A traffic flow prediction method based on a space-time graph convolution network

Abstract

The application discloses a traffic flow prediction method based on a space-time graph convolution network, comprising the following steps: S1), collecting traffic flow data of a target area by using a road sensor and preprocessing the traffic flow data; S2), collecting additional information and encoding the additional information, wherein the additional information comprises climate data, whether the day is a weekday, and the type and quantity of public places around the node; S3), generating time series data according to the preprocessed traffic flow data obtained in the step S1); S4), constructing two adaptive vectors E S ,E T generating a space correlation matrix as a spatial embedding representation of the node; and S5), constructing a DSRNN network model based on a DSGRU unit, inputting the time series data obtained in the step S3) as input and a label, and inputting the additional information vector obtained in the step S2) into the model, so that the model can better capture the dynamic space dependence of the traffic flow data and achieve better prediction effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation, in particular to the technical field of a traffic flow prediction method based on a space-time graph convolution network.

BACKGROUND TECHNIQUE

[0002] With the process of rapid urbanization, more and more population flows into the city, which leads to the traffic system of the city facing more and more serious challenges. At the same time, thanks to the deployment of a large number of economic and practical traffic sensors, the explosive traffic data pushes us into the era of traffic big data. In order to better solve the mobility challenges faced by the traffic system and better utilize the massive traffic data, the intelligent transportation system (ITS) has become an active research field in recent years, and its development can assist people in utilizing traffic big data to help control and plan urban traffic, and is an indispensable part of building a smart city. Traffic flow prediction has become an important part of this system, which aims to predict future traffic conditions (such as traffic flow, vehicle speed, traffic density) through observed historical data, and thus help to alleviate urban congestion and other problems.

[0003] For example, Figure 1As shown, the traffic flow prediction task can be abstracted as predicting the future traffic flow information of the traffic network given the historical traffic flow information and traffic network information. However, the complex spatial and temporal correlation of traffic flow prediction makes this task extremely challenging. Most existing methods only consider temporal correlation, such as ARIMA model, support vector regression (SVR) and K-nearest neighbor model (KNN). With the advent of the big data era and the rise of deep learning, more and more people try to use deep learning methods to study traffic flow prediction problems. Recent studies mostly model the spatial correlation of traffic data using graph neural networks (GNN), model the temporal correlation using recurrent neural networks (RNN) or convolutional neural networks (CNN), and build spatio-temporal graph convolutional networks for traffic flow prediction. For example, Yu et al. proposed a spatio-temporal graph convolutional network (STGCN) model to model spatial correlation through Chebyshev approximation graph convolution and to model temporal correlation through gated convolution (GLU), which improves the calculation speed of the model (see Yu B, Yin H, Zhu Z. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting [J]. arXiv preprint arXiv:1709.04875, 2017.). Zhao et al. proposed a temporal graph convolutional network (T-GCN) model, which relies on graph convolutional networks to model the spatial correlation of traffic data and uses gated recurrent units to model the temporal correlation, achieving good prediction results (see Zhao L, Song Y, Zhang C, et al. T-gcn: A temporal graph convolutional network for traffic prediction [J]. IEEE Transactions on Intelligent Transportation Systems, 2019, 21(9): 3848-3858.).

[0004] However, the above methods generally have the following problems: 1. Too dependent on traffic network information, which cannot fully describe the spatial information of traffic data, resulting in a model that can only model single spatial correlation, lacks the ability to model hidden spatial correlation, and lacks robustness; 2. Traffic network information changes over time, resulting in a model that can only model static spatial dependence, lacking dynamic spatial dependence; 3. Urban traffic flow prediction does not consider additional information, such as whether it is a weekday, weather conditions, and the distribution of public places (restaurants, cinemas, stations, etc.) near road intersections.

SUMMARY

[0005] The purpose of the present application is to solve the problems in the prior art, and propose a traffic flow prediction method based on a space-time graph convolution network, which has high robustness and can better capture the dynamic spatial dependence of traffic flow data and achieve better prediction results.

