CN115409256B - Route recommendation method for avoiding congestion area based on travel time prediction - Google Patents

Abstract

The present invention provides a route recommendation method based on travel time prediction to avoid congestion areas, including three steps: a DGCN-based road segment travel time prediction model, a road network area state recognition model based on travel time prediction, and a route recommendation model based on the Floyd algorithm. The present invention is based on the route recommendation model of travel time prediction and congested area avoidance, based on historical checkpoint data and road network topology structure, uses dynamic graph neural network (DGCN) to carry out short-term prediction on road section travel time of urban road network, and combines Based on the dynamic actual conditions and traffic characteristics of the road network, a road network partition model based on the weighted GN algorithm is established, and the regional state of the road network is identified according to the predicted travel time results; finally, combined with the predicted travel time and the congestion status of the road network, based on the Floyd algorithm Recommendations for avoidance routes can, to a certain extent, reduce the degree of congestion on congested roads, provide travelers with shortest route recommendations, and improve travel experience.

Description

Route recommendation method for avoiding congestion areas based on travel time prediction

Technical Field

The invention belongs to the technical field of intelligent transportation and relates to a path recommendation method for avoiding congestion areas based on travel time prediction.

Background Art

As the number of cars increases year by year, traffic congestion in some first-tier cities has spread from the central urban areas to the peripheral urban areas, and the reliability of vehicle travel time is also decreasing with the uncertainty of traffic congestion. Route recommendation is one of the most widely used location-based services today. It is essential for a good driving experience and smooth public transportation. It can provide users with the best travel experience based on given standards. Therefore, travelers and traffic managers are also very concerned about the problem of vehicle travel paths.

The problem of vehicle travel path selection is the focus of intelligent transportation system research. The acquisition of dynamic traffic information is also the premise of travel time prediction and optimized path recommendation. Yang Feng et al., for the problem of emergency vehicle path selection, combined with dynamic road network and emergency rescue, established a two-stage optimization model with maximum path reliability as the first goal and shortest travel time as the second goal, and designed a hybrid cuckoo search algorithm to narrow the global search space of the algorithm and improve the global optimization ability of the optimization model. In order to reduce the operating cost and path cost of AGV, an airport AGV path optimization model based on Dijkstra ant colony optimization algorithm (ACO-DA) was proposed. Considering the environment with obstacles, the baggage collection sorting was carried out through ant colony optimization, and the AGV path planning was integrated with the Dijkstra algorithm. Wang Chaoxiong et al. proposed a neural network model based on A* fastest path recommendation. The traffic prediction model for route travel time prediction and the multi-task representation learning model for OD travel time estimation were combined with the A* algorithm. High-precision travel time prediction can be achieved without actual path information, and the fastest real-time route can be found. In order to avoid ship collision, Li Ning established an optimization model for ship collision avoidance path planning based on particle swarm optimization algorithm, genetic algorithm and nonlinear programming theory. Liu Binbin; analyzed the research status of logistics distribution path determination scheme, and applied recursive fuzzy neural network algorithm to select e-commerce logistics distribution path scheme. Dengkai HOU aimed at the multi-path dynamic multi-yard refrigerated vehicle path problem based on real-time traffic information. Based on the idea of pre-optimization and real-time adjustment, a two-stage optimization model with the goal of minimizing the total cost was established, and a hybrid chaos genetic algorithm with variable neighborhood search (HCGAVNS) was designed to generate the initial path. An Dongdong et al., based on the actual situation of humpback transportation, used the traffic network diagram to simulate the relationship between nodes and routes in the road network, and considered the time cost, expense cost and traffic environment impact cost for function analysis; in terms of algorithm, the transportation cost was added to the virtual route, the shortest path Floyd algorithm was optimized, and the optimal path was obtained, which provided reference ideas for the path selection of intermodal enterprises. Yu Quan et al. proposed a multi-objective path selection model based on the shortest time, shortest distance and lowest congestion. Starting from the two aspects of path selection and path planning, combined with the real-time feedback of road conditions in the system, real-time multi-objective dynamic re-planning of travel paths is realized. Wang Jingyuan et al. integrated the cost from the starting point to the candidate location and the cost from the candidate location to the destination using the value network based on attention RNN network modeling, proposed to use neural networks to automatically learn the cost function of the classic heuristic algorithm, and used the A* algorithm in the PRR task to obtain a more accurate cost for searching candidate locations. Zhilu Pingping studied random dynamic road networks from a microscopic level, established a road section travel time reliability calculation model, and proposed a path selection algorithm using travel time reliability as a key control variable for the vehicle path selection problem of simple and complex networks to improve the efficiency of path selection. Zhang Chuanqi et al. divided the road network into different time periods according to the degree of congestion based on the traffic characteristics of vehicles in the road network, considered the characteristics of vehicle service order, path and optimal travel time, established a dynamic congested distribution path optimization model, and combined with an improved genetic algorithm to solve the objective function of the minimum distribution cost. To solve the problem that it is difficult to utilize useful contextual information during the path search process, Ying Shen et al. integrated the improved DBSCAN algorithm into hotspot extraction and the density-based ε-distance trajectory clustering algorithm to identify recommended potential paths. Factors such as driving time, speed, distance, and endpoint attractiveness defined a weighted tree to help cruising taxi drivers find potential passengers with the best route.

