CN114694379A - Traffic flow prediction method and system based on self-adaptive dynamic graph convolution - Google Patents

Abstract

The invention discloses a traffic flow prediction method and system based on self-adaptive dynamic graph convolution. The method includes: acquiring historical traffic data and preprocessing the historical traffic data; constructing a static adjacency relation graph; constructing an adaptive dynamic adjacency relation Graph tensor; construct an adaptive dynamic graph convolution prediction model; train the adaptive dynamic graph convolution prediction model based on preprocessed historical traffic data; input the traffic flow data of the data to be measured into the trained prediction model to obtain the prediction result . The system includes: a preprocessing module, a first building module, a second building module, a model building module, a training module and a prediction module. By using the present invention, the accuracy of traffic flow prediction can be improved. As a traffic flow prediction method and system based on adaptive dynamic graph convolution, the present invention can be widely used in the field of traffic prediction.

Description

A traffic flow prediction method and system based on adaptive dynamic graph convolution

technical field

The invention relates to the field of traffic prediction, in particular to a traffic flow prediction method and system based on adaptive dynamic graph convolution.

Background technique

Traffic flow prediction aims to predict the future traffic flow conditions (such as traffic speed, traffic volume, etc.) in the road network based on historical traffic observations. Accurate traffic prediction is an important foundation for building an intelligent transportation system, and is of great significance to various downstream applications such as traffic time estimation, route planning, and traffic light control. Accurate traffic forecasting remains a challenge due to the highly dynamic and complex spatiotemporal dependencies of urban transportation networks.

Traditional statistical signal processing methods such as ARIMA model and Support Vector Regression (SVR) model model traffic forecasting with univariate time signal regression. They rely on the assumption of signal stationarity and ignore the interrelationships between traffic nodes, making it difficult to capture the complex traffic patterns in the real world. With the development of deep learning technology, models such as convolutional neural networks have emerged, but they can only process spatial information in the way of Euclidean spatial rasterization, and cannot deal with irregular traffic network topology.

SUMMARY OF THE INVENTION

In order to solve the above technical problems, the purpose of the present invention is to provide a traffic flow prediction method and system based on adaptive dynamic graph convolution, which can improve the accuracy of traffic flow prediction.

The first technical solution adopted by the present invention is: a traffic flow prediction method based on adaptive dynamic graph convolution, comprising the following steps:

Obtain historical traffic data and preprocess the historical traffic data to obtain preprocessed historical traffic data;

Obtain the geospatial distances of traffic nodes and build a static adjacency graph;

Characterize traffic nodes and build adaptive dynamic adjacency graph tensors;

Build an adaptive dynamic graph convolution prediction model based on the static adjacency graph and the adaptive dynamic adjacency graph tensors;

The adaptive dynamic graph convolution prediction model is trained based on the preprocessed historical traffic data, and the trained prediction model is obtained;

Input the traffic flow data of the data to be tested into the trained prediction model to obtain the prediction result.

Further, the step of obtaining historical traffic data and preprocessing the historical traffic data to obtain the preprocessing historical traffic data specifically includes:

Set the time step interval, the maximum number of historical observation time steps and the maximum number of forecast time steps;

Divide one day into equal-length periods according to the time step interval, and obtain the period index sequence;

According to the time step interval, the maximum number of historical observation time steps and the maximum number of predicted time steps, the historical traffic data is sliced by sliding window, and the traffic flow feature sequence is obtained;

The feature index pair is constructed according to the traffic flow feature and the time period index sequence, and the preprocessed historical traffic data is obtained.

Further, the expression of the static adjacency graph is as follows:

表示交通节点v_i与v_j的地理空间距离，σ表示各节点间距离的标准差。In the above formula, the static adjacency graph,

represents the geographic spatial distance between traffic nodes v _i and v _j , and σ represents the standard deviation of the distance between each node.

