Urban traffic flow prediction method, system and device based on generative adversarial network

Abstract

The application discloses a kind of urban traffic flow prediction method, system and equipment based on generative adversarial network, it is related to urban traffic flow prediction field, the method includes: according to the travel demand data of each traffic flow area in the current stage research city, using Node2Vec graph embedding algorithm, determines the embedding matrix of the current stage research city;The embedding matrix of the current stage research city and the random noise conforming to Gaussian distribution are input into the trained generator as input value, to predict the traffic flow of future stage research city.The application can capture the basic mode that how traffic flow evolves with the change of travel demand, so as to realize accurate prediction of traffic flow.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of urban traffic flow prediction, and particularly to a method, system and device for urban traffic flow prediction based on a generative adversarial network. BACKGROUND

[0002] With the continuous advancement of urbanization, intelligent transportation systems (ITS) play an important role in facilitating urban travel and promoting economic and cultural development in cities. Accurate traffic flow prediction is crucial for the development of ITS, and can provide valuable insights for traffic planning, resident travel and operational decision-making of transportation departments. Therefore, more and more traffic flow prediction models have been proposed by researchers.

[0003] There are two main traffic flow prediction methods initially: statistical models and machine learning-based models. The former includes methods such as historical average (HA) model, autoregressive integrated moving average (ARIMA) model and its various variants. The latter includes methods such as K-Nearest Neighbors (KNN), Random Forest and Support Vector Machine (SVM) model.

[0004] In recent years, the rapid development of deep learning has shown significant advantages in solving complex nonlinear regression problems. Deep learning models can effectively solve the inherent defects of multi-module models (achieving local optimal performance does not guarantee global optimal performance) by using end-to-end learning methods. Currently, traffic flow prediction models are mainly based on deep learning techniques, such as Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Graph Convolutional Network (GCN), Graph Attention Network (GAT), Generative Adversarial Network (GAN), etc.

[0005] Some researchers have proposed using deep learning methods to address the challenges in traffic flow prediction. For example, CNN is used to capture spatial dependencies and RNN to capture temporal correlations. Zhang et al. proposed ST-ResNet to predict crowd flow. Convolutional operations are commonly used to process Euclidean structure data. However, traffic data is non-Euclidean structure. Therefore, traditional convolutional operations may not effectively capture the complex spatial relationships and patterns inherent in traffic data. To address this challenge, researchers have developed GCN. Some studies use GCN to extract spatial and temporal features in traffic prediction. Zhang et al. proposed STGCN to solve the time series prediction problem in the field of traffic. Lu et al. proposed a spatio-temporal adaptive gated convolutional network to predict traffic conditions after several time steps. Peng et al. designed a dynamic graph recurrent convolutional network for urban traffic flow prediction. Zhang et al. proposed a dynamic graph convolutional network based on spatio-temporal data embedding to more accurately and stably predict traffic flow. Lv et al. proposed T-MGCN, which uses a multi-graph convolutional network and a recurrent neural network for traffic flow prediction. Wang et al. introduced a graph convolutional neural network based on hypergraph structure to capture the spatio-temporal features of traffic data, while Li et al. combined a gating mechanism with a graph convolutional neural network for traffic prediction. Ni et al. proposed a one-dimensional convolutional neural network and a channel attention mechanism for extracting temporal features, and a multi-graph convolutional network framework and a gating mechanism for capturing spatial features. However, traffic patterns have inherent complexity and dynamic variability, and therefore cannot obtain the inherent spatio-temporal dependency features between traffic flow regions based on certain deterministic states. The above models using GCN for traffic prediction ignore the changes in traffic patterns in historical data. SUMMARY

[0006] The purpose of the present application is to provide a city traffic flow prediction method, system and device based on a generative adversarial network, which can capture the basic pattern of how traffic flow evolves with changes in travel demand, thereby achieving accurate prediction of traffic flow.