In summary, for the problem of vehicle path selection in urban road networks, scholars have mostly studied it from the perspectives of travel time, travel cost or road network travel time reliability, and have not given enough consideration to the prediction of dynamic road network travel time and the fact that the shortest path obtained by the shortest path algorithm is not necessarily the optimal one. The present invention fully considers the real-time changes of the dynamic travel time of road network traffic in small and medium-sized cities, partitions the road network by traffic weights through the weighted GN algorithm, and predicts the travel time with a dynamic graph neural network. The regional status is identified based on the predicted travel time results combined with the average travel speed. Finally, based on the predicted travel time and the congested area identification results, the Floyd algorithm is used to recommend paths that avoid congested areas.

Summary of the invention

The purpose of the present invention is to provide a route recommendation method for avoiding congestion areas based on travel time prediction, so as to solve the problems raised in the above background technology.

The object of the present invention can be achieved by the following technical scheme: a route recommendation method for avoiding congested areas based on travel time prediction, characterized in that it includes three steps: a road section travel time prediction model based on DGCN, a road network area state recognition model based on travel time prediction, and a route recommendation model based on Floyd algorithm, which are specifically as follows:

The road section travel time prediction model based on DGCN includes a Laplace matrix latent network module and a traffic prediction module based on a graph convolutional network, where the Laplace matrix prediction unit consists of three parts:

Feature sampling: Sampling the data of the nearest neighbor 15min, 30min, and 60min every day, sampling the nearest traffic while reducing the data dimension of the feature; Spatial attention mechanism: In order to establish the spatial relationship of the dynamic traffic road network, the attention mechanism is used to estimate the adjacency matrix of the current neighbor time of the road network; LSTM unit: LSTM is used to learn time correlation and explore the intrinsic correlation between the sequences of the adjacency matrix Ld;

The GCN-based traffic prediction module consists of two parts:

(1) Temporal convolution layer: It mainly includes four two-dimensional convolution layers with a convolution kernel of 1x3, which are used to extract high-dimensional local time in the data. The temporal convolution of the traffic data segment X(1:k*T)=(X1........Xt......Xk*T) is: TC=Conv _1×t (X(1:k*T));

(2) Graph-time convolutional layer: GCN and TCL are stacked into a spatiotemporal module, but there is a problem of more calculations. This problem can be well solved by integrating these two functions into a GTCL. The GTCL after replacement can be expressed as:

Road network regional state recognition model based on travel time prediction: The GN algorithm in the community partition algorithm based on hierarchical partitioning of complex networks is adopted. Community discovery refers to the directed connection of each node in a complex network, and all nodes are divided into communities according to the partition judgment. Nodes in the same community have similar characteristics. The GN algorithm is more efficient. The directed relationship of all nodes and edges is taken into account in the network partitioning. The partitioning results of the experimental results are hierarchical. The judgment standard of the road network partitioning results is modularity. The larger the modularity, the more obvious the community structure of the partition. The modularity Q is defined as:

Where: a _vw is the weight of the edge between nodes v and w, k _w is the degree of node w;

According to the edge attributes in the complex network, that is, the edge value in the GN algorithm, the flow attributes of the road sections in the road network are weighted. When the maximum edge betweenness is removed, especially when the edge betweenness is equal, the weight of the corresponding edge is obtained according to the flow ratio between the edges in the current time period, thereby obtaining the edge weight ratio, and the correlation of the edges in the road network can be obtained. The obtained community division result is better and the correlation between two edges in the actual road network is more reliable.