Further, the step of characterizing traffic nodes and constructing an adaptive dynamic adjacency graph tensor specifically includes:

Set the representation dimensions of traffic nodes and time periods, and construct a traffic node representation matrix and a time period representation matrix;

Based on the tensor synthesis method, the tensor is calculated according to the traffic node representation matrix and the time period representation matrix;

The tensors are nonlinearly mapped and normalized to obtain adaptive dynamic adjacency graph tensors.

Further, the step of constructing an adaptive dynamic graph convolution prediction model according to the static adjacency graph and the adaptive dynamic adjacency graph tensor specifically includes:

Obtain the adaptive dynamic adjacency graph according to the adaptive dynamic adjacency graph tensor;

Build an adaptive dynamic graph convolution module and use static adjacency graph and dynamic adjacency graph to perform graph convolution operations;

Embed the adaptive dynamic graph convolution module into the gated cyclic unit and replace the full connection calculation to obtain the gated cyclic unit with adaptive dynamic graph convolution;

Based on the gated recurrent unit with adaptive dynamic graph convolution, a model constituting the encoder-decoder structure is constructed, and an adaptive dynamic graph convolution prediction model is obtained.

Further, the step of training the adaptive dynamic graph convolution prediction model based on the preprocessing historical traffic data to obtain the trained prediction model specifically includes:

Based on the planned sampling method, the actual value of historical traffic flow features is used as input with probability ε, and the output estimated value of the previous time step is used as input with probability 1-ε. Train to get the trained prediction model.

The second technical solution adopted by the present invention is: a traffic flow prediction system based on adaptive dynamic graph convolution, comprising:

The preprocessing module is used to obtain the historical traffic data and preprocess the historical traffic data to obtain the preprocessed historical traffic data;

The first building module is used to obtain the geospatial distance of the traffic node and construct a static adjacency graph;

The second building module is used to characterize the traffic nodes and construct an adaptive dynamic adjacency graph tensor;

A model building module for constructing an adaptive dynamic graph convolution prediction model based on the static adjacency graph and the adaptive dynamic adjacency graph tensors;

The training module trains the adaptive dynamic graph convolution prediction model based on the preprocessed historical traffic data, and obtains the trained prediction model;

The prediction module is used to input the traffic flow data of the data to be measured into the trained prediction model to obtain the prediction result.

The beneficial effects of the method and system of the present invention are as follows: the present invention uses different adaptive adjacency graphs at different time points to perform dynamic graph convolution on the representation of traffic nodes, so as to mine the complex dynamic patterns of the traffic network and improve the accuracy of traffic flow prediction. In addition, the trainable node representation matrix and the time period representation matrix are respectively set, and the dynamic adjacency graph of nodes in different time periods is generated by tensor synthesis, which avoids defining a node representation in each time period. When it is huge, it effectively reduces the number of parameters of the prediction model.

Description of drawings

1 is a flow chart of the steps of a traffic flow prediction method based on adaptive dynamic graph convolution of the present invention;

2 is a structural block diagram of a traffic flow prediction system based on adaptive dynamic graph convolution of the present invention;

3 is a schematic diagram of the present invention implementing an adaptive dynamic graph convolution prediction model;

FIG. 4 is a comparison chart of the prediction performance between the method of the present invention and the existing typical traffic prediction method based on graph convolution.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The numbers of the steps in the following embodiments are set only for the convenience of description, and the sequence between the steps is not limited in any way, and the execution sequence of each step in the embodiments can be adapted according to the understanding of those skilled in the art Sexual adjustment.

As shown in Figure 1, the present invention provides a traffic flow prediction method based on adaptive dynamic graph convolution, the method comprising the following steps:

S1. Obtain historical traffic data and preprocess the historical traffic data to obtain preprocessed historical traffic data;

S1.1. Set the time step interval, the maximum number of historical observation time steps and the maximum number of prediction time steps;

Specifically, set the time step interval ΔT to 5 minutes, set the maximum number of historical observation time steps P=12 (that is, the duration of 1 hour), and the maximum number of predicted time steps Q=12 (that is, the duration of 1 hour);

S1.2. Divide one day into equal-length periods according to the time step interval, and obtain the period index sequence;

Specifically, a day is divided into L=288 time periods of equal length according to ΔT of 5 minutes, and the indices l of each time period in the day correspond to 0, 1, . . . , 287 respectively.

