic spot traffic accessibility feature map;
step 2: obtaining historical traffic flow of a scenic spot, extracting characteristics of the historical traffic flow of the scenic spot, and obtaining a historical traffic flow matrix of the scenic spot;
and step 3: establishing a multi-map convolution recurrent neural network model, taking a historical traffic flow matrix, a scenic spot popularity characteristic diagram, a scenic spot functional similarity characteristic diagram, a scenic spot distance characteristic diagram and a scenic spot traffic accessibility characteristic diagram as inputs, taking scenic spot predicted traffic flow as an output, training the model, and taking the trained model as a scenic spot traffic flow prediction model;
the multi-map convolution recurrent neural network model comprises a map convolution network and a recurrent gate control unit, wherein the map convolution network is used for outputting the depth characteristics of a scenic region characteristic map according to a historical traffic flow matrix, the scenic region popularity characteristic map, a scenic region function similarity characteristic map, a scenic region distance characteristic map and a scenic region traffic accessibility characteristic map and establishing a multi-feature fusion matrix according to the depth characteristics of the scenic region characteristic map, and the recurrent gate control unit is used for outputting a scenic region predicted traffic flow according to the multi-feature fusion matrix.
Further, the adjacency matrix P of the scenic spot popularity feature map is obtained by equation 1:
Figure BDA0002709959740000031
wherein, ViAnd VjFor any ith and jth scenic spot in the set of scenic spots,
Figure BDA00027099597400000318
and
Figure BDA00027099597400000319
are each ViAnd VjThe number of user comments (a) to be made,
Figure BDA0002709959740000032
is a ViAnd VjThe degree of similarity in popularity of (a) a,
Figure BDA0002709959740000033
has a value range of [0,1 ]]And i and j are positive integers.
Further, the adjacency matrix F of the scenic spot functional similarity feature map is obtained by equation 2:
Figure BDA0002709959740000034
wherein, ViAnd VjFor any ith and jth scenic spot in the set of scenic spots,
Figure BDA0002709959740000035
and
Figure BDA0002709959740000036
are each ViAnd VjA vector consisting of the number of POIs of each category, r represents the total number of categories of POIs, k represents the kth category,
Figure BDA0002709959740000037
represents ViThe number of POI class k in the image,
Figure BDA0002709959740000038
represents VjThe number of POI class k in the image,
Figure BDA0002709959740000039
i. j, r and k are positive integers.
Further, the adjacency matrix L of the scenic distance feature map is obtained by equation 3:
Figure BDA00027099597400000310
wherein, ViAnd VjFor any ith and jth scenic spot in the set of scenic spots, dist (V)i,Vj) Represents ViAnd VjAnd a distance therebetween
Figure BDA00027099597400000311
Figure BDA00027099597400000312
R is the radius of the earth,
Figure BDA00027099597400000313
and
Figure BDA00027099597400000314
are each ViAnd VjThe longitude of (a) is determined,
Figure BDA00027099597400000315
and
Figure BDA00027099597400000316
are each ViAnd VjThe latitude of (c), maxmin represents the maximum and minimum normalization, n represents the number of scenic spots, i and j are belonged to n, and i, j and n are positive integers.
Further, the adjacency matrix T of the scenic spot traffic accessibility feature map is obtained by equation 4:
Figure BDA00027099597400000317
wherein, Ti,jThe traffic access degree of any ith scenic spot and jth scenic spot is represented, and the contementtransmission means that the ith scenic spot and the jth scenic spot are on the same road section.
Further, the historical traffic flow matrix of the scenic spot is obtained by adopting a formula 5
Figure BDA0002709959740000041
Wherein x ismnRepresents the traffic flow value of the nth scenic spot at the mth time interval, and m represents the time intervalThe number of intervals, n, indicates the number of scenic spots.
Further, the loss function in the step 3 training is:
Figure BDA0002709959740000042
wherein, YtRepresenting the actual traffic flow during the t period,
Figure BDA0002709959740000043
indicating predicted traffic flow, L, over a period of tregIs the L2 regularization term, μ is to [0,1 ]]Is determined.
A scenic spot traffic volume prediction method based on a multi-graph convolution network comprises the following steps:
step a: acquiring historical traffic flow of a target scene and multi-source heterogeneous data of the target scene;
step b: extracting features from multi-source heterogeneous data of a target scenic spot to obtain a scenic spot popularity feature map of the target scenic spot, a scenic spot function similarity feature map of the target scenic spot, a scenic spot distance feature map of the target scenic spot and a scenic spot traffic accessibility feature map of the target scenic spot; extracting characteristics of historical traffic flow of a target scenic spot to obtain a historical traffic flow matrix of the target scenic spot;
step c: inputting a scenic spot popularity characteristic diagram of a target scenic spot, a scenic spot function similarity characteristic diagram of the target scenic spot, a scenic spot distance characteristic diagram of the target scenic spot, a scenic spot traffic accessibility characteristic diagram of the target scenic spot and a historical traffic flow matrix of the target scenic spot into a scenic spot predicted traffic flow, and outputting the predicted traffic flow of the target scenic spot.