[0007] To achieve the above purpose, the present application provides the following solutions:

[0008] In a first aspect, the present application provides a city traffic flow prediction method based on a generative adversarial network, comprising:

[0009] Obtaining travel demand data of each traffic flow region in the city under study at the current stage;

[0010] According to the travel demand data of each traffic flow region in the city under study at the current stage, using a Node2Vec graph embedding algorithm to determine the embedding matrix of the city under study at the current stage;

[0011] The embedding matrix of the current stage research city and random noise conforming to a Gaussian distribution are input into the trained generator as input values to predict the traffic flow of the future stage research city.

[0012] In a second aspect, the present application provides a city traffic flow prediction system based on a generative adversarial network, comprising:

[0013] A travel demand data acquisition module is configured to acquire travel demand data of each traffic flow region in the current stage research city.

[0014] An embedding matrix calculation module is configured to determine the embedding matrix of the current stage research city by using a Node2Vec graph embedding algorithm based on the travel demand data of each traffic flow region in the current stage research city.

[0015] A traffic flow prediction module is configured to input the embedding matrix of the current stage research city and random noise conforming to a Gaussian distribution as input values into the trained generator to predict the traffic flow of the future stage research city.

[0016] In a third aspect, the present application provides an electronic device comprising a memory and a processor, wherein the memory is configured to store a computer program, and the processor is configured to run the computer program to enable the electronic device to perform the city traffic flow prediction method based on the generative adversarial network according to the first aspect.

[0017] According to the embodiments of the present application, the following technical effects are achieved:

[0018] The present application introduces a method for constructing a traffic embedding graph based on a Node2Vec graph embedding algorithm, converts the node relationship of a graph into a two-dimensional embedding matrix to make the traffic data suitable for processing by a convolutional neural network, and specifically comprises: using a convolutional neural network based on the Node2Vec graph embedding algorithm to capture deep spatiotemporal dependencies, and using a generative adversarial network framework to capture the basic mode of how traffic flow evolves with changes in travel demand, thereby achieving accurate prediction of traffic flow. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0020] Figure 1 A flowchart of a city traffic flow prediction method based on a generative adversarial network according to an embodiment of the present application is shown in the figure.

[0021] Figure 2 A schematic diagram of the structure of the urban traffic flow prediction model (TGAN) based on generative adversarial networks provided in an embodiment of the present invention; Figure 2 (a) is a schematic diagram of the generator provided in an embodiment of the present invention; Figure 2 (b) is a schematic diagram of the discriminator provided in an embodiment of the present invention;

[0022] Figure 3 Heatmaps showing traffic flow distributions generated by different models and actual traffic flow distributions provided in embodiments of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Example 1

[0026] With the rapid development of intelligent transportation systems, accurate traffic flow prediction has become a crucial aspect of traffic planning and management. However, most existing methods neglect to capture changes in traffic patterns within historical data. To address this issue, this embodiment presents a deep learning model that combines a convolutional neural network based on the Node2Vec graph embedding algorithm with a generative adversarial network framework. This ingenious combination effectively captures deep spatiotemporal dependencies and fundamental patterns in traffic flow evolution driven by travel demand. Experimental results on two real-world traffic datasets demonstrate that the proposed model outperforms other baseline models, showcasing its potential as a tool for traffic management and urban planning.

[0027] like Figure 1 As shown in the figure, this embodiment provides a method for predicting urban traffic flow based on generative adversarial networks, including:

[0028] Step 100: Obtain travel demand data for each traffic flow area in the city under study in the current stage.

[0029] Step 200: Based on the travel demand data of each traffic flow area in the city under study in the current stage, the Node2Vec graph embedding algorithm is used to determine the embedding matrix of the city under study in the current stage.

[0030] Step 300: input the embedding matrix of the current stage research city and the random noise conforming to the Gaussian distribution as input values into the trained generator to predict the traffic flow of the future stage research city.

[0031] In this embodiment, step 200 specifically includes:

[0032] First, according to the travel demand data of each traffic flow area in the current stage research city, a current stage research city traffic graph is constructed; second, according to the current stage research city traffic graph and the Node2Vec graph embedding algorithm, the embedding matrix of the current stage research city is determined.