S1.3. Perform sliding window slicing on the historical traffic data according to the time step interval, the maximum number of historical observation time steps and the maximum number of predicted time steps to obtain a traffic flow feature sequence;

207为交通节点数量，1为每个节点的特征数(即仅采用交通速度一种特征)，对应时段在一天中的索引分别为l_t-11,…,l_t,l_t+1,…,l_t+12；Specifically, the set time step interval ΔT is 5 minutes, the maximum number of historical observation time steps P=12, and the maximum number of predicted time steps Q=12, the historical traffic flow characteristic data is sliced by sliding window, and each window slice is The traffic flow feature sequence X _t-11 ,…,X _t ,X _t+1 ,…,X _t+12 of length 24, in which the

207 is the number of traffic nodes, 1 is the number of features of each node (that is, only one feature of traffic speed is used), and the indices of the corresponding time periods in one day are l _t-11 ,...,l _t ,l _t+1 ,... ,l _t+12 ;

S1.4, construct a feature index pair according to the traffic flow feature and the time period index sequence, and obtain the preprocessed historical traffic data.

Specifically, the traffic flow feature sequence and the corresponding time period index sequence are combined into a "traffic feature-time period index pair" sequence to obtain preprocessed historical traffic flow data samples.

A single sample has the form [(X _t-11 ,l _t-11 ),…,(X _t ,l _t ),(X _t+1 ,l _t+1 ),…,(X _t+12 ,l _{t +12} )] is a sequence of length 24.

S2. Obtain the geospatial distance of traffic nodes and construct a static adjacency graph;

Specifically, a static adjacency graph is obtained by calculating the proximity of traffic nodes in the form of a Gaussian kernel function, and the expression is as follows:

表示交通节点v_i与v_j的地理空间距离，σ表示各节点间距离的标准差。In the above formula,

represents the geographic spatial distance between traffic nodes v _i and v _j , and σ represents the standard deviation of the distance between each node.

S3. Characterize the traffic nodes and construct an adaptive dynamic adjacency graph tensor;

S3.1. Set the representation dimensions of traffic nodes and time periods, and construct a traffic node representation matrix and a time period representation matrix;

终端交通节点表征矩阵

一天中各时段表征矩阵

核张量

E^s,E^t,E^o,C随机初始化；Specifically, set the dimension d=30 of the representation of traffic nodes and the representation of each time period in a day, and construct the representation matrix of source-end traffic nodes accordingly

Terminal Traffic Node Characterization Matrix

Time-of-day representation matrix

kernel tensor

E ^s , E ^t , E ^o , C are initialized randomly;

S3.2. Based on the tensor synthesis method, the tensor is calculated according to the traffic node representation matrix and the time period representation matrix;

计算表达式如下：Specifically, according to E ^s , E ^t , E ^o , C, the tensor is calculated by tensor synthesis

The calculation expression is as follows:

A ^d = C × ₁ E ^t × ₂ E ^s × ₃ E ^e

S3.3. Perform nonlinear mapping and normalization on the tensor to obtain an adaptive dynamic adjacency graph tensor.

计算表达式如下：Specifically, perform nonlinear mapping on A ^d and normalize it to obtain the final traffic node dynamic adjacency graph tensor

The calculation expression is as follows:

In the above formula, the nonlinear mapping uses the leakyReLU function, and the softmax function normalizes the last dimension of the tensor;

S4, constructing an adaptive dynamic graph convolution prediction model according to the static adjacency graph and the adaptive dynamic adjacency graph tensor;

S4.1. Obtain an adaptive dynamic adjacency graph according to the adaptive dynamic adjacency graph tensor;

在其第一个维度的第l个切片

的含义为在一天中第l个时段的交通节点动态邻接关系图。Specifically, the node dynamic adjacency graph tensor

the lth slice in its first dimension

The meaning of is the dynamic adjacency graph of traffic nodes in the lth period of the day.