Further, the scenic spot multi-source heterogeneous data comprises scenic spot comment data, scenic spot POI data, geographical coordinates of the scenic spot and road network data.
Compared with the prior art, the invention has the following technical characteristics:
(1) the method and the device combine the spatial characteristics and the time characteristics to accurately predict the traffic flow of the scene area. The scheme uses a Graph Convolution Network (GCN) to extract complex spatial correlation characteristics among different scenic spots; and (3) extracting time characteristics of historical traffic flow of the scenic spot by using a gated loop unit (GRU), and finally combining the time characteristics and the GRU to realize accurate prediction of the scenic spot traffic flow.
(2) In the invention, in the aspect of extracting the spatial correlation between scenic region areas, a Graph Convolution Network (GCN) promoted by the CNN can process data with any graph structure. The GCN model has good application effects in a plurality of frontier fields, such as document classification, unsupervised learning, image classification and the like. Therefore, the present solution uses a GCN model to learn spatial correlation features between scenic regions.
(3) In the aspect of extracting the time characteristics of historical traffic flow of scenic spots, the most widely used neural network model in the trend prediction of sequence data is a Recurrent Neural Network (RNN). However, the conventional recurrent neural network has a limitation in long-term prediction due to the defect of gradient disappearance or gradient explosion existing in the Recurrent Neural Network (RNN) itself. The LSTM model and GRU model are variants of the recurrent neural network and have been shown to solve the above problems. Compared with an LSTM model which is complex in structure, has more parameters and needs longer training time, the GRU model is simpler in structure and less in parameters, and can learn time sequence characteristics more quickly. Therefore, the method selects a GRU model to capture the time correlation from the historical traffic data of the scenic spot.
Drawings
FIG. 1 is a schematic diagram showing the distribution of 4 famous tourist sites in the city of Xian in the example;
FIG. 2 is a traffic flow trend chart of 4 famous tourist attractions in the city of Xian in the example;
FIG. 3 is a block diagram of a spatial feature model of a traffic flow prediction model for a multi-map convolutional recurrent neural network scenic region;
FIG. 4 is a graph of the effect of the number of Gated Round Units (GRUs) on the accuracy of traffic flow prediction;
FIG. 5 is an illustration of the effect of an input historical traffic flow duration on the accuracy of traffic flow predictions;
FIG. 6 is a graph of the effect of the number of gated round robin unit (GRU) hidden units on the accuracy of traffic flow prediction;
FIG. 7 is a graph of traffic flow prediction accuracy versus predicted future time duration for a multi-map convolution recurrent neural network;
FIG. 8 is a graph comparing training time cost of traffic flow prediction model of multi-graph convolution recurrent neural network scenic spot with DCRNN and T-GCN;
FIG. 9 is a graph of the effect of different profiles on the prediction accuracy of scenic traffic flow;
FIG. 10 is a traffic flow prediction result of the multi-graph convolution recurrent neural network prediction model in the area of museum in Changan district in accordance with the embodiment;
FIG. 11 is a traffic flow prediction result of the multi-graph convolutional recurrent neural network prediction model in the region where the museum in history of Shaanxi is located in the embodiment.
Detailed Description
Graph Convolution Network (GCN): the GCN model has good application effects in a plurality of frontier fields, such as document classification, unsupervised learning, image classification and the like. The method is used for extracting complex spatial correlation characteristics among different scenic spots.
Gated cycle unit (GRU): the GRU model has a simpler structure and fewer parameters, and can learn time sequence characteristics more quickly. To extract the time characteristics of the historical traffic flow of the scenic spot.
Adjacency matrix: the nodes on the feature graph are scenic spots in the invention by using the adjacent relation among the nodes representing the feature graph.
In this embodiment, the scenic spot set refers to a set V of different scenic spots, and the scenic spot set includes a plurality of different scenic spots ViThe scenic spots and the scenic spots have the same meaning, and each scenic spot is embodied as a specific coordinate value on the map.
Traffic flow: refers to the number of vehicles passing by in a certain time.
The embodiment discloses a scenic spot traffic volume prediction model building method based on a multi-graph convolution network, which comprises the following steps:
step 1: acquiring multi-source heterogeneous data of a scenic spot, extracting features of the multi-source heterogeneous data of the scenic spot, and acquiring a scenic spot popularity feature map, a scenic spot function similarity feature map, a scenic spot distance feature map and a scenic spot traffic accessibility feature map;