[0006] To achieve the above purpose, the present application proposes a traffic flow prediction method based on a space-time graph convolution network, comprising the following steps:

[0007] S1) Collecting traffic flow data of a target area by using a road sensor and performing preprocessing;

[0008] S2) Collecting additional information and encoding, the additional information including climate data, whether it is a weekday, and the types and quantities of public places around the node, and the climate data is preprocessed before encoding;

[0009] S3) Generating time series data according to the preprocessed traffic flow data obtained in step S1);

[0010] S4) Constructing two adaptive vectors E S ,E T as the spatial embedding representation of the node, and generating a spatial correlation matrix;

[0011] wherein E S ,E T are trainable parameters, which are randomly initialized by Gaussian distribution and are trained with the model;

[0012] S5) Constructing a DSRNN network model based on a DSGRU unit, taking the time series data obtained in step S3) as input and label, and inputting the additional information vector obtained in step S2) into the model to assist the model in predicting the traffic flow of the target area.

[0013] As a preferred, in the step S1), the collected traffic flow data is aggregated and normalized by using the Z-score normalization method, and the normalization formula is:

[0014]

[0015] wherein X J represents the traffic flow data to be normalized, represents the average value of the traffic flow data to be normalized, and S J represents the standard deviation of the traffic flow data to be normalized.

[0016] As a preferred, the step S2) comprises the following steps:

[0017] S21): For climate data, the collected climate data are aggregated and preprocessed using the Z-score normalization method. The normalization formula is:

[0018]

[0019] Among them, X Q This represents climate data to be normalized. S represents the average value of the climate data to be normalized. Q This represents the standard deviation of the climate data to be normalized.

[0020] Then concatenate the data into a matrix after the vector.

[0021] Each row represents a climate data vector, T represents the number of climate data vectors collected, and T represents the time length.

[0022] S22): For whether the day is a working day, if the day is a working day, it is recorded as 1, otherwise it is recorded as 0;

[0023] S23): Regarding the types and quantities of public places around a node, count the types and quantities of public places around the node, encode them to generate a public place distribution vector for each node, and then concatenate them into a matrix.

[0024] Where N represents the number of nodes.

[0025] Preferably, in step S3), a sliding window with a length of 24 and an interval of 1 slides along the time dimension of the traffic flow data preprocessed in step S1) to extract and generate time series data.

[0026] Preferably, in step S4), an adaptive spatial embedding representation is constructed based on the number of nodes N. As a low-dimensional spatial embedding representation of nodes, E is then... S and E T Multiplication yields the spatial correlation matrix

[0027] Preferably, step S5) includes the following steps:

[0028] S51): Divide the time series data generated in step S3). For each data point, use the first L data points as input and the last 24-L data points as labels. Denote the input data as {X1, X2, X3, ..., X...} L}, the label data is denoted as {X L+1 ,X L+2 ,X L+3,…X 23};

[0029] S52): Construct a DSRNN network model with DSGRU as the recurrent unit, and input data {X1,X2,X3,…X}. L The data is input into the DSGRU unit in chronological order to model the temporal and spatial dependencies. The final output, along with the additional information obtained in step S2), is encoded and fed into the Fusion hybrid module to output the prediction result Y. The predicted traffic flow is compared with the actual traffic flow, and the parameters of the DSRNN network model are trained and adjusted using the mean absolute error (MAE) as the loss function to achieve the expected accuracy. The formula for calculating the mean absolute error is:

[0030]

[0031] Among them, y i Forecast traffic volume This represents the actual traffic volume.

[0032] Preferably, the DSGRU unit consists of a dynamic spatial graph convolution unit (DSGC) and a gated recurrent unit (GRU).

[0033] As a preferred option, for traffic flow data X at time t t The dynamic spatial dependencies are captured by the Dynamic Spatial Graph Convolutional Unit (DSGC), and the adaptive spatial embedding representation E constructed in step S4) is used. S E T The dynamic spatial graph convolution operation is defined in the spatial domain as follows:

[0034]

[0035]

[0036]

[0037] in, and The coupling mapping function is used to construct spatial embedding representations at different times. W t , and b t These are the trainable parameters for dynamic spatial graph convolution.