S4.2. Build an adaptive dynamic graph convolution module dgconv( ) and use static adjacency graph and dynamic adjacency graph geometry to perform graph convolution operations, and the graph convolution order used is K=2;

Specifically, the calculation formula is:

为一天中第l个时段对应的动态邻接关系图，W均为可训练权重矩阵。In the above formula, H _in and H _out are the input representation of the traffic node and the output representation of the adaptive graph convolution module, respectively, K is the graph convolution order, D ^f is the degree matrix of the static adjacency graph A ^s , and D ^b is the degree matrix of the transposed matrix of A ^s ,

is the dynamic adjacency graph corresponding to the l-th time period in a day, and W is a trainable weight matrix.

S4.3. Embed the adaptive dynamic graph convolution module into the gated cyclic unit and replace the fully connected calculation to obtain a gated cyclic unit with adaptive dynamic graph convolution;

Specifically, the dgconv( ) is embedded in the gated cyclic unit GRU to replace the fully connected calculation, so as to obtain a GRU unit containing adaptive dynamic graph convolution; that is, for the GRU unit at time step t, it is calculated according to the following expression:

H _t =u _t ⊙H _t-1 +(1-u _t )⊙c _t ,

为

在其第一个维度的第l_t个切片；σ(·)表示sigmoid函数，||代表矩阵拼接操作，⊙代表矩阵Hadamard积操作。In the above formula, X _t and H _t are the input traffic flow feature and output hidden state of the current time step t respectively, H _t-1 is the hidden state of the previous time step; l _t is the input sample and the traffic flow feature X _t the index of the corresponding time of day,

for

The _t -th slice in its first dimension; σ( ) represents the sigmoid function, || represents the matrix concatenation operation, and ⊙ represents the matrix Hadamard product operation.

S4.4, constructing a model constituting an encoder-decoder structure based on a gated recurrent unit including an adaptive dynamic graph convolution, and obtaining an adaptive dynamic graph convolution prediction model.

Specifically, the length of the GRU encoder is 12 equal to P, the length of the GRU decoder is 12 equal to Q, and the number of layers of the encoder-decoder structure is 2, and the GRU unit containing the adaptive dynamic graph convolution is used to form the GRU encoding The decoder-decoder prediction model is shown in Figure 3 for a schematic diagram of the model structure.

S5. Train the adaptive dynamic graph convolution prediction model based on the preprocessed historical traffic data, and obtain the trained prediction model;

Based on the planned sampling method, the actual value of historical traffic flow features is used as input with probability ε, and the output estimated value of the previous time step is used as input with probability 1-ε. Train to get the trained prediction model.

Specifically, when training the model, the first 12 "traffic feature-period index pairs" of the sequence in each sample are input to the encoder in the model, and the last 12 "traffic feature-period index pairs" are input to the decoder in the model; The minimum mean absolute error (MAE) criterion is adopted; the Adam optimizer is adopted; the learning rate starts at 0.01, and decays at a rate of 0.1 at the 20th, 30th, 40th, and 50th rounds of training; during the ith iteration of the training process The planned sampling probability _εi of the GRU decoder in the model is calculated by the following function:

where τ is taken as 2000;

Train based on model output and ground truth error. The model adopts the GRU encoder-decoder structure: (1) During training, the generated input sequence actually corresponds to the two segments of the encoder and the decoder. (2) The input of the encoder is fixed with the input of the first segment. The hidden vector of the encoder only in the last step will be input to the encoder as the initial hidden vector of the first step of the encoder, but without the output of the model, the decoder can only output. (3) During training, the input of each step of the decoder is selected according to the probability, either the value corresponding to the current time step in the second segment of the input sequence (1) is input, or the output of the previous time step of the decoder is input. forecast/estimate value.