step 2: obtaining historical traffic flow of a scenic spot, extracting characteristics of the historical traffic flow of the scenic spot, and obtaining a historical traffic flow matrix of the scenic spot;
and step 3: establishing a multi-map convolution recurrent neural network model, taking a historical traffic flow matrix, a scenic spot popularity characteristic diagram, a scenic spot functional similarity characteristic diagram, a scenic spot distance characteristic diagram and a scenic spot traffic accessibility characteristic diagram as inputs, taking scenic spot predicted traffic flow as an output, training the model, and taking the trained model as a scenic spot traffic flow prediction model;
the multi-map convolution recurrent neural network model comprises a map convolution network and a recurrent gate control unit, wherein the map convolution network is used for outputting the depth characteristics of a scenic region characteristic map according to a historical traffic flow matrix, the scenic region popularity characteristic map, a scenic region function similarity characteristic map, a scenic region distance characteristic map and a scenic region traffic accessibility characteristic map and establishing a multi-feature fusion matrix according to the depth characteristics of the scenic region characteristic map, and the recurrent gate control unit is used for outputting a scenic region predicted traffic flow according to the multi-feature fusion matrix.
Specifically, the scenic spot multi-source heterogeneous data comprises scenic spot comment data, scenic spot POI data, geographical coordinates of scenic spots and road network data.
The scenic spot comment data are cultural scenic spot comment data obtained through American group comment, the scenic spot POI data are POI data in a to-be-predicted range around a scenic spot obtained through an open platform of an Baidu map, the geographical coordinates of the scenic spot are geographical coordinate data of the scenic spot obtained through the open platform of the Baidu map, and the road network data are urban road network data obtained through OpenStreetMap.
Preferably, the adjacency matrix P of the scenic spot popularity feature map is obtained by equation 1:
Figure BDA0002709959740000081
wherein, ViAnd VjFor any ith and jth scenic spot in the set of scenic spots,
Figure BDA0002709959740000082
and
Figure BDA0002709959740000083
are each ViAnd VjThe number of user comments (a) to be made,
Figure BDA0002709959740000084
is a ViAnd VjI and j are positive integers.
In the scenic spot popularity characteristic graph, scenic spots are taken as nodes of the graph, and the popularity similarity matrix P is taken as an adjacent matrix of the graph, and as the scenic spot popularity is directly expressed by visitors to the likeness of the scenic spots, more popular areas or POIs are often expressed with higher traffic. Similarly, the traffic flow in the scenic spot is also affected by the heat of the scenic spot. Popularity similarity between two scenic spots
Figure BDA0002709959740000085
And forming a popularity similarity matrix P for the ratio of the number of the few comments to the number of the many comments, and the popularity similarities between every two scenic spots.
Preferably, the adjacency matrix F of the scenic spot functional similarity feature map is obtained by equation 2:
Figure BDA0002709959740000086
wherein,
Figure BDA0002709959740000087
and
Figure BDA0002709959740000088
are each ViAnd VjA vector consisting of the number of POIs of each category, r represents the total number of categories of POIs, k represents the kth category,
Figure BDA0002709959740000089
represents ViThe number of POI class k in the image,
Figure BDA00027099597400000810
represents VjThe number of POI class k in the image,
Figure BDA00027099597400000811
r and k are both positive integers.
In the scenic spot functional similarity feature map, scenic spots are taken as nodes of the map, a functional similarity matrix F between the scenic spots is taken as an adjacent matrix of the map, the functionality of the scenic spots is represented by the quantity distribution of various POI in the area where the scenic spots are located, and the functional similarity between different scenic spots is measured by cosine similarity.
Figure BDA00027099597400000812
And expressing the cosine similarity between the two vectors, taking the cosine similarity as the functional similarity of the two scenic spots, and forming a functional similarity matrix F by the functional similarity between every two scenic spots.
Preferably, the adjacency matrix L of the scenic distance feature map is obtained by equation 3:
Figure BDA0002709959740000091
wherein, dist (V)i,Vj) Represents ViAnd VjAnd a distance therebetween
Figure BDA0002709959740000092
Figure BDA0002709959740000093
R is the radius of the earth,
Figure BDA0002709959740000094
and
Figure BDA0002709959740000095
are each ViAnd VjThe longitude of (a) is determined,
Figure BDA0002709959740000096
and
Figure BDA0002709959740000097
are each ViAnd VjAnd (3) the latitude of (a), maxmin represents the maximum and minimum normalization, i and j belong to n, and n is a positive integer.
In the scenic spot distance feature map, the scenic spots are taken as nodes of the map, the distance feature matrix L between the scenic spots is taken as an adjacency matrix of the map,
Figure BDA0002709959740000098
the closer its value is to 1, the closer the distance between the two scenic spots.
Preferably, the adjacency matrix T of the scenic spot traffic accessibility feature map is obtained by equation 4:
