Multi-graph convolutional traffic propagation flow prediction method and system considering periodic fluctuation

Abstract

The application relates to a multi-graph convolution traffic propagation flow prediction method and system considering periodic fluctuation, and belongs to the technical field of traffic prediction. The method comprises the following steps: first, identifying traffic propagation flow information and constructing a traffic propagation matrix; second, constructing an adjacent propagation graph, a homologous propagation graph and a same-destination propagation graph by using road network information and historical data; third, mining periodic fluctuation of traffic propagation flow based on an input time sequence; fourth, constructing a multi-graph convolution traffic propagation flow prediction model considering periodic fluctuation; and fifth, training and evaluating the model. The system comprises a matrix construction module, a propagation graph construction module, a time feature mining module, a model construction module and a training and evaluation module. The influence of time periodicity and fluctuation and the influence of multiple spatial correlation features are fully considered.

Description

TECHNICAL FIELD

[0001] The present application relates to a multi-graph convolution traffic propagation flow prediction method and system considering periodic fluctuation, belonging to the technical field of traffic prediction. BACKGROUND

[0002] Traffic propagation flow refers to the flow of vehicles in the road network, and it is one of the important problems in the field of transportation. It refers to the propagation and diffusion process of traffic flow in the traffic network, which describes the propagation process of traffic flow from the starting point to the ending point.

[0003] Traffic propagation flow is similar to other traffic flows in the time dimension and has a certain periodicity, that is, traffic propagation flow presents a repeated pattern and rule within a certain time period, but this periodicity has a certain fluctuation and does not have a completely fixed time window. For example, the peak of the morning peak will appear around 8 o'clock every day, which shows that traffic propagation flow has periodicity; but at the same time, the specific occurrence time of the morning peak every day may be slightly different, which shows that the periodicity of traffic propagation flow has a certain fluctuation. This fluctuation will affect the accuracy and reliability of traffic propagation flow prediction.

[0004] Space is a very important factor in traffic prediction, which can have a significant impact on traffic propagation flow. Traditional convolutional neural networks (CNN), graph convolutional networks (GCN) and other convolutional neural networks can mine the complex spatial features of traffic data, and thus have been widely applied in the field of traffic prediction, but they cannot fully capture the influence of multi-spatial correlation features.

[0005] For example, Chinese patent CN202210832846.6 proposes a dynamic spatio-temporal multi-graph convolutional neural network (MGCN) based on graph convolution and attention mechanism. The model mainly includes a spatio-temporal graph convolution module for capturing comprehensive and dynamic spatio-temporal correlation of traffic network. Then, the future traffic conditions are analyzed along the spatio-temporal dimension on two decoders, respectively, to aggregate multi-dimensional information.

[0006] The model proposed by it adopts a gating fusion mechanism to fuse the results of different graph convolutions when mining spatio-temporal correlation features, and when mining time features, the input time series only contains the most recent time series, which cannot fully consider the influence of time periodicity and fluctuation, nor can it fully capture the influence of multi-spatial correlation features.

[0007] For example, Chinese patent CN201911148344.6 discloses a method for irregular area traffic prediction based on multi-graph convolution and GRU, including the following steps: step one, divide the region into N non-connected irregular areas; step two, simplify the historical trajectory data in space and time, and calculate the entering amount and outflow amount of all areas at each time step; step three, establish multiple association graphs between areas, construct the corresponding adjacency matrix, and represent the diversified spatial association between irregular areas; step four, design a multi-graph convolution neural network based on the association graph between areas, fuse the diversified spatial association features between areas, and obtain the multi-graph convolution fusion result; step five, based on the multi-graph convolution fusion result, use GRU neural network to capture time association; step six, select a suitable loss function, train to obtain a prediction model, and predict the entering amount and outflow amount of each area through the prediction model.

[0008] The focus is on predicting the inflow and outflow of the region in the corresponding period by using the inflow, outflow, date attribute, hour attribute, weather attribute and temperature attribute in the irregular area. This invention predicts the inflow and outflow in the irregular area, and does not mine the periodic fluctuation of traffic information.