Path recommendation model based on Floyd algorithm: Floyd algorithm is used to solve the shortest path algorithm between any two points, and to solve the shortest path problem of directed or undirected graphs. The core of the algorithm is to solve the global optimum through local optimum and perform dynamic planning. First, find the shortest path length of the target OD, and then record the path of this length to find the recommended route. Remove the sections in the path that are in a severely congested state, and repeat the above steps to find the recommended path matrix and routing matrix based on congestion avoidance;

The idea of the Floyd algorithm based on congestion avoidance is:

(1) Initial state: Based on the travel time data predicted by DGCN, write the initial distance-based adjacency matrix W and the initial routing matrix R ₀ = [ _{ri j} ] _n×n of the graph G. For each pair of vertices v _i and v _j , if there is an edge from v to v, then the length is the weight of the edge. If there is no edge, the length is set to infinity Q;

(2) k = 0: that is, for each pair of vertices _vi and _vj , the subscript of the passing vertex is not greater than k. In fact, only _vo can be passed here. The path can be divided into two segments, namely ( _vi , _vo ) and ( _vo , _vj ). The length of this path is the sum of the lengths of the two paths. By comparing this new path with the previous path ( _vi , _vj ), we can determine the shortest path from _vi to _vj with a subscript not greater than k.

(3) k = 1: Similarly, the path can be split into two segments (v _i , …, v _k ) and (v _k , …, v _j ). The lengths of these two segments are determined when k = 0. By comparing the new path with the previously known path (v _i , v _j ), the shortest path with an index not greater than k can be determined. At this time, k = 1;

(4) Repeat the above steps until k = n-1, at which point all possible shortest travel time paths from _vi and _vj and the corresponding routing matrices have been determined;

(5) Select some fixed start-end paths, remove the edges containing heavily congested sections, form a new road network graph G’, repeat steps (1)-(4), obtain the path distance matrix and routing matrix based on congestion avoidance, and obtain all the shortest travel time paths after avoidance.

In the above-mentioned congestion area avoidance route recommendation method based on travel time prediction, the specific process of the weighted GN algorithm is as follows:

Step 1: In the initial state, each node is regarded as an independent community, that is, the number of communities in this state is the same as the number of nodes;

Step 2: Ignore the weights of the edges between two nodes in the data, and calculate the edge betweenness of each edge using the unweighted network calculation method;

Step 3: Divide the edge betweenness by the weight of the corresponding edge to get the edge weight ratio;

Step 4: Remove the edge with the largest edge-weight ratio and calculate the modularity Q of the current network;

Step 5: Repeat steps 2 to 4 for the remaining edges, and calculate the edge weight ratio and modularity of each step until all edges in the network are removed;

Step 6: After the operation is completed, the number of community divisions and the partition results corresponding to the maximum Q are obtained.

Compared with the prior art, the advantages of the route recommendation method for avoiding congested areas based on travel time prediction of the present invention are: a route recommendation model based on travel time prediction and congested area avoidance. Based on historical checkpoint data and road network topology, a dynamic graph neural network (DGCN) is used to make short-term predictions on the travel time of sections of the urban road network, and by combining the actual dynamic conditions of the road network and traffic characteristics, a road network partitioning model based on the weighted GN algorithm is established, and the regional status of the road network is identified according to the predicted travel time results; finally, based on the predicted travel time and the congestion status of the road network, avoidance routes are recommended based on the Floyd algorithm. Taking the road network of a certain area in a certain city as an example, the checkpoint data and road network GIS-T data of the area are applied to the model. The analysis results show that the route recommendation for vehicles to avoid congested areas takes into account the global road network, which, to a certain extent, reduces the congestion level of congested sections, provides travelers with the shortest route recommendation, and improves the travel experience.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of a route recommendation model based on travel time prediction and congestion avoidance according to the present invention.

FIG. 2 is a heat map of travel time predicted for the next 12 time periods according to the present invention.

FIG3 is a schematic diagram of the road network adjacency moment result based on travel time according to the present invention.