Figure BDA0002709959740000099
wherein conjenienttransfer represents ViAnd VjOn the same road segment.
In the scenic spot traffic accessibility feature map, scenic spots are taken as nodes of the map, a traffic accessibility matrix T is taken as an adjacent matrix of the map, ViAnd VjIf the traffic reaches the same road section, the degree of reach is set to be 1, otherwise, the degree of reach is set to be 0, and then the degree of reach between every two scenic spots forms a traffic degree of reach matrix T.
Specifically, in step 2, the historical traffic flow of the scenic spot is the traffic flow of the scenic spot in a certain range (generally, a prediction area) in each scenic spot, and the number of taxis in a period of time is used as the traffic flow of the scenic spot in the period of time. The scenic spots are distributed differently, and the traffic flow of the regions in which the scenic spots are located is also different. The scenic spot traffic flow data may be viewed as a time-varying signal, and the spatial location of the scenic spot is also important information to consider in order to predict the traffic flow of the scenic spot. Thus, the flow values and positions are coupled into a spatiotemporal traffic flow matrix as input to the model.
Specifically, the historical traffic flow matrix of the scenic spot is obtained by adopting a formula 5
Figure BDA00027099597400000910
Wherein the rows of the matrix represent time intervals, the columns of the matrix represent scenic spots, the elements in the matrix represent flow values of the scenic spots in the time period, xmnThe traffic flow value of the nth scenic spot at the mth time interval is shown, the number of the mth time intervals is shown, and n is shown as the number of the scenic spots.
Specifically, in the present embodiment, the time interval is one hour.
Specifically, the graph-rolling network (GCN) can capture a plurality of spatial correlations such as location proximity, traffic accessibility, popularity similarity, functional similarity and the like between each scenic spot and other scenic spots so as to represent the mutual influence of traffic between the scenic spots.
Specifically, in this embodiment, the Graph Convolution Network (GCN) is two layers, the output of the first layer graph convolution network is used as the input of the second layer graph convolution network, and the output of the second layer graph convolution network is the extracted depth features L of 4 different scenic region feature maps*、T*、P*And F*。
Specifically, the multi-feature fusion matrix is obtained by equation 5
H*＝λ1×L*+λ2×T*+λ3×P*+λ4×F*Formula 5
Wherein L is*、T*、P*And F*Respectively represents the popularity depth characteristic of the scenic spot, the function similarity depth characteristic of the area where the scenic spot is located, the distance depth characteristic of the scenic spot and the traffic access depth characteristic of the scenic spot, and lambda1、λ2、λ3And λ4Is a weight matrix in which the value range of the elements [0,1 ]]And continuously updating in a self-adaptive manner in the training process.
In particular, the cyclic gate unit (GRU)) For mining time characteristics in historical traffic flow data, the GRU inputs H in the current time interval in each iteration process*The output of GRU is the traffic flow prediction value H of the current time intervaltObtained using formula 5:
H*＝λ1×L*+λ2×T*+λ3×P*+λ4×F*formula 5
Wherein u ist＝σ(Wu*H*+Wu*Ht-1+bu)，ct＝tanh(Wc*(rt*Ht-1)+Wc*H*+bc)，rt＝σ(Wr*H*+Wr*Ht-1+br)，utAnd rtUpdate gate and reset gate representing the current time interval, respectively, ctThe state of the cell, H, representing the current time intervalt-1Represents the output, W, of the last time interval t-1u、WrAnd WcWeight representing parameter, bu、brAnd bcRespectively, the offset in the parameters, the symbol x the convolution operation, and σ and tanh the activation function.
Preferably, the loss function in step 3 training is:
Figure BDA0002709959740000111
wherein, YtRepresenting the actual traffic flow during the t period,
Figure BDA0002709959740000112
indicating predicted traffic flow, L, over a period of tregIs the L2 regularization term, μ is to [0,1 ]]Is determined.
The advantages of choosing this loss function are: because the traffic flow of scenic spots is influenced by a plurality of factors, the model represents the factors by constructing a plurality of characteristic graphs, however, the more the characteristics are, the more the overfitting phenomenon is easily generated, and therefore, the regularization term is added into the loss function to relieve the phenomenon; in addition, the loss function can be more easily converged to a relatively stable value by adding the regularization term, and unnecessary iteration times caused by oscillation are reduced.
In the actual construction of the model, the invention adjusts the relevant parameters of GRUs in the model by a grid search method, and performs test evaluation on a test set to determine the optimal relevant parameters. The number of GRU layers was searched in the range (1, 2, 3, 4, 5), and as a result, as shown in fig. 4, it was determined that the number of GRU layers of this model was 3. For the time length of the historical traffic flow of the batch input, a search is carried out in (6, 12, 18, 24, 30, 36), and the result is shown in fig. 5, and the time length of the historical flow input of the model is determined to be 24. For the number of hidden units of the GRU in the model, searching is performed in the range of (16, 32, 64, 100, 128), and finally, the model selects the 64 hidden units with the best effect, as shown in fig. 6.