[0009] Also, Chinese patent CN202011641479.9 discloses a long-term traffic prediction method, acquires the first historical traffic data of the nodes in the constructed traffic network graph, and convolves the first historical traffic data through the first convolution layer in the pre-set traffic prediction model; the first historical traffic data after convolution is processed through the first iterative RNN operator in the model to output the traffic prediction result of the first time step, and the traffic prediction result of the first time step is input to the next iterative RNN operator for traffic prediction until the T p th iterative RNN operator outputs the traffic prediction result of the T p th time step, and the traffic prediction result of the time step is input to the second convolution layer for convolution processing through the splicing module in the model, and the final traffic prediction result is output.

[0010] The model proposed by it convolves the data twice, uses multiple RNNs to obtain time features in between, and uses the same weight adjacency matrix for both convolutions, which cannot well mine the influence of multiple spatial association features.

[0011] To solve the problems in the background art, the application provides a traffic propagation flow prediction method considering periodic fluctuation multi-graph convolution, which fully considers the influence of time periodicity and fluctuation and the influence of multi-space correlation features. The traffic propagation flow information is accurately predicted mainly by using vehicle GPS positioning data combined with road network topology information. First, the traffic propagation flow information is calculated to construct a traffic propagation matrix. Second, the adjacent propagation graph, the homologous propagation graph and the same destination propagation graph are constructed by using the road network information and the historical data. Third, the periodic fluctuation of the traffic propagation flow is mined based on the input time series. Fourth, a traffic propagation flow prediction model considering periodic fluctuation multi-graph convolution is constructed. Finally, the model is trained and evaluated. SUMMARY

[0012] The application aims to provide a traffic propagation flow prediction method and system considering periodic fluctuation multi-graph convolution, which fully considers the influence of time periodicity and fluctuation and the influence of multi-space correlation features.

[0013] The traffic propagation flow prediction method considering periodic fluctuation multi-graph convolution comprises the following steps:

[0014] First, the traffic propagation flow information is identified to construct a traffic propagation matrix.

[0015] Second, the adjacent propagation graph, the homologous propagation graph and the same destination propagation graph are constructed by using the road network information and the historical data.

[0016] Third, the periodic fluctuation of the traffic propagation flow is mined based on the input time series.

[0017] Fourth, a traffic propagation flow prediction model considering periodic fluctuation multi-graph convolution is constructed.

[0018] Fifth, the model is trained and evaluated.

[0019] Preferably, the traffic propagation flow information is counted by vehicle GPS positioning data. The vehicle can be matched to the corresponding road segment according to the latitude, longitude and direction angle, so that the traffic propagation flow information can be obtained by knowing that a vehicle passes through a specific road segment at a specific time.

[0020] For a vehicle V, the latitude and longitude of the vehicle are matched to the corresponding road segment and sorted by time, so that the road segment sequence information of the vehicle in a day can be obtained, that is:

[0021] LINK=link1→link2→link3→...→link n →...

[0022] At the same time, there is a corresponding time series:

[0023] T = t1→t2→t3→...→t n →...

[0024] In this sequence, link n The sequence points around link n The identification method is as follows:

[0025] After calculating the LINK of a specific vehicle, it is traversed. If at time t n-1 link n-1 is not equal to link n , and t n -t n-1 <60s, it is considered that at time t n-1 , there is 1 propagation from link n-1 to link n ;

[0026] All data is counted, i.e. the flow propagation matrix Q N*H between links in each time interval is obtained, where N represents the number of historical time intervals, and H represents the number of propagation relationships between links.

[0027] Preferably, the adjacency propagation graph A is constructed as follows:

[0028] TPT j The relationship between adjacent flow propagation transactions is considered by using the adjacency propagation graph. If there are links link a , link b , link c , and the following relationship exists: link a →link b →link c , and TPT i represents the flow propagation transaction of link a →link b , and TPT j represents the flow propagation transaction of link b →link c ; because the downstream link of TPT i and the upstream link of TPT j are both link b , the value of the adjacency graph A 1 at position A ij is 1, otherwise it is 0.

[0029] The propagation adjacency graph is defined as:

[0030]

[0031] A traffic propagation transaction TPT is defined as follows:

[0032] TPT = <link u , link d , Q ud >

[0033] where link u , link d , Q ud represent the upstream link, the downstream link and the propagation volume from link u to link d , respectively.

[0034] Preferably, the construction method of the homogenous propagation graph G is as follows:

[0035]

[0036] where Q i , Q j ∈ Q N*H , Q i represent the propagation relationship TPT i the propagation volume in each time interval; cor(Q i , Q j ) calculates the Pearson correlation coefficient between Q i and Q j . The larger the Pearson correlation coefficient is, the stronger the correlation between the two propagation transactions is.