S6. Input the traffic flow data of the data to be measured into the trained prediction model to obtain a prediction result.

The traffic flow prediction performance of the present invention is further described below in conjunction with Fig. 4:

The method of the present invention is compared with the prediction performance of the existing typical traffic prediction method based on graph convolution, and the compared methods include: DCRNN (diffusion convolutional recurrent neural network) model, STGCN (space-time graph convolutional network) model, Graph-WaveNet model. Among them, the DCRNN model uses a fixed static distance adjacency graph for diffusion graph convolution, and combines the encoder-decoder structure for traffic flow prediction; the STGCN model uses a fixed static distance adjacency graph in the form of Chebyshev polynomials. Convolution, and combined with 1D-CNN time domain convolution for traffic flow prediction; Graph-WaveNet model, based on STGCN model, adds a static adaptive adjacency graph shared at all times for graph convolution, and combines multiple scales of Atrous temporal convolution for traffic flow prediction. Figure 4 is a comparison chart of prediction performance, showing the error of each method for the traffic flow prediction in the next 15 minutes (3 steps), 30 minutes (6 steps) and 60 minutes (12 steps); where MAE stands for mean absolute error, RMSE stands for Root Mean Squared Error, MAPE stands for Mean Absolute Percentage Error. It can be seen that the present invention obtains an overall better traffic flow by using different adaptive adjacency graphs at different time points to perform dynamic graph convolution, compared with the methods based on fixed static adjacency graphs and adaptive static adjacency graphs prediction accuracy.

As shown in Figure 2, a traffic flow prediction system based on adaptive dynamic graph convolution includes:

The preprocessing module is used to obtain the historical traffic data and preprocess the historical traffic data to obtain the preprocessed historical traffic data;

The first building module is used to obtain the geospatial distance of the traffic node and construct a static adjacency graph;

The second building module is used to characterize the traffic nodes and construct an adaptive dynamic adjacency graph tensor;

A model building module for constructing an adaptive dynamic graph convolution prediction model based on the static adjacency graph and the adaptive dynamic adjacency graph tensors;

The training module trains the adaptive dynamic graph convolution prediction model based on the preprocessed historical traffic data, and obtains the trained prediction model;

The prediction module is used to input the traffic flow data of the data to be measured into the trained prediction model to obtain the prediction result.

The contents in the above method embodiments are all applicable to the present system embodiments, the specific functions implemented by the present system embodiments are the same as the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

A traffic flow prediction device based on adaptive dynamic graph convolution:

at least one processor;

at least one memory for storing at least one program;

When the at least one program is executed by the at least one processor, the at least one processor implements the above-mentioned traffic flow prediction method based on adaptive dynamic graph convolution.

The contents in the above method embodiments are all applicable to the present device embodiments, the specific functions implemented by the present device embodiments are the same as the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

A storage medium storing processor-executable instructions, wherein the processor-executable instructions, when executed by the processor, are used to implement the above-mentioned traffic based on adaptive dynamic graph convolution Flow prediction method.

The contents in the above method embodiments are all applicable to the present storage medium embodiments, the specific functions implemented by the present storage medium embodiments are the same as the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments. same.

The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the described embodiments, and those skilled in the art can make various equivalent deformations or replacements without departing from the spirit of the present invention. , these equivalent modifications or substitutions are all included within the scope defined by the claims of the present application.