The embodiment also discloses a scenic spot traffic volume prediction method based on the multi-graph convolution network, which comprises the following steps:
step a: acquiring historical traffic flow of a target scene and multi-source heterogeneous data of the target scene;
step b: extracting features from multi-source heterogeneous data of a target scenic spot to obtain a scenic spot popularity feature map of the target scenic spot, a scenic spot function similarity feature map of the target scenic spot, a scenic spot distance feature map of the target scenic spot and a scenic spot traffic accessibility feature map of the target scenic spot; extracting characteristics of historical traffic flow of a target scenic spot to obtain a historical traffic flow matrix of the target scenic spot;
step c: inputting a scenic spot popularity characteristic diagram of a target scenic spot, a scenic spot function similarity characteristic diagram of the target scenic spot, a scenic spot distance characteristic diagram of the target scenic spot, a scenic spot traffic accessibility characteristic diagram of the target scenic spot and a historical traffic flow matrix of the target scenic spot into a scenic spot predicted traffic flow, and outputting the predicted traffic flow of the target scenic spot.
Example 1
In the embodiment, traffic flow prediction is performed on four famous tourist attractions in the west city, including a west security art gallery, a big goose tower, a large Ci Ensi and a large Tang lotus garden, wherein the scenic region comment data are obtained by using the four famous tourist attractions which are obtained from the American college, the scenic region POI data are POI data in prediction ranges of the four famous tourist attractions obtained by crawling from an Baidu map open platform, the geographical coordinates of the scenic region are obtained by obtaining the geographical coordinate data of the four famous tourist attractions from the Baidu map open platform, and the road network data are the road network data of the city in the west city of the west city obtained from OpenStreetMap. In the embodiment, the traffic flow data of the scenic spot is from a GPS track data set of a taxi in the city of Western-Ann provided by a traffic office in the city of Western-Ann, and the sampling interval of GPS points is between 2S and 120S.
Compared with other traffic flow prediction models, the multi-graph convolution recurrent neural network model has higher prediction accuracy in scenic spot traffic flow prediction. As shown in Table 1, the accuracy of the multi-graph convolutional recurrent neural network was improved by 35.81%, 9.12%, 1.57%, 0.71% and 1.73% respectively when the prediction duration was 1 hour compared to the baseline methods ARMIA, SVR, GRU, DCRNN and T-GCN. The model uses 4 different feature maps, considers rich domain knowledge, further considers the functional similarity between regions and the popularity similarity between scenic spots on the basis of the connectivity between the regional features and the scenic spots, and captures important features influencing traffic variation. The multi-graph convolution recurrent neural network model has stronger robustness. As shown in fig. 7, no matter how the predicted duration changes, the multi-graph convolution cyclic neural network model can obtain the best predicted performance through training, and the variation trend of the predicted result is relatively stable, which indicates that the model is relatively insensitive to the predicted duration. Therefore, the multi-graph convolution recurrent neural network provided by the scheme can be used for short-term traffic flow prediction and is more suitable for long-term traffic flow prediction. The multi-graph convolution cyclic neural network model also has great advantages in the time cost of training the model. As shown in fig. 8, compared with the T-GCN model which requires 5000 iterations to converge, the multi-graph convolution cyclic neural network model incorporates multiple feature graphs, and only 500 iterations are required to obtain good results, so that the time cost for training the model is greatly reduced.
TABLE 1 traffic flow prediction Performance indicators for different methods at different prediction durations
Figure BDA0002709959740000131
Figure BDA0002709959740000141
In order to further analyze the effect of the spatial correlation modeling, the scheme evaluates the influence of a single characteristic map in the multi-map convolution cyclic neural network through an ablation experiment, wherein the influence comprises a scenic spot bitmap, a scenic spot traffic accessibility map, a scenic spot popularity similarity map and a regional function similarity map of the scenic spot. As shown in fig. 9, each of the profiles contributes to the traffic flow prediction task to a certain extent as compared to the case where no profile is used. In the characteristic diagrams, the scenic spot traffic accessibility characteristic diagram has the most obvious contribution to traffic prediction, and the scenic spot functional similarity diagram has relatively small contribution to two prediction tasks of traffic, so that the important innovation significance of introducing various characteristic diagrams in the multi-diagram convolution cyclic neural network is verified.