[0037] The formula of the Pearson correlation coefficient is as follows:

[0038]

[0039] where r represents the Pearson correlation coefficient between Q i and Q j . n represents the total number of time intervals. x k , y k represent the propagation volume of Q i , Q j in the kth time interval, respectively. represents the average propagation volume of Q i , Q j .

[0040] Preferably, the construction method of the homogenous destination propagation graph G is as follows:

[0041]

[0042] Preferably, the traffic propagation flow periodicity fluctuation based on the input time series is mined by the latest time series, the previous day time series, the previous week time series and the previous month time series as follows:

[0043] X h = [X t-a , X t-a+1 ,..., X t-a+2 ,..., X t-1 ]

[0044] X d = [X t-T-b , X t-T-b+1 ,..., X t-T ,..., X t-T+b-1 , X t-T+b ]

[0045] X w = [X t-7T-c , X t-7T-c+1 ,..., X t-7T ,..., X t-7T+c-1 , X t-7T+c ]

[0046] X m = [X t-nT-e , X t-nT-e+1 ,..., X t-nT ,..., X t-nT+e-1 , X t-nT+e ]

[0047] wherein X t ∈ Q N*H , represents the propagation quantity of all propagation relationships in the tth time interval; a represents the length of the latest time series, b, c and e respectively represent the time series length taken before the corresponding period before the prediction period, i.e. the length of the previous day time series, the previous week time series and the previous month time series are 2b+1, 2c+1 and 2e+1 respectively; T represents the number of time intervals in a day, and N represents the number of days in the corresponding month.

[0048] Preferably, the multi-graph convolution traffic propagation flow prediction model passes through multi-graph convolution to mine spatial features, then passes through a gated recurrent unit to mine time features, and finally fuses the results through a fully connected layer to obtain a prediction result.

[0049] Preferably, the general formula of the multi-graph convolution is as follows:

[0050]

[0051] wherein, denotes the degree matrix, H (l+1) and H (l) denotes all parameters of layer l and l+1, W (l) is the weight parameter matrix of the l-th layer; for are defined for the three propagation graphs, respectively:

[0052]

[0053]

[0054]

[0055] The outputs of the multi-graph convolutional network layers corresponding to the four time series will enter the gated recurrent unit respectively to obtain the time features, and the gated recurrent unit is expressed by the following formula:

[0056] r t = σ(W r [GC(A 1 ,A 2 ,A 3 ,X t ), h (t-1) + b r )

[0057] z t = σ(W u [GC(A 1 ,A 2 ,A 3 ,X t ), h (t-1) + b N )

[0058]

[0059]

[0060] X t denotes the traffic propagation flow at time interval t, GC(A 1 ,A 2 ,A 3 ,X t ) denotes the multi-graph convolution process, that is, the output of the multi-graph convolution part will be input to the gated recurrent unit; σ denotes the sigmoid activation function, W r , W u , W h and b r , b N , b hThe weight parameters and bias parameters in the training process are the weight parameters and bias parameters

[0061] Preferably, in the traffic propagation matrix, 75% is used as a training set, and 25% is used as a test set; the evaluation is performed by the mean absolute percentage error (MAPE), and the evaluation formula is as follows:

[0062]

[0063] After each training round, the evaluation index of the model on the training set and the test set is evaluated, and the best test set performance is recorded. By comparing the evaluation indexes of different training rounds, the model at the best training round is retained;

[0064] The evaluation index value on the test set and its effect are as follows:

[0065] [0.50,0.80] excellent prediction effect

[0066] [0.80,1.00] better prediction effect

[0067] [1.00,1.20] general prediction effect

[0068] [1.20, + ∞] poor prediction effect

[0069] After achieving "excellent" and "better" results on the test set, the output is used; after achieving "general" and "poor" results on the test set, the data set is reorganized for model training.

[0070] The multi-graph convolution traffic propagation flow prediction system considering periodic fluctuation according to the application is used for the above method, comprising:

[0071] The matrix construction module identifies traffic propagation flow information and constructs a traffic propagation matrix;

[0072] The propagation graph construction module uses road network information and historical data to construct an adjacent propagation graph, a homogenous propagation graph and a same-destination propagation graph;

[0073] The time feature mining module mines the periodic fluctuation of traffic propagation flow based on the input time sequence;

[0074] The model construction module constructs a multi-graph convolution traffic propagation flow prediction model considering periodic fluctuation;

[0075] The training and evaluation module trains and evaluates the model.