FIG. 4 is a diagram of the 39th round path adjacency matrix (a) and a diagram of the 39th round routing matrix (b) of the present invention.

FIG. 5 is a schematic diagram of actual checkpoint locations of the road network of the present invention.

DETAILED DESCRIPTION

The following are specific embodiments of the present invention and the accompanying drawings to further describe the technical solution of the present invention, but the present invention is not limited to these embodiments.

As shown in Figure 1, the road section travel time prediction model based on DGCN:

基于行程时间预测的路网区域状态识别模型More accurate prediction of travel time is the basis for route recommendation. Sampled graph neural network can predict travel time, especially for unstructured data such as traffic data, which can well solve the correlation between two nodes and the learning problem of complex structure and dynamic performance of road network. Therefore, the dynamic graph convolutional network (DGCN) model is used to predict travel time. The model mainly includes: Laplacian matrix latent network (LMLN) module and traffic prediction module based on graph convolutional network (GCN). The Laplacian matrix prediction unit (LMLN) consists of three parts: (1) Feature sampling: sampling the data of the nearest neighbor 15min, 30min, and 60min every day, sampling the nearest traffic while reducing the data dimension of the feature; (2) Spatial attention mechanism: in order to establish the spatial relationship of the dynamic traffic network, the attention mechanism is used to estimate the adjacency matrix of the current neighbor time of the road network; (3) LSTM unit: LSTM is used to learn time correlation and explore the intrinsic correlation between the sequences of the adjacency matrix Ld. The traffic prediction module based on GCN includes: (1) Temporal convolution layer: It mainly includes four two-dimensional convolution layers with convolution kernels of (1x3), which are used to extract high-dimensional local time in the data. The temporal convolution of the traffic data segment X(1:k*T)=( _X1 ........ _Xt. .....Xk _*T ) is: TC=Conv1xt(X(1:k*T)). (2) Graph temporal convolution layer: GCN and TCL are stacked into a spatiotemporal module, but there is a problem of more calculations. This problem can be well solved by integrating these two functions into a GTCL. The GTCL after replacement can be expressed as:

Road network area status identification model based on travel time prediction

The traffic network has the structure and characteristics of a complex network. The models for dividing the road network include grid division areas, division according to the road network, division according to the density of the research object, and division according to the reference point. This study adopts the GN algorithm in the community division algorithm based on hierarchical division of complex networks. Community discovery refers to the directed connection of each node in a complex network, and all nodes are divided into communities according to the division judgment. Nodes in the same community have similar characteristics. The GN algorithm is more efficient and takes the directed relationships of all nodes and edges into account in the network division. The division results of the experimental results are hierarchical. The criterion for judging the results of road network partitioning is modularity. The larger the modularity, the more obvious the structure of the divided community. The definition of modularity is:

Where: a _vw is the weight of the edge between nodes v and w, k _w is the degree of node w;

According to the edge attributes in the complex network, that is, the edge value in the GN algorithm, the flow attributes of the road sections in the road network are weighted. When the maximum edge betweenness is removed, especially when the edge betweenness is equal, the weight of the corresponding edge is obtained according to the flow ratio between the edges in the current time period, thereby obtaining the edge weight ratio, and the correlation of the edges in the road network can be obtained. The obtained community division result is better and the correlation between two edges in the actual road network is more reliable.

The specific process of the weighted GN algorithm is as follows: ① In the initial state, each node is regarded as an independent community, that is, the number of communities is the same as the number of nodes in this state; ② Ignore the weights of the edges composed of two nodes in the data, and calculate the edge betweenness of each edge according to the unweighted network calculation method; ③ Divide the edge betweenness by the weight of the corresponding edge to obtain the edge weight ratio; ④ Remove the edge with the largest edge weight ratio and calculate the modularity Q of the current network; ⑤ Repeat steps ② to ④ for the remaining edges, and calculate the edge weight ratio and modularity of each step until all edges in the network are removed; ⑥ At the end of the operation, take the number of community divisions and partition results corresponding to the maximum Q.

Method for calculating regional average velocity:

(1) According to the evaluation method, the average travel speed V _i of a certain road section is calculated as follows:

Where: L is the length of the road section in km; t is the time it takes for vehicle i to pass through the section in h; n is the number of measured vehicles.