Claims (2)

1. A scenic spot traffic prediction model building method based on a multi-graph convolution network is characterized by comprising the following steps:
step 1: acquiring multi-source heterogeneous data of a scenic spot, extracting features of the multi-source heterogeneous data of the scenic spot, and acquiring a scenic spot popularity feature map, a scenic spot function similarity feature map, a scenic spot distance feature map and a scenic spot traffic accessibility feature map;
step 2: obtaining historical traffic flow of a scenic spot, extracting characteristics of the historical traffic flow of the scenic spot, and obtaining a historical traffic flow matrix of the scenic spot;
and step 3: establishing a multi-map convolution recurrent neural network model, taking a historical traffic flow matrix, a scenic spot popularity characteristic diagram, a scenic spot functional similarity characteristic diagram, a scenic spot distance characteristic diagram and a scenic spot traffic accessibility characteristic diagram as inputs, taking scenic spot predicted traffic flow as an output, training the model, and taking the trained model as a scenic spot traffic flow prediction model;
the multi-map convolution recurrent neural network model comprises a map convolution network and a recurrent gate control unit, wherein the map convolution network is used for outputting the depth characteristics of a scenic spot characteristic map according to a historical traffic flow matrix, a scenic spot popularity characteristic map, a scenic spot function similarity characteristic map, a scenic spot distance characteristic map and a scenic spot traffic accessibility characteristic map and establishing a multi-characteristic fusion matrix according to the depth characteristics of the scenic spot characteristic map, and the recurrent gate control unit is used for outputting a scenic spot predicted traffic flow according to the multi-characteristic fusion matrix;
the adjacent matrix P of the scenic spot popularity characteristic map is obtained by the following formula 1:
Figure FDA0003394841010000011
wherein, ViAnd VjFor any ith and jth scenic spot in the set of scenic spots,
Figure FDA0003394841010000012
and
Figure FDA0003394841010000013
are each ViAnd VjThe number of user comments (a) to be made,
Figure FDA0003394841010000014
is a ViAnd VjThe degree of similarity in popularity of (a) a,
Figure FDA0003394841010000015
has a value range of [0,1 ]]I and j are positive integers;
the adjacent matrix F of the scenic spot functional similarity feature map is obtained by the following formula 2:
Figure FDA0003394841010000021
wherein, ViAnd VjFor any ith and jth scenic spot in the set of scenic spots,
Figure FDA0003394841010000022
and
Figure FDA0003394841010000023
are each ViAnd VjA vector consisting of the number of POIs of each category, r represents the total number of categories of POIs, k represents the kth category,
Figure FDA0003394841010000024
represents ViThe number of POI class k in the image,
Figure FDA0003394841010000025
represents VjThe number of POI class k in the image,
Figure FDA0003394841010000026
i. j, r and k are positive integers;
the adjacent matrix L of the scenic spot distance feature map is obtained by equation 3:
Figure FDA0003394841010000027
wherein, ViAnd VjFor any ith and jth scenic spot in the set of scenic spots, dist (V)i,Vj) Represents ViAnd VjAnd a distance therebetween
Figure FDA0003394841010000028
Figure FDA0003394841010000029
R is the radius of the earth,
Figure FDA00033948410100000210
and
Figure FDA00033948410100000211
are each ViAnd VjThe longitude of (a) is determined,
Figure FDA00033948410100000212
and
Figure FDA00033948410100000213
are each ViAnd VjThe latitude of the system is represented by maxmin, the maximum and minimum normalization is represented by n, the number of scenic spots is represented by n, i and j belong to n, and i, j and n are positive integers;
an adjacent matrix T of the scenic spot traffic accessibility feature map is obtained by equation 4:
Figure FDA00033948410100000214
wherein, Ti,jRepresenting the traffic access degrees of any ith scenic spot and jth scenic spot, and the conditional transmission representing that the ith scenic spot and the jth scenic spot are on the same road segment;
obtaining the historical traffic flow matrix of the scenic spot by adopting a formula 5
Figure FDA0003394841010000031
Wherein x ismnThe traffic flow value of the nth scenic spot at the mth time interval is shown, m represents the number of the time intervals, and n represents the number of the scenic spots;
the loss function during the training of the step 3 is as follows:
Figure FDA0003394841010000032
wherein, YtRepresenting the actual traffic flow during the t period,
Figure FDA0003394841010000033
indicating predicted traffic flow, L, over a period of tregIs the L2 regularization term, μ is to [0,1 ]]Is determined.
2. A scenic spot traffic volume prediction method based on a multi-graph convolution network is characterized by comprising the following steps:
step a: acquiring historical traffic flow of a target scene and multi-source heterogeneous data of the target scene;
step b: extracting features from multi-source heterogeneous data of a target scenic spot to obtain a scenic spot popularity feature map of the target scenic spot, a scenic spot function similarity feature map of the target scenic spot, a scenic spot distance feature map of the target scenic spot and a scenic spot traffic accessibility feature map of the target scenic spot; extracting characteristics of historical traffic flow of a target scenic spot to obtain a historical traffic flow matrix of the target scenic spot;
step c: inputting a scenic spot popularity characteristic diagram of a target scenic spot, a scenic spot functional similarity characteristic diagram of the target scenic spot, a scenic spot distance characteristic diagram of the target scenic spot, a scenic spot traffic accessibility characteristic diagram of the target scenic spot and a historical traffic flow matrix of the target scenic spot into a scenic spot traffic prediction model to predict traffic flow, wherein the scenic spot traffic prediction model is obtained by adopting the scenic spot traffic prediction model establishing method based on the multi-map convolution network according to claim 1, and outputting the predicted traffic flow of the target scenic spot;
the scenic spot multi-source heterogeneous data comprises scenic spot comment data, scenic spot POI data, geographical coordinates of scenic spots and road network data.
CN202011052408.5A 2020-09-29 2020-09-29 Scenic spot traffic volume prediction model establishing and predicting method based on multi-graph convolution network Active CN112365708B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011052408.5A CN112365708B (en) 2020-09-29 2020-09-29 Scenic spot traffic volume prediction model establishing and predicting method based on multi-graph convolution network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011052408.5A CN112365708B (en) 2020-09-29 2020-09-29 Scenic spot traffic volume prediction model establishing and predicting method based on multi-graph convolution network