[0076] Compared with the prior art, the application has the following beneficial effects:

[0077] (1) The application proposes a method for establishing a traffic propagation flow matrix, which can accurately reflect the evolution characteristics of the traffic propagation flow of the road network in the time and space dimensions; (2) The periodicity of traffic propagation flow information has certain volatility, and the application mines the volatility of the periodicity of traffic propagation flow based on the input time sequence, so that the model can have certain prediction of the volatility and improve the accuracy of the model; (3) The application designs an adjacent propagation graph, a homologous propagation graph and a same-destination propagation graph for traffic propagation flow, and the three graph structures are respectively used to describe the connection between different propagation relationships, and the spatial characteristics of traffic propagation flow information are mined through multi-graph convolution. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 A flowchart of a multi-graph convolution traffic propagation flow prediction method considering periodic volatility according to the application;

[0079] Figure 2 A framework diagram of a multi-graph convolution traffic propagation flow prediction model considering periodic volatility according to the application;

[0080] Figure 3 A framework diagram of a multi-graph convolution neural network according to the application. DETAILED DESCRIPTION

[0081] In order to more clearly illustrate the embodiments or technical solutions of the application, the following will briefly introduce the description of the embodiments or technical solutions. Obviously, the following description is only some embodiments of the application, and other embodiments can be obtained by those skilled in the art without creative labor on the basis of these embodiments.

[0082] Embodiment 1

[0083] As shown in the Figure 1 application, a multi-graph convolution traffic propagation flow prediction method considering periodic volatility is proposed, which mainly uses vehicle driving data, combines with road network information, and accurately predicts traffic propagation flow. It contains five parts, first, identify the traffic propagation flow information, and construct the flow propagation matrix. Second, use the road network information and historical data to construct the adjacent propagation graph, the homologous propagation graph and the same-destination propagation graph. Third, mine the volatility of the periodicity of traffic propagation flow based on the input time sequence. Fourth, construct a multi-graph convolution traffic propagation flow prediction model considering periodic volatility. Fifth, train and evaluate the model.

[0084] The following will introduce each part in detail:

[0085] First, identify the traffic propagation flow information, and construct the flow propagation matrix

[0086] Vehicle GPS positioning data includes information such as vehicle license plate number, latitude and longitude, time, and direction angle. This invention uses vehicle GPS positioning data to statistically analyze traffic information propagation flow between road segments. By using latitude, longitude, and direction angle, vehicles can be matched to corresponding road segments, revealing that a vehicle passed through a specific road segment at a specific time. Thus, traffic propagation flow information can be obtained. The following definitions are provided:

[0087] For a vehicle V, by matching its latitude and longitude to the corresponding road segments and sorting them by time, we can obtain the sequence of road segments the vehicle traversed throughout the day, i.e.:

[0088] LINK=link1→link2→link3→...→link n →...

[0089] There is also a corresponding time series:

[0090] T = t1→t2→t3→...→t n →...

[0091] In this sequence, link n Surrounding sequence points may be related to links n The road segments represented are the same; the most crucial aspect is identifying when the propagation occurred. The identification method of this invention is as follows:

[0092] After calculating the LINK for a specific vehicle, iterate through it. If in t n-1 Moment, link n-1 Not equal to link n , and t n -t n-1 <60s, then it is considered that at t n-1 At that moment, a road link occurred once. n-1 To the link n The spread of.

[0093] By statistically analyzing all the data, we can obtain the traffic propagation matrix Q between road segments at each time interval. N*H Where N represents the number of historical time intervals, and H represents the number of propagation relationships between road segments.

[0094] Traffic flow information calculation uses vehicle GPS positioning data to calculate traffic flow information between road segments and constructs a traffic flow matrix, which can reflect the process of traffic information propagation between road segments.

[0095] Second, using road network information and historical data, construct adjacency propagation graphs, source propagation graphs, and destination propagation graphs.

[0096] Define the traffic propagation transaction TPT as follows:

[0097] TPT = <link u ,link d Q ud >

[0098] Where link u , link d Q ud These represent the upstream segment, downstream segment, and link, respectively. u to link d The amount of dissemination.