[0038] As a preferred approach, after establishing the dynamic spatial dependencies, a gated recurrent unit (GRU) is used to capture the temporal dependencies, as shown in the following formula:

[0039] r = Sigmod([X′) t ,h t-1 Wr +b r )

[0040] u = Sigmod([X′) t ,h t-1 W u +b u )

[0041] c = Tanh([X′) t ,r⊙h t-1 W c +b c )

[0042] h t =u⊙h t-1 +(1-u)⊙c

[0043] Among them, h t h is the output at the current time step and serves as the input information state for the next time step. t-1 This represents the information state from the previous time step, where r represents the reset gate, u represents the update gate, c is the memory unit, and W... r W u W c b r b u and b c represents the trainable parameters of the gated recurrent unit (GRU), and ⊙ represents the Hadamard product operation.

[0044] As a preferred option, use X extra This represents the encoding of additional information, which is then combined with h. L The input is fed into the Fusion fusion module for feature mixing, and then decoded to obtain the model's prediction result, as shown in the following formula:

[0045] Y = Relu([h L ,X extra W1+b1)W2+b2

[0046] Among them, W1, W2, b1, and b2 are trainable parameters.

[0047] The beneficial effects of this invention are:

[0048] 1. Construct two adaptive vectors E S E T As a spatial embedding representation of nodes, E generates a spatial correlation matrix. S E TThese are trainable parameters, randomly initialized by a Gaussian distribution. As the model learns and trains, they can fully describe the spatial information of traffic flow data. They can replace traffic network information without the need to obtain traffic network information in advance. This avoids the defect that traffic network information cannot fully describe the spatial information of traffic data, which causes the model to only model a single spatial correlation, and improves the robustness of the model.

[0049] 2. Design a mapping function φ at each time step. S ,φ T and respectively E S E T Perform mapping, and then map the φ. S (E S ),φ T (E T Perform dynamic spatial graph convolution operation, φ S ,φ T All parameters are trainable and randomly initialized by a Gaussian distribution. As the model learns and trains, it models dynamic spatial dependencies and further combines with gated recurrent units (GRUs) to model temporal dependencies, building DSGRU units. The DSGRU units then become recurrent units, and with the addition of extra information encoding techniques and Fusion modules, a DSRNN network model is constructed for traffic flow prediction. This better reflects the temporal dynamics of traffic network information, thus better capturing the dynamic spatial dependencies of traffic flow data and achieving better prediction results.

[0050] 3. By incorporating additional information such as weather conditions, whether it is a weekday, and the distribution of public spaces, and feeding this information into the model through encoding and combining it with traffic data, the model can better reflect changes in urban traffic flow and further improve its ability to capture changes in urban traffic flow, thus adapting to the needs of the city.

[0051] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. [Attached Image Description]

[0052] Figure 1 This is a schematic diagram of traffic flow data recorded by sensor nodes;

[0053] Figure 2 This is a diagram of the DSRNN model architecture of the present invention;

[0054] Figure 3 This is a structural diagram of the DSGRU unit of the present invention.

Detailed Implementation Methods

[0055] See Figure 1 , Figure 2 and Figure 3This invention discloses a traffic flow prediction method based on spatiotemporal graph convolutional networks, comprising the following steps:

[0056] S1) Collect traffic flow data of the target area using road sensors and preprocess it:

[0057] In step S1), the collected traffic flow data is aggregated every 5 minutes and preprocessed using the Z-score normalization method. The normalization formula is:

[0058]

[0059] Among them, X J This represents traffic flow data to be normalized. S represents the average value of the traffic flow data to be normalized. J This represents the standard deviation of the traffic flow data to be normalized.

[0060] S2) Collect and encode additional information, including climate data, whether the day is a workday, and the types and number of public places around the node. The climate data is preprocessed before encoding.