[0033] The detailed process is as follows: the traffic flow network is defined as a graph G(V, E, A), where V represents the node set, which actually represents the traffic flow areas obtained after the research city is divided; E represents the edge set, which actually represents the travel demand between traffic flow areas; A represents the adjacency matrix, which is actually the weight of the connection degree between traffic flow areas. Let Xi(t) represents the feature value of node i at time interval t, which actually represents the flow value of the ith traffic flow area at time interval t; Xi(t) represents the feature value of node i at time interval t, which actually represents the flow value of the ith traffic flow area at time interval t; t t-P+1 t-P+2 t , …, X N×P , …, X t , …, X s , …, X N×N represent the feature values of all nodes in the past P time intervals, which actually represent the flow values of the research city in the past P time intervals. Traffic flow prediction is to predict future traffic flow according to historical traffic data, and the specific representation is as shown in formula (1), where F() can represent some deep neural network model, and Y represents the model output result.

[0034] Y = F(H t ) (1).

[0035] A s ∈ R N×N is the embedding matrix of graph G obtained according to the Node2Vec algorithm, and N represents the number of nodes in the node set. Node2Vec is a graph embedding algorithm, which aims to learn a mapping that maps nodes from a high-dimensional space to a low-dimensional space while maximizing the feature representation of neighboring nodes. Node2Vec walk is a biased random walk that combines depth-first sampling and breadth-first sampling strategies. Breadth-first sampling requires that the sampled node is a direct neighbor of the source node, while depth-first sampling is a continuous sampling method that gradually increases the distance of the sampled node to the source node.

[0036] Table 1 Node2Vec graph embedding algorithm

[0037]

[0038]

[0039] In the present embodiment, the trained generator is a generator in a city traffic flow prediction model based on a generative adversarial network (TGAN), the structure of which is as shown in Figure 2

[0040] The determination process of the city traffic flow prediction model based on the generative adversarial network is as follows:

[0041] (1) Construct a generative adversarial network framework; the generative adversarial network framework comprises a generator and a discriminator.

[0042] (2) Construct sample data; the sample data comprises an embedding matrix of a city in a first historical stage and real traffic flow of the city in a second historical stage; the second historical stage is a future stage of the first historical stage.

[0043] (3) Input the embedding matrix of the city in the first historical stage and random noise conforming to a Gaussian distribution as input values into the generator in the generative adversarial network framework, to obtain predicted traffic flow of the city in the second historical stage.

[0044] (4) Input the real traffic flow of the city in the second historical stage and the predicted traffic flow of the city in the second historical stage as input values into the discriminator in the generative adversarial network framework, to obtain a discrimination result, and adjust the network parameters of the generator in the generative adversarial network framework according to the discrimination result, until the loss values of the generator and the discriminator meet the set requirements, and thus obtain the city traffic flow prediction model based on the generative adversarial network.

[0045] The generator is composed of four convolutional layers based on the Node2Vec graph embedding algorithm (Node2Vec-convolutional layers) and one fully connected layer, while the discriminator is composed of two Node2Vec-convolutional layers and one fully connected layer.

[0046] Further, the objective of the Node2Vec-convolutional layer is to learn a function aiming to extract spatiotemporal features. The representation of the function is as follows:

[0047] H l+1 =f(H l ,A S )=ReLU(A S H l W l ) (2)。

[0048] A​S = Node2Vec(G(V, E, A)) (3).

[0049] where H l represents the input data of the l-th layer, A S represents the embedding matrix obtained by the Node2Vec graph embedding algorithm, W i represents the learnable parameter matrix of the l-th layer, ReLU() represents the activation function, and H l+1 represents the output of the l-th layer.

[0050] The activation function of the Node2Vec-convolutional layer in the generator is ReLU(), and Batch Normalization is used for layer normalization between layers. The activation function of the Node2Vec-convolutional layer in the discriminator is LeakyReLU(), and Batch Normalization is used for layer normalization between layers.