[0099] 2.1 Propagation Adjacency Graph

[0100] Adjacent traffic propagation transactions are more relevant than traffic propagation transactions that are farther apart. Therefore, this invention uses an adjacency propagation graph to consider the relationships between adjacent traffic propagation transactions. If a link exists... a , link b , link c And the following relationship exists: link a →link b →link c At the same time, use TPT i Indicates link a →link b Traffic propagation transactions, TPT j Indicates link b →link c Traffic propagation transactions. Because TPT i Downstream section and TPT j The upstream sections are all links b Therefore, adjacency graph A 1 Corresponding position A ij The value is 1 if it is not 1, otherwise it is 0.

[0101] Therefore, the propagation adjacency graph can be defined as:

[0102]

[0103] 2.2 Homologous Propagation Diagram

[0104] Same origin refers to two traffic propagation transactions having the same upstream path, such as the existence of a path link. a , link b , link c And the following relationship exists: link a →link b , link a →link c At the same time, use TPTi represents link a → link b traffic propagation transaction, TPT j represents link a → link c traffic propagation transaction, TPT i and TPT j Because they have the same upstream road segment, assuming the number of vehicles on link a is constant, if TPT i increases, the number of TPT j will decrease, and vice versa. Therefore, the number of TPT i is small, and the number of TPT j is large, so the two traffic propagation transactions with the same upstream road segment are approximately negatively correlated with each other.

[0105] Therefore, the following definition is made for the homologous propagation graph:

[0106]

[0107] wherein Q i , Q j ∈ Q N*H , Q i represents the propagation relationship TPT i and the propagation traffic corresponding to each time interval. cor(Q i , Q j ) calculates the Pearson correlation coefficient between Q i and Q j . The larger the value is, the stronger the correlation between the two propagation transactions is.

[0108] The formula of the Pearson correlation coefficient is:

[0109]

[0110] wherein r represents the Pearson correlation coefficient between Q i and Q j . n represents the total number of time intervals. x k , y k represent the propagation traffic of Q i , Q j in the kth time interval, respectively. represents the average propagation traffic of Q i , Q j .

[0111] 2.3 Destination-oriented propagation graph

[0112] Similar to the homogenous propagation graph, the homed propagation graph refers to two traffic propagation transactions having the same downstream road segment. Traffic propagation transactions having the same downstream road segment also exhibit a certain negative correlation with each other.

[0113] Therefore, the homed propagation graph is defined as follows:

[0114]

[0115] Thus, the adjacent propagation graph, the homogenous propagation graph and the homed propagation graph are constructed.

[0116] Based on the multi-graph convolutional neural network, a multi-space feature mining model is constructed, and for the traffic propagation flow, a real road network structure and vehicle GPS positioning data are used to propose an adjacent propagation graph, a homogenous propagation graph and a homed propagation graph Figure 3 The topological graph is used for multi-graph convolution to mine multi-space features.

[0117] Thirdly, the periodic fluctuation of the traffic propagation flow is mined based on the input time series

[0118] Considering the periodicity and fluctuation of the traffic propagation flow information, the periodic fluctuation of the traffic propagation flow is mined based on the input time series. The time series input by the model mainly includes four parts: the recent time series, the previous day time series, the previous week time series and the previous month time series. The recent time series refers to the part of the time series closest to the prediction period, and the previous day time series, the previous week time series and the previous month time series refer to the time series near the previous day, the previous week and the previous month of the prediction period, rather than only including the time series before the previous day, the previous week and the previous month of the prediction period. This is to mine the periodicity of the traffic propagation flow and obtain the fluctuation of the periodicity of the traffic propagation flow information.

[0119] The recent time series, the previous day time series, the previous week time series and the previous month time series can be represented as follows:

[0120] X h =[X t-a ,X t-a+1 ,X t-a+2 ,....,X t-1 ]

[0121] X d =[X t-T-b ,X t-T-b+1 ,....,X t-T ,...,X t-T+b-1 ,X t-T+b ]

[0122] X w =[X t-7T-c ,Xt-7T-c+1 ,....,X t-7T ,...,X t-7T+c-1 ,X t-7T+c ]

[0123] X m =[X t-nT-e ,X t-nT-e+1 ,....,X t-nT ,...,X t-nT+e-1 ,X t-nT+e ]

[0124] Among them, X t ∈Q N*H , representing the number of propagations of all propagation relationships in the t-th time interval. 'a' represents the length of the most recent time series, and 'b', 'c', and 'e' represent the lengths of the time series taken before the corresponding time periods of the day, week, and month preceding the prediction period, respectively. That is, the lengths of the previous day's time series, the previous week's time series, and the previous month's time series are 2b+1, 2c+1, and 2e+1, respectively. 'T' represents the number of time intervals in a day. 'N' represents the number of days in the corresponding month.