Publications (2)

Publication Number Publication Date
CN112365708A CN112365708A (en) 2021-02-12
CN112365708B true CN112365708B (en) 2022-04-19

Family

ID=74507609

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011052408.5A Active CN112365708B (en) 2020-09-29 2020-09-29 Scenic spot traffic volume prediction model establishing and predicting method based on multi-graph convolution network

Country Status (1)

Country Link
CN (1) CN112365708B (en)

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112766240B (en) * 2021-03-18 2022-08-12 福州大学 Residual multi-graph convolution crowd distribution prediction method and system based on space-time relationship
CN113240904B (en) * 2021-05-08 2022-06-14 福州大学 Traffic flow prediction method based on feature fusion
CN113505536A (en) * 2021-07-09 2021-10-15 兰州理工大学 Optimized traffic flow prediction model based on space-time diagram convolution network
CN115331425B (en) * 2022-06-30 2023-12-19 银江技术股份有限公司 Traffic early warning method, device and system
CN115641718B (en) * 2022-10-24 2023-12-08 重庆邮电大学 Short-time traffic flow prediction method based on bayonet flow similarity and semantic association

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109389244A (en) * 2018-09-06 2019-02-26 浙江鸿程计算机系统有限公司 Tourist's number prediction technique in a kind of short-term scenic spot of multifactor perception based on GRU
CN109993668A (en) * 2019-04-09 2019-07-09 桂林电子科技大学 A kind of recommending scenery spot method based on gating cycle unit neural network
CN110245285A (en) * 2019-04-30 2019-09-17 中国科学院信息工程研究所 A kind of personalized recommendation method based on Heterogeneous Information network
CN110969854A (en) * 2019-12-13 2020-04-07 深圳先进技术研究院 Traffic flow prediction method, system and terminal equipment
CN111126704A (en) * 2019-12-25 2020-05-08 河海大学常州校区 Multi-region precipitation prediction model construction method based on multi-graph convolution and memory network
CN111274493A (en) * 2020-01-17 2020-06-12 电子科技大学 Grading prediction method based on multi-source user comments
CN111489013A (en) * 2020-03-11 2020-08-04 北京交通大学 Traffic station flow prediction method based on space-time multi-graph convolution network
CN111612489A (en) * 2019-02-25 2020-09-01 北京嘀嘀无限科技发展有限公司 Order quantity prediction method and device and electronic equipment
CN111612122A (en) * 2019-02-25 2020-09-01 北京嘀嘀无限科技发展有限公司 Method and device for predicting real-time demand and electronic equipment