[0051] The generative adversarial network framework is a kind of generative model, which uses a generator and a discriminator to learn the distribution of training data. In TGAN, the generator takes random noise as input and outputs traffic data, aiming to make the generated data very close to the distribution of real historical traffic data. The discriminator takes real traffic data and traffic data generated by the generator as input and outputs a probability value between 0 and 1, indicating the likelihood that the input data is real data. By optimizing the generator and discriminator in an adversarial manner, TGAN can generate results very close to the distribution of real historical traffic data.

[0052] Specifically, according to the historical travel demand of each traffic flow area in the city, the Node2Vec graph embedding algorithm is used to obtain the embedding matrix A s ∈R N×N , N represents the number of traffic flow areas. The noise data conforming to the Gaussian distribution and the embedding matrix A s are taken as the input of the generator to obtain the predicted traffic flow of the city; wherein the format of the noise data is Noise∈R B×N×1 , B represents the size of the batch training dimension. The predicted traffic flow obtained by the generator and the real traffic flow in the training set are taken as the input of the discriminator. The discriminator outputs: the output is a probability value, which is the probability of judging whether the input data is the predicted traffic flow of the generator or the real traffic flow.

[0053] Table 2 Training process of city traffic flow prediction model based on generative adversarial network

[0054]

[0055] The loss functions of the generator and the discriminator are shown in equations (4) to (6):

[0056] V(D) = E z～P(Z) [log(D(z))] + E M～P(G) [log(1-D(M))] (4).

[0057] V(G) = E M～P(G) [log(1-D(M))] + μMAELoss (5).

[0058]

[0059] wherein, denotes the output of the generator, y i denotes the real traffic flow, and μ is a constant that controls the proportion of MAELoss in the loss function of the generator.

[0060] The above technical solutions are described below through a specific experiment.

[0061] In this experiment, the TGAN model provided in this embodiment is used to predict traffic flow using the data set of Chengdu's online taxi in November 2016. The study of Chengdu is a region with longitude 104.042E to 104.130E and latitude 30.652N to 30.728N, which is divided into 81 traffic flow regions of 1km x 1km. Traffic flow distribution reflects the traffic conditions of the traffic flow region, and in this experiment, the inflow and outflow of traffic flow are used to quantify. Since it is difficult to obtain the total flow of all traffic modes in each traffic flow region or grid, this experiment represents the flow value of the traffic flow region or grid as the inflow and outflow of taxis or online taxis, which has been proven to be effective before. In this experiment, a day is divided into 144 10-minute time intervals, and the total flow of each traffic flow region is represented by the number of arriving and departing vehicles in each time interval, and the data set obtained is in the form of Data ∈ R 4320×81×1 Since this experiment aims to capture the changes in flow patterns in historical data, unlike traditional flow prediction methods, it no longer predicts the traffic flow values of future time intervals based on the traffic flow values of past time intervals, but learns from all historical data. This experiment divides the data set into a training set (70%) and a test set (30%), and the training set is in the form of Data_train ∈ R 3024×81×1 , and the test set is in the form of Data_test ∈ R 1296×81×1 .

[0062] In this experiment, the initial learning rate of the discriminator and the generator is set to 0.001, and is reduced to 0.0001 after 10 rounds. The batch size is set to 156, and the total number of training rounds is 200. The number of neurons in the Node2Vec-convolutional layer of the discriminator is set to 32, while the number of neurons in the generator is set to 64. Adam is selected as the optimizer, and the constant μ that controls the proportion of MAELoss in the generator loss function is set to 0.75.

[0063] In this experiment, three evaluation indicators are used to measure the difference between the generated traffic flow distribution and the real traffic flow distribution, namely the mean absolute error (MAE), the root mean square error (RMSE), and the definition D1, as shown in equations (7) to (9).

[0064]

[0065]

[0066]

[0067] where N represents the total number of samples, y i represents the true value, represents the predicted value output by the model. D1 refers to the Euclidean distance between the generated traffic flow distribution and the real traffic flow distribution.