[0125] For a, b, c, and e, larger values ​​allow the model to capture richer information and perform better, but also increase model parameters, slow down training, and increase model size. Therefore, in this invention, a five-minute time interval is used, a is 12, meaning the length of the most recent time series is 12, and b, c, and e are 2, 2, and 1 respectively, meaning the lengths of the previous day's time series, the previous week's time series, and the previous month's time series are 5, 5, and 3 respectively.

[0126] Based on the mining of time features of periodicity and periodic fluctuation, this invention selects time series near the prediction period for the input data, such as the time series of the previous day, the previous week, and the previous month, instead of only selecting time series before the prediction period. This is beneficial for mining the periodicity and periodic fluctuation of traffic transmission information.

[0127] Fourth, construct a traffic propagation flow prediction model that considers periodic fluctuations using multi-graph convolution.

[0128] like Figure 2 As shown, the four time series proposed in this invention, namely the most recent time series, the previous day time series, the previous week time series, and the previous month time series, will be mined for spatial features through multi-graph convolution, and then mined for temporal features through gated recurrent units. Finally, the results will be fused through a fully connected layer to obtain the prediction result.

[0129] like Figure 3As shown, the multi-graph convolutional neural network is composed of multiple graph convolutional neural networks and a full connection layer. The input time series are respectively subjected to graph convolutional neural networks and activation operations to extract features, and meanwhile, three kinds of propagation graphs are subjected to the same graph convolutional layer and activation operation to extract features, so that four spatial level features are obtained, which are further subjected to a full connection layer to fuse and learn spatial correlation features.

[0130] The general formula of multi-graph convolution can be defined as follows:

[0131]

[0132] Among them, denotes a degree matrix, H (l+1) and H (l) denote all parameters of l and l+1 layers, W (l) is a weight parameter matrix of the lth layer. For The present application defines three kinds of propagation graphs, which are respectively:

[0133]

[0134]

[0135]

[0136] The outputs of the multi-graph convolutional network layers corresponding to the four time series will respectively enter the gated recurrent unit to obtain time features. The gated recurrent unit can be expressed by the following formula

[0137] r t =σ(W r [GC(A 1 ,A 2 ,A 3 ,X t ),h (t-1) ]+b r )

[0138] z t =σ(W u [GC(A 1 ,A 2 ,A 3 ,X t ),h (t-1) ]+b N )

[0139]

[0140]

[0141] X t denotes the traffic propagation flow at the time interval t, GC(A 1A 2 ,A 3 ,X t ) represents a multi-graph convolution process, that is, the output of the multi-graph convolution part will be used as the input of the gated recurrent unit. σ represents the sigmoid activation function, W r , W u , W h and b r , b N , b h are weight parameters and bias parameters in the training process. The output of the gated recurrent unit is fused through the full connection layer to obtain the prediction result

[0142] The application proposes a traffic propagation flow prediction model based on space-time multi-graph convolution. The model adopts multi-graph convolution and gated recurrent unit. The multi-graph convolution layer is composed of multiple graph convolution neural networks and full connection layers. The input time series are respectively extracted through the graph convolution neural network and the activation operation, and at the same time, three propagation graphs are extracted through the same graph convolution layer and the activation operation, and four spatial level features are obtained.

[0143] Fifth, the model is trained and evaluated

[0144] 5.1 Data preprocessing

[0145] The data set used in the application is vehicle GPS positioning data, which contains vehicle license plate number, latitude and longitude, time, direction angle and other information. Using the data set and the method for establishing traffic propagation flow matrix proposed in the application, the flow propagation matrix Q N*H is constructed. Among them, 75% is used as the training set, and 25% is used as the test set. According to Q N*H and the road network topology information, the adjacent propagation graph, the homologous propagation graph and the same destination propagation graph are constructed.

[0146] 5.2 Training parameter setting

[0147] When training the model, the following parameter settings are adopted:

[0148] ·Training rounds: 300

[0149] ·Batch size: 32

[0150] ·Loss function: mean square error (MSE)

[0151] ·Optimizer: Adam

[0152] ·Learning rate: 0.001

[0153] The above parameters should be adjusted according to the actual situation to achieve the best performance.