[0061] Step S2) includes the following steps:

[0062] S21): For climate data (including real-time temperature data, real-time wind speed data, and real-time rainfall data; if there is no rain, it is recorded as 0), climate data is collected every 5 minutes. The collected climate data are aggregated and preprocessed using the Z-score normalization method. The normalization formula is:

[0063]

[0064] Among them, X Q This represents climate data to be normalized. S represents the average value of the climate data to be normalized. Q This represents the standard deviation of the climate data to be normalized.

[0065] Then concatenate the data into a matrix after the vector.

[0066] Each row represents a climate data vector, T represents the number of climate data vectors collected, and T represents the time length.

[0067] S22): For whether the day is a working day, if the day is a working day, it is recorded as 1, otherwise it is recorded as 0;

[0068] S23): For the types (e.g., if there are only hotels and restaurants near the node, it is recorded as 2) and quantities (e.g., if there are 2 hotels near the node, it is recorded as 2 in the hotel information; if there are no hotels, it is recorded as 0) of public places (including restaurants, shopping malls, schools, hospitals, cinemas, hotels, and stations) around the node, the types and quantities of public places around the node are counted, and the distribution vector of public places for each node is encoded and then concatenated into a matrix.

[0069] Where N represents the number of nodes.

[0070] S3) Generate time series data based on the preprocessed traffic flow data obtained in step S1):

[0071] In step S3), a sliding window with a length of 24 and an interval of 1 slides along the time dimension of the preprocessed traffic flow data Z from step S1), thereby extracting and generating time series data H. Each data point has a length of 24 and an interval of 5, representing traffic data for each node within 2 hours.

[0072] S4) Initialize the adaptive spatial embedding representation of the node E S E T :

[0073] In step S4), an adaptive spatial embedding representation E is constructed based on the number of nodes N. S , E is a low-dimensional spatial embedding representation of nodes. S E T These are trainable parameters, randomly initialized using a Gaussian distribution, and are learned during model training to characterize the complete spatial information of traffic data. The two parameters can be multiplied to obtain the spatial correlation matrix.

[0074] S5) Construct a DSRNN network model based on DSGRU units, using the time series data obtained in step S3) as input and labels, and then inputting the additional information vector obtained in step S2) into the model to assist the model in predicting traffic flow in the target area:

[0075] Step S5) includes the following steps:

[0076] S51): Divide the time series data H generated in step S3) into segments. For each data segment, use the first L data segments as input and the last 24-L data segments as labels. (The value of L depends on the specific prediction needs. For example, if L = 12, it means that the data from the previous hour is used as input and the data from the next hour is used as labels for training.) The input data is denoted as {X1, X2, X3, ... X... L}, the label data is denoted as {XL+1 ,X L+2 ,X L+3 ,…X 23};

[0077] S52): Construct a DSRNN network model with DSGRU as the recurrent unit, such as... Figure 2 The input data {X1,X2,X3,…X} L The data is input into the DSGRU unit in chronological order to model the temporal and spatial dependencies. The final output, along with the additional information obtained in step S2), is encoded and fed into the Fusion module to output the prediction result Y. The predicted traffic flow is compared with the actual traffic flow, and the parameters of the DSRNN network model are trained and adjusted using the mean absolute error (MAE) as the loss function to achieve the expected accuracy. The formula for calculating the mean absolute error is:

[0078]

[0079] Among them, y i Forecast traffic volume This represents the actual traffic volume.

[0080] Next, we will introduce the DSGRU unit and the Fusion (hybrid) module.

[0081] DSGRU unit: Composed of Dynamic Spatial Graph Convolutional Unit (DSGC) and Gated Recurrent Unit (GRU). For example... Figure 3 As shown, for traffic flow data X at time t t The dynamic spatial dependencies are captured by the Dynamic Spatial Graph Convolutional Unit (DSGC), and the adaptive spatial embedding representation E constructed in step S4) is utilized. S E T The dynamic spatial graph convolution operation is defined in the spatial domain as follows:

[0082]

[0083]

[0084]