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109389244A (en) * 2018-09-06 2019-02-26 浙江鸿程计算机系统有限公司 Tourist's number prediction technique in a kind of short-term scenic spot of multifactor perception based on GRU
CN111612489A (en) * 2019-02-25 2020-09-01 北京嘀嘀无限科技发展有限公司 Order quantity prediction method and device and electronic equipment
CN111612122A (en) * 2019-02-25 2020-09-01 北京嘀嘀无限科技发展有限公司 Method and device for predicting real-time demand and electronic equipment
CN109993668A (en) * 2019-04-09 2019-07-09 桂林电子科技大学 A kind of recommending scenery spot method based on gating cycle unit neural network
CN110245285A (en) * 2019-04-30 2019-09-17 中国科学院信息工程研究所 A kind of personalized recommendation method based on Heterogeneous Information network
CN110969854A (en) * 2019-12-13 2020-04-07 深圳先进技术研究院 Traffic flow prediction method, system and terminal equipment
CN111126704A (en) * 2019-12-25 2020-05-08 河海大学常州校区 Multi-region precipitation prediction model construction method based on multi-graph convolution and memory network
CN111274493A (en) * 2020-01-17 2020-06-12 电子科技大学 Grading prediction method based on multi-source user comments
CN111489013A (en) * 2020-03-11 2020-08-04 北京交通大学 Traffic station flow prediction method based on space-time multi-graph convolution network

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
基于多特征时空图卷积网络的水运通航密度预测;董伟 等;《物联网学报》;20200930;第4卷(第3期);第78-85页 *
基于时空多图卷积网络的交通站点流量预测;荣斌 等;《计算机工程》;20200531;第46卷(第5期);第26-33页 *
多源异构数据融合的城市私家车流量预测研究;刘晨曦 等;《通信学报》;20210331;第42卷(第3期);第54-64页 *

Also Published As

Publication number Publication date
CN112365708A (en) 2021-02-12

Similar Documents

Publication Publication Date Title
CN112365708B (en) Scenic spot traffic volume prediction model establishing and predicting method based on multi-graph convolution network
CN111080029B (en) Urban traffic road speed prediction method and system based on multi-path segment space-time correlation
CN110598953A (en) Space-time correlation air quality prediction method
CN104462190B (en) A kind of online position predicting method excavated based on magnanimity space tracking
CN108629978A (en) A kind of traffic trajectory predictions method based on higher-dimension road network and Recognition with Recurrent Neural Network
CN115578852B (en) DSTGCN-based traffic prediction method
Ou et al. Learn, assign, and search: real-time estimation of dynamic origin-destination flows using machine learning algorithms
Xu et al. HighAir: A hierarchical graph neural network-based air quality forecasting method
CN114944053A (en) Traffic flow prediction method based on spatio-temporal hypergraph neural network
CN115545758B (en) Method and system for self-adaptive incremental site selection of urban service facilities
CN112419711B (en) Closed parking lot parking demand prediction method based on improved GMDH algorithm
CN112884014A (en) Traffic speed short-time prediction method based on road section topological structure classification
CN116486624A (en) Traffic flow prediction method and system based on space-time diagram convolutional neural network
CN117271899A (en) Interest point recommendation method based on space-time perception
CN117037483A (en) Traffic flow prediction method based on multi-head attention mechanism
CN113887704A (en) Traffic information prediction method, device, equipment and storage medium
Yang et al. Feature selection in conditional random fields for map matching of GPS trajectories
Wang et al. Reconstruction of missing trajectory data: a deep learning approach
CN115293427A (en) Planning method and system for optimal travel route of city early peak
Chu et al. Simulating human mobility with a trajectory generation framework based on diffusion model
CN110263250A (en) A kind of generation method and device of recommended models
CN118133084A (en) Method and device for predicting air quality of area without monitoring station based on hierarchical graph convolutional network
CN117275222A (en) Traffic flow prediction method integrating one-dimensional convolution and attribute enhancement units
Chen et al. Deep citywide multisource data fusion-based air quality estimation
Xiao et al. Streaming Traffic Flow Prediction Based on Continuous Reinforcement Learning

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