[0154] 5.3 Model evaluation

[0155] To evaluate the performance of the model, we calculate the error between the predicted results and the true results on the training set and the test set, respectively. In this patent application, we choose the following evaluation indicators:

[0156] · Mean Absolute Percentage Error (MAPE)

[0157] The formula is as follows:

[0158]

[0159] After each training round, the model's evaluation indicators on the training set and the test set are calculated, and the best test set performance is recorded. By comparing the evaluation indicators of different training rounds, the model at the best training round is retained.

[0160] 5.4 Result analysis

[0161] After experiments and parameter adjustment, the model achieves good performance on the training set and the test set. The recommended evaluation indicators (MAPE) on the test set and their effects are as follows:

[0162] [0.50,0.80] Excellent prediction effect

[0163] [0.80,1.00] Good prediction effect

[0164] [1.00,1.20] General prediction effect

[0165] [1.20, +∞] Poor prediction effect

[0166] After achieving "excellent" and "good" results on the test set, it can be considered that the model has achieved the expected effect and can be output for use. After achieving "general" and "poor" results on the test set, it is recommended to reorganize the data set for model training.

[0167] The present application uses vehicle GPS positioning data to construct a traffic propagation flow prediction method based on spatiotemporal multi-graph convolution for the process quantity of traffic propagation flow. The present application has the following advantages: (1) the present application proposes a method for establishing a traffic propagation flow matrix, which can accurately reflect the evolution characteristics of traffic propagation flow in the space-time dimension; (2) the periodicity of traffic propagation flow information has certain volatility, and the present application mines the periodicity and volatility of traffic propagation flow based on input time series, so that the model can have certain prediction for the periodicity and volatility of traffic propagation flow, improving the accuracy of the model; (3) the present application designs adjacent propagation graph, homologous propagation graph, and same-destination propagation graph for traffic propagation flow, which are used to describe the relationship between different propagation relationships, and the spatial characteristics of traffic propagation flow information are mined through multi-graph convolution.

[0168] Embodiment 2

[0169] The application also provides a multi-graph convolution traffic propagation flow prediction system considering periodic fluctuation, which is used for the method in Embodiment 1, and comprises:

[0170] a matrix construction module, which identifies traffic propagation flow information and constructs a traffic propagation matrix;

[0171] a propagation graph construction module, which constructs an adjacent propagation graph, a homogenous propagation graph and a same-destination propagation graph by using road network information and historical data;

[0172] a time feature mining module, which mines periodic fluctuation of traffic propagation flow based on an input time sequence;

[0173] a model construction module, which constructs a multi-graph convolution traffic propagation flow prediction model considering periodic fluctuation;

[0174] a training and evaluation module, which trains and evaluates the model.

[0175] The above describes in detail a multi-graph convolution traffic propagation flow prediction method and system considering periodic fluctuation. The principle and implementation mode of the application are described by using specific examples, and the above description of the embodiments is only used to help understand the method and core idea of the application. It should be pointed out that, for those skilled in the art, some improvements and modifications can be made to the application without departing from the principle of the application, and these improvements and modifications also fall within the protection scope of the claims of the application.