[0085] in, and The coupling mapping function is used to construct spatial embedding representations at different times. W t , and b t These are trainable parameters for dynamic spatial graph convolution, and their specific values ​​are determined through the training process. This invention utilizes... Constructing the spatial correlation of the traffic road network at time t, and the spatial correlation of different times. and Different parameters are used to meet the dynamic nature of traffic network information. After the dynamic spatial dependencies are established, a gated recurrent unit (GRU) is used to capture temporal dependencies. It controls the flow of information states and captures temporal dependencies through learnable reset and update gate functions. The specific formula is as follows:

[0086] r = Sigmod([X′) t ,h t-1 W r +b r )

[0087] u = Sigmod([X′) t ,h t-1 W u +b u )

[0088] c = Tanh([X′) t ,r⊙h t-1 W c +b c )

[0089] h t =u⊙h t-1 +(1-u)⊙c

[0090] Among them, h t h is the output at the current time step and serves as the input information state for the next time step. t-1 This represents the information state from the previous time step, where r represents the reset gate, u represents the update gate, c is the memory unit, and W... r W u W c b r b u and b c represents the trainable parameters of the gated recurrent unit (GRU), whose specific values ​​are determined through the training process; ⊙ represents the Hadamard product operation.

[0091] Fusion module: using X extra This represents the encoding of additional information, which is then combined with h. L The input is fed into the Fusion module for feature fusion, and then decoded to obtain the model's prediction result, as shown in the following formula:

[0092] Y = Relu([h L ,X extra W1+b1)W2+b2

[0093] Among them, W1, W2, b1, and b2 are trainable parameters, and their specific values ​​are determined through the training process.

[0094] This invention proposes a DSRNN network model for traffic flow prediction, based on graph neural networks and using DSGRU as the recurrent unit. This invention constructs two learnable vectors E... S E T As a spatial embedding representation of nodes, a spatial correlation matrix is ​​generated to break away from the dependence of past models on traffic network information, thereby improving the model's ability to capture hidden spatial dependencies and its robustness. Furthermore, since graph convolutional networks can only model static spatial dependencies, resulting in a lack of dynamic spatial dependencies, this invention designs a mapping function φ. S ,φ T At each time step, E S E T Perform mapping, the mapped φ S (E S ),φ T (E T This invention performs dynamic spatial graph convolution operations to improve the model's ability to model dynamic spatial dependencies. Traffic prediction in cities depends not only on real-time traffic conditions but also on additional information such as weekday status, weather conditions, and the distribution of public places near road intersections. Therefore, this invention encodes this information and incorporates it into the model for final prediction, thereby improving the model's prediction accuracy.

[0095] The above embodiments are illustrative of the present invention and are not intended to limit the present invention. Any simple modifications to the present invention are within the scope of protection of the present invention.

Claims

1. A traffic flow prediction method based on spatiotemporal graph convolutional networks, characterized in that, Includes the following steps: S1): Collect traffic flow data of the target area using road sensors and perform preprocessing; S2): Collect and encode additional information, including climate data, whether the day is a workday, and the types and number of public places around the node. The climate data is preprocessed before being encoded. S3): Generate time series data based on the preprocessed traffic flow data obtained in step S1); S4): Construct two adaptive vectors E S E T As a spatial embedding representation of nodes, a spatial correlation matrix is ​​generated; Among them, E S E T These are trainable parameters, randomly initialized by a Gaussian distribution, and learned and trained as the model progresses. S5): Construct a DSRNN network model based on DSGRU units according to the spatial correlation matrix obtained in step S4). The DSRNN network model includes DSGRU units and Fusion modules. The time series data obtained in step S3) is used as input and label, and the additional information vector obtained in step S2) is input into the model to assist the model in predicting traffic flow in the target area. Step S5) includes the following steps: S51): Divide the time series data generated in step S3). For each data point, use the first L data points as input and the last 24-L data points as labels. Denote the input data as {X1, X2, X3, ..., X...} L }, the label data is denoted as {X L+1 ,X L+2 ,X L+3 ,…X 23 }; S52): Construct a DSRNN network model with DSGRU as the recurrent unit, and input data {X1,X2,X3,…X}. L The data is input into the DSGRU unit in chronological order to model the temporal and spatial dependencies. The final output, along with the additional information obtained in step S2), is encoded and fed into the Fusion hybrid module to output the prediction result Y. The predicted traffic flow is compared with the actual traffic flow, and the parameters of the DSRNN network model are trained and adjusted using the mean absolute error (MAE) as the loss function to achieve the expected accuracy. The formula for calculating the mean absolute error is: Among them, y i Forecast traffic volume This represents the actual traffic volume. The DSGRU unit consists of a dynamic spatial graph convolutional unit (DSGC) and a gated recurrent unit (GRU). For traffic flow data X at time t t The dynamic spatial dependencies are captured by the Dynamic Spatial Graph Convolutional Unit (DSGC), and the adaptive spatial embedding representation E constructed in step S4) is used. S E T The dynamic spatial graph convolution operation is defined in the spatial domain as follows: in, and The coupling mapping function is used to construct spatial embedding representations at different times. W t , and b t These are the trainable parameters for dynamic spatial graph convolution.