Claims (7)

1. a traffic flow prediction method based on adaptive dynamic graph convolution, is characterized in that, comprises the following steps: Obtain historical traffic data and preprocess the historical traffic data to obtain preprocessed historical traffic data; Obtain the geospatial distances of traffic nodes and build a static adjacency graph; Characterize traffic nodes and build adaptive dynamic adjacency graph tensors; Build an adaptive dynamic graph convolution prediction model based on the static adjacency graph and the adaptive dynamic adjacency graph tensors; The adaptive dynamic graph convolution prediction model is trained based on the preprocessed historical traffic data, and the trained prediction model is obtained; Input the traffic flow data of the data to be tested into the trained prediction model to obtain the prediction result. 2. a kind of traffic flow prediction method based on adaptive dynamic graph convolution according to claim 1, is characterized in that, described acquisition historical traffic data and historical traffic data are preprocessed, obtain preprocessing historical traffic data: steps, which specifically include: Set the time step interval, the maximum number of historical observation time steps and the maximum number of forecast time steps; Divide one day into equal-length periods according to the time step interval, and obtain the period index sequence; According to the time step interval, the maximum number of historical observation time steps and the maximum number of predicted time steps, the historical traffic data is sliced by sliding window, and the traffic flow feature sequence is obtained; The feature index pair is constructed according to the traffic flow feature and the time period index sequence, and the preprocessed historical traffic data is obtained. 3. a kind of traffic flow prediction method based on adaptive dynamic graph convolution according to claim 2, is characterized in that, the expression of described static adjacency relation graph is as follows:

In the above formula,

represents the geographic spatial distance between traffic nodes v _i and v _j , and σ represents the standard deviation of the distance between each node. 4. A kind of traffic flow prediction method based on adaptive dynamic graph convolution according to claim 3, is characterized in that, the described step of characterizing traffic nodes and constructing adaptive dynamic adjacency relation graph tensor, its specific include: Set the representation dimensions of traffic nodes and time periods, and construct a traffic node representation matrix and a time period representation matrix; Based on the tensor synthesis method, the tensor is calculated according to the traffic node representation matrix and the time period representation matrix; The tensors are nonlinearly mapped and normalized to obtain adaptive dynamic adjacency graph tensors. 5. The traffic flow prediction method based on adaptive dynamic graph convolution according to claim 4, wherein the adaptive dynamic graph convolution is constructed according to static adjacency graph and adaptive dynamic adjacency graph tensor. This step of the prediction model includes: Obtain the adaptive dynamic adjacency graph according to the adaptive dynamic adjacency graph tensor; Build an adaptive dynamic graph convolution module and use static adjacency graph and dynamic adjacency graph to perform graph convolution operations; Embed the adaptive dynamic graph convolution module into the gated cyclic unit and replace the full connection calculation to obtain the gated cyclic unit with adaptive dynamic graph convolution; Based on the gated recurrent unit with adaptive dynamic graph convolution, a model constituting the encoder-decoder structure is constructed, and an adaptive dynamic graph convolution prediction model is obtained. 6. A kind of traffic flow prediction method based on adaptive dynamic graph convolution according to claim 5, characterized in that, the adaptive dynamic graph convolution prediction model is trained based on preprocessing historical traffic data, and the training is completed. This step of the prediction model, which specifically includes: Based on the planned sampling method, the actual value of historical traffic flow features is used as input with probability ε, and the output estimated value of the previous time step is used as input with probability 1-ε. Train to get the trained prediction model. 7. A traffic flow prediction system based on adaptive dynamic graph convolution, characterized in that, comprising: The preprocessing module is used to obtain the historical traffic data and preprocess the historical traffic data to obtain the preprocessed historical traffic data; The first building module is used to obtain the geospatial distance of the traffic node and construct a static adjacency graph; The second building module is used to characterize the traffic nodes and construct an adaptive dynamic adjacency graph tensor; A model building module for constructing an adaptive dynamic graph convolution prediction model based on the static adjacency graph and the adaptive dynamic adjacency graph tensors; The training module trains the adaptive dynamic graph convolution prediction model based on the preprocessed historical traffic data, and obtains the trained prediction model; The prediction module is used to input the traffic flow data of the data to be measured into the trained prediction model to obtain the prediction result.

Images