[0068] In this experiment, the Chengdu online taxi data set is used to compare TGAN with TrafficGAN, CGAN, WGAN, and DCGAN, and the performance comparison results of TGAN and various baseline methods are obtained, as shown in Table 3.

[0069] Table 3 Comparison of baseline models (Chengdu online taxi data set)

[0070]

[0071]

[0072] In order to better understand the prediction performance of TGAN, the true value and the prediction result are visualized, as shown in Figure 3 From the heat map of the generated traffic flow distribution and the real traffic flow distribution, it can be seen that TGAN has the highest fitting degree with the true value, and can accurately capture the traffic flow area with dense traffic flow. This shows that TGAN can effectively learn the distribution of the original data and generate accurate data according to the learned distribution.

[0073] The application proposes a deep learning model TGAN which combines a convolutional neural network based on a Node2Vec graph embedding algorithm and a generative adversarial network framework to solve the limitation of ignoring capturing traffic pattern changes in historical data in most existing prediction methods. The experimental results on the Chengdu online taxi dataset show that the MAE, RMSE and D1 of TGAN are the minimum values among the proposed baseline models, and the MAE, RMSE and D1 of TGAN are reduced by 37.15%, 31.81% and 31.80% compared with CGAN, respectively; the MAE, RMSE and D1 of TGAN are reduced by 73.51%, 57.46% and 57.45% compared with WGAN, respectively; the MAE, RMSE and D1 of TGAN are reduced by 41.53%, 36.51% and 36.49% compared with DCGAN, respectively; and the MAE, RMSE and D1 of TGAN are reduced by 41.60%, 36.17% and 36.17% compared with TrafficGAN, respectively.

[0074] DCGAN, WGAN and CGAN use a traditional convolutional neural network to extract spatio-temporal features from traffic data, while TrafficGAN uses a graph convolutional neural network to extract spatio-temporal features. TGAN uses a Node2Vec convolutional layer to capture deep spatio-temporal dependencies. Based on the comparison of evaluation indicators and the visualization of the heat map between TGAN and DCGAN, WGAN, CGAN and TrafficGAN, it can be concluded that TGAN introduces the Node2Vec graph embedding algorithm to construct a traffic embedding graph, so that the convolutional neural network can better process traffic data, and introduces the generative adversarial network framework to capture changes in traffic patterns in historical data, thereby improving prediction accuracy and providing a valuable tool for traffic management and urban planning.

[0075] Embodiment two

[0076] In order to perform the method corresponding to the above-mentioned embodiment one to realize the corresponding functions and technical effects, a city traffic flow prediction system based on a generative adversarial network is provided below.

[0077] The city traffic flow prediction system based on the generative adversarial network provided in this embodiment comprises:

[0078] The travel demand data acquisition module is configured to acquire travel demand data of each traffic flow area in the city under study in the current stage.

[0079] The embedding matrix calculation module is configured to determine an embedding matrix of the city under study in the current stage by using a Node2Vec graph embedding algorithm according to the travel demand data of each traffic flow area in the city under study in the current stage.

[0080] The traffic flow prediction module takes the embedding matrix of the current stage of the city and the random noise conforming to the Gaussian distribution as input values, inputs them into the trained generator, and predicts the traffic flow of the future stage of the city.

[0081] Embodiment three

[0082] The embodiment of the present application provides an electronic device including a memory for storing a computer program and a processor for running the computer program to enable the electronic device to execute the city traffic flow prediction method based on the generative adversarial network.

[0083] Optionally, the electronic device can be a server.

[0084] In addition, the embodiment of the present application further provides a computer readable storage medium storing a computer program, and the computer program is executed by a processor to implement the city traffic flow prediction method based on the generative adversarial network.

[0085] In the specification, each embodiment is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0086] The principles and implementation manners of the present application are described by using specific examples in the present application. The above embodiment is only used to help understand the method of the present application and its core idea; meanwhile, for the general technical personnel in the art, according to the idea of the present application, the specific implementation manner and application range can be changed. In conclusion, the content of the specification should not be understood as the limitation of the present application.