2. The traffic flow prediction method based on spatiotemporal graph convolutional networks as described in claim 1, characterized in that: In step S1), the collected traffic flow data is aggregated and preprocessed using the Z-score normalization method. The normalization formula is as follows: Among them, X J This represents traffic flow data to be normalized. S represents the average value of the traffic flow data to be normalized. J This represents the standard deviation of the traffic flow data to be normalized.

3. The traffic flow prediction method based on spatiotemporal graph convolutional networks as described in claim 1, characterized in that, Step S2) includes the following steps: S21): For climate data, the collected climate data are aggregated and preprocessed using the Z-score normalization method. The normalization formula is: Among them, X Q This represents climate data to be normalized. S represents the average value of the climate data to be normalized. Q This represents the standard deviation of the climate data to be normalized. Then concatenate the data into a matrix after the vector. Each row represents a climate data vector, T represents the number of climate data vectors collected, and T represents the time length. S22): For whether the day is a working day, if the day is a working day, it is recorded as 1, otherwise it is recorded as 0; S23): Regarding the types and quantities of public places around a node, count the types and quantities of public places around the node, encode them to generate a public place distribution vector for each node, and then concatenate them into a matrix. Where N represents the number of nodes.

4. The traffic flow prediction method based on spatiotemporal graph convolutional networks as described in claim 1, characterized in that: In step S3), a sliding window with a length of 24 and an interval of 1 slides along the time dimension of the traffic flow data preprocessed in step S1) to extract and generate time series data.

5. The traffic flow prediction method based on spatiotemporal graph convolutional networks as described in claim 1, characterized in that: In step S4), an adaptive spatial embedding representation is constructed based on the number of nodes N. As a low-dimensional spatial embedding representation of nodes, E is then... S and E T Multiplication yields the spatial correlation matrix 6. The traffic flow prediction method based on spatiotemporal graph convolutional networks as described in claim 1, characterized in that: After establishing the dynamic spatial dependencies, a gated recurrent unit (GRU) is used to capture the temporal dependencies. The specific formula is as follows: r=Sigmod([X′ t ,h t-1 ]W r +b r ) u=Sigmod([X′ t ,h t-1 ]W u +b u ) c=Tanh([X′ t ,r⊙h t-1 ]W c +b c ) h t =u⊙h t-1 +(1-u)⊙c Among them, h t h is the output at the current time step and serves as the input information state for the next time step. t-1 This represents the information state from the previous time step, where r represents the reset gate, u represents the update gate, c is the memory unit, and W... r W u W c b r b u and b c represents the trainable parameters of the gated recurrent unit (GRU), and ⊙ represents the Hadamard product operation.

7. The traffic flow prediction method based on spatiotemporal graph convolutional networks as described in claim 1, characterized in that: Use X extra This represents the encoding of additional information, which is then combined with h. L The input is fed into the Fusion fusion module for feature mixing, and then decoded to obtain the model's prediction result, as shown in the following formula: Y=Relu([h L X extra (W1+b1)W2+b2 Among them, W1, W2, b1, and b2 are trainable parameters.

Images