Claims

1. A multi-graph convolutional traffic propagation flow prediction method considering periodic fluctuation, characterized in that, Comprise: First, identify the traffic propagation flow information, construct the traffic propagation matrix; Second, use the road network information and historical data to construct the adjacent propagation graph, the homologous propagation graph and the same destination propagation graph; Contiguous propagation graph The construction method is as follows: Consider the relationship between adjacent traffic propagation transactions with the adjacency propagation graph, if there is a road segment , , , and there is a relationship as follows: At the same time, let represent the traffic propagation transaction of , represent the traffic propagation transaction of ; Because the downstream road segment of and the upstream road segment of are both , therefore, the adjacency graph The value of the corresponding position is 1, otherwise, it is 0; The adjacent propagation graph is defined as: Defining traffic propagation transactions As follows: wherein , , respectively represent the upstream road section, the downstream road section and the propagation amount from to ; Homologous propagation chart The construction method is as follows: wherein, , represents a propagation relationship propagation flow corresponding to each time interval; calculate and the Pearson correlation coefficient between them; The greater, the stronger the correlation between the two propagation transactions; The formula of the Pearson correlation coefficient is: wherein r denotes the propagation flow and the Pearson correlation coefficient between; n denotes the total number of time intervals; , denote the propagation flow , in the kth time interval, , denote the average propagation flow of the propagation flow , . Destination propagation graph The construction method is as follows: ； Third, based on the input time series, the periodic fluctuation of the traffic propagation flow is mined; Fourth, a multi-graph convolution traffic propagation flow prediction model considering periodic fluctuation is constructed; The multi-graph convolution traffic propagation flow prediction model passes through multi-graph convolution to mine spatial features, respectively, and then passes through a gated recurrent unit to mine time features, respectively, and finally the results are fused through a fully connected layer to obtain the prediction results; The general formula of the multi-graph convolution is as follows: in, Degree matrix, and express and All parameters of the layer, For the first The weight parameter matrix of the layer; for Three types of propagation graphs are defined as follows: The outputs of the multi-graph convolution network layers corresponding to the four time series respectively enter the gated recurrent unit to obtain the time features, and the gated recurrent unit is represented by the following formula: denotes the traffic propagation flow at time interval t, denotes the multi-graph convolution process, i.e., the output of the multi-graph convolution part will be used as the input of the gated recurrent unit; denotes the sigmoid activation function, , , and , , are the weight parameters and bias parameters in the training process; the output of the gated recurrent unit is fused through the full connection layer to obtain the prediction result ; Fifth, the model is trained and evaluated.

2. The multi-graph convolutional traffic propagation flow prediction method considering periodic fluctuations according to claim 1, characterized in that, The traffic propagation flow information is obtained by vehicle GPS positioning data, which can match vehicles to corresponding road segments according to latitude, longitude and direction angle, and obtain traffic propagation flow information by knowing that a vehicle passes through a specific road segment at a specific time; For a vehicle V, after matching the latitude and longitude to the corresponding road segment and sorting by time, the road segment sequence information passed by the vehicle in a day can be obtained, that is: There is a corresponding time series: In this sequence, if the sequence point around the same section as represented, it is identified when the propagation occurs, and the identification method is as follows: The number of times of the propagation of the road segment to the road segment is calculated for each vehicle . If the time is not equal to the time , it is considered that the propagation of the road segment to the road segment occurred once at the time Statistical all data, namely, get each time interval road section and road section between the traffic propagation matrix , wherein represents the number of historical time intervals, represents the number of propagation relationships between road sections. 3.The multi-graph convolutional traffic propagation flow prediction method considering periodic fluctuation of claim 1, wherein, The periodic fluctuation of the traffic propagation flow is mined based on the input time series, which is represented by the following formula: wherein, , represents the propagation number of all propagation relationships in the tth time interval; a represents the length of the latest time series, b, c, e respectively represent the length of the time series taken before the corresponding period of the day before, the week before and the month before the prediction period, that is, the lengths of the day before time series, the week before time series and the month before time series are 2b+1, 2c+1, 2e+1 respectively; T represents the number of time intervals in a day, and N represents the number of days in the corresponding month.

4. The multi-graph convolutional traffic propagation flow prediction method considering periodic fluctuations according to claim 1, characterized in that, In the traffic propagation matrix, 75% is used as the training set and 25% is used as the test set; the mean absolute percentage error MAPE is used for evaluation, and the evaluation formula is as follows: After each training round, the evaluation index of the model on the training set and the test set is obtained, and the best test set performance is recorded; by comparing the evaluation indexes of different training rounds, the model at the best training round is retained; The evaluation index values on the test set and their effects are as follows: [0.50,0.80] excellent prediction effect [0.80,1.00] good prediction effect [1.00,1.20] general prediction effect [1.20,+∞] poor prediction effect After achieving "excellent" and "good" results on the test set, the output is used; after achieving "general" and "poor" results on the test set, the data set is reorganized for model training.

5. A multi-graph convolutional traffic propagation flow prediction system considering periodic fluctuation, for the method of any one of claims 1-4, characterized in that, Include: The matrix construction module identifies the traffic propagation flow information and constructs the traffic propagation matrix; The propagation graph construction module uses the road network information and historical data to construct the adjacent propagation graph, the homologous propagation graph and the same destination propagation graph; The time feature mining module mines the periodic fluctuation of the traffic propagation flow based on the input time series; The model construction module constructs a multi-graph convolution traffic propagation flow prediction model considering periodic fluctuation; The training and evaluation module trains and evaluates the model.

Images