Claims

1. A method for urban traffic flow prediction based on a generative adversarial network, characterized in that, The method comprises the following steps: obtaining travel demand data of each traffic flow area in a current stage of a city under study; determining an embedding matrix of the city under study in the current stage according to the travel demand data of each traffic flow area in the city under study in the current stage, and using a Node2Vec graph embedding algorithm; inputting the embedding matrix of the city under study in the current stage and random noise conforming to a Gaussian distribution as input values into a trained generator to predict traffic flow of the city under study in a future stage; the trained generator is a generator in a city traffic flow prediction model based on a generative adversarial network; a determination process of the city traffic flow prediction model based on the generative adversarial network comprises the following steps: constructing a generative adversarial network framework, wherein the generative adversarial network framework comprises a generator and a discriminator; constructing sample data, wherein the sample data comprises an embedding matrix of a city under study in a first historical stage and real traffic flow of the city under study in a second historical stage; the second historical stage is a future stage of the first historical stage; inputting the embedding matrix of the city under study in the first historical stage and random noise conforming to the Gaussian distribution as input values into the generator in the generative adversarial network framework to obtain predicted traffic flow of the city under study in the second historical stage; inputting the real traffic flow of the city under study in the second historical stage and the predicted traffic flow of the city under study in the second historical stage as input values into the discriminator in the generative adversarial network framework to obtain a discrimination result, and adjusting network parameters of the generator in the generative adversarial network framework according to the discrimination result until loss values of the generator and the discriminator meet a set requirement, so as to obtain the city traffic flow prediction model based on the generative adversarial network. 2.The urban traffic flow prediction method based on a generative adversarial network according to claim 1, characterized in that, The determination of the embedding matrix of the city under study in the current stage according to the travel demand data of each traffic flow area in the city under study in the current stage and using the Node2Vec graph embedding algorithm comprises the following steps: constructing a city traffic graph of the city under study in the current stage according to the travel demand data of each traffic flow area in the city under study in the current stage; determining the embedding matrix of the city under study in the current stage according to the city traffic graph of the city under study in the current stage and the Node2Vec graph embedding algorithm. 3.The urban traffic flow prediction method based on a generative adversarial network according to claim 1, characterized in that, The generator comprises four Node2Vec-convolution layers and one fully connected layer; the Node2Vec-convolution layer is a convolution layer based on the Node2Vec graph embedding algorithm. 4.The urban traffic flow prediction method based on a generative adversarial network according to claim 3, characterized in that, The activation function of the Node2Vec-convolution layer in the generator is ReLU(), and Batch Normalization is used for layer normalization between layers. 5.The urban traffic flow prediction method based on a generative adversarial network according to claim 3, characterized in that, The discriminator comprises two Node2Vec-convolution layers and one fully connected layer. 6.The urban traffic flow prediction method based on a generative adversarial network according to claim 5, characterized in that, The activation function of the Node2Vec-convolution layer in the discriminator is LeakyReLU(), and Batch Normalization is used for layer normalization between layers.

7. A system for urban traffic flow prediction based on generative adversarial networks, operating based on a method for urban traffic flow prediction based on generative adversarial networks according to any one of claims 1-6, characterized in that, The method comprises the following steps: an travel demand data obtaining module is configured to obtain travel demand data of each traffic flow area in a current stage of a city under study; an embedding matrix calculating module is configured to determine an embedding matrix of the city under study in the current stage according to the travel demand data of each traffic flow area in the city under study in the current stage, and using a Node2Vec graph embedding algorithm; The traffic flow prediction module is configured to take the embedding matrix of the current stage of the research city and the random noise conforming to the Gaussian distribution as input values, and input the input values into the trained generator to predict the traffic flow of the future stage of the research city.

8. An electronic device, comprising: The electronic device comprises a memory and a processor, the memory is configured to store a computer program, and the processor is configured to run the computer program to enable the electronic device to perform the method for predicting urban traffic flow based on the generative adversarial network according to any one of claims 1 to 6.

Images