(2) Calculate the average travel speed of a road network or a certain area:

Where: V ₁ ........V _n represents the average travel speed of the road network, in km/h, and n is the number of road sections.

Route recommendation model based on Floyd algorithm

According to the needs, the optimal path problem can be divided into the shortest distance and the minimum weight. Among them, the shortest distance method can be divided into free path and restricted path. This study uses the Floyd algorithm, which is an algorithm to solve the shortest path between any two points. It can solve the shortest path problem of directed or undirected graphs. The core of this algorithm is to solve the global optimum through local optimum and perform dynamic planning. First, find the shortest path length of the target OD, and then record the path of this length to find the recommended route, remove the sections in the path that are in a severely congested state, and repeat the above steps to find the recommended path matrix and routing matrix based on congestion avoidance.

The idea of the Floyd algorithm based on congestion avoidance is:

(1) Initial state: Based on the travel time data predicted by DGCN, write the initial distance-based adjacency matrix W and the initial routing matrix R ₀ = [ _{ri j} ] _n×n of the graph G. For each pair of vertices _vi and v _j , if there is an edge from v to v, then the length is the weight of the edge. If there is no edge, the length is set to infinity Q.

(2) k = 0: that is, for each pair of vertices _vi and _vj , the subscript of the vertex passed through is not greater than k. In fact, it can only pass through _vo . The path can be divided into two segments, namely ( _vi , _vo ) and ( _vo , _vj ). The length of this path is the sum of the lengths of the two paths. By comparing this new path with the previous path ( _vi , _vj ), we can determine the shortest path from _vi to _vj with a subscript not greater than k.

(3) k = 1: Similarly, the path can be divided into two segments: (v _i ,…, v _k ) and (v _k ,…, v _j ). The lengths of these two segments are determined when k = 0. By comparing the new path with the previously known path (v _i , v _j ), the shortest path with an index no greater than k can be determined. At this time, k = 1.

(4) Repeat the above steps until k = n-1, at which point all possible shortest travel time paths from _vi and _vj and the corresponding routing matrices have been determined.

(5) Select some fixed start-end paths, remove the edges containing heavily congested sections, form a new road network graph G’, repeat steps (1)-(4), obtain the path distance matrix and routing matrix based on congestion avoidance, and obtain all the shortest travel time paths after avoidance.

Embodiment 1

Travel time prediction based on DGCN

3.1.1 Dataset Introduction

The dataset used is from the public checkpoint dataset in the central urban area, which is organized and released by the openITS platform. The total road length in the area is 174km, including the checkpoint vehicle passing data and checkpoint GIS-T distribution data from December 3 to 9, 2017. There are about 830,000 vehicle passing data in one day in the checkpoint vehicle passing data. After deleting the unrecognized license plate numbers, the data volume is about 740,000. After screening according to the obtained section travel time, the data volume is about 560,000. The specific fields are shown in the following table. The daily travel time data contains a traffic feature, namely the section travel time. The predicted travel time is used as the output in the model; the time slot is divided into 15 minutes, and the data of each day contains 96 time slots. There are 4 samples of all data per hour, so T=4 is set, and 60% of the data set is selected as the training set, 20% as the validation set, and 20% as the test set. At the same time, the best model parameters on the validation set are selected, and all models are evaluated on the test set. The data samples of each section are normalized.

Parameter settings

The experiment was applied to the Linux operating system ubuntu18.04, the GPU was RTX3090*2, the video memory was 48G, the code was written in python3.7, and the model architecture was developed based on pytorch1.2.0 deep learning.

Four 1×3 convolution kernels are used in graph convolution and temporal convolution, the prediction time step c is 12, the learning rate is set to 0.0005, the decay rate of each round is 0.92, the batch size is 8, l2_loss is used as the loss function, and Adam is used as the optimizer for optimization. In GTCL, the Chebyshev polynomial M=3, the convolution kernel ts=3, the multi-head attention mechanism k=4, and the prediction time interval Tp=12, that is, one hour in the future, that is, we use 60% of the samples for training and predicting 12 samples.

Travel time prediction results

To estimate the impact of different inputs, we also built our input using the nearest neighbor data, denoted by DGCN_R. To further evaluate the efficiency of different Laplacian matrices for GCN, especially GAT, we compared our method with four GCN-based methods: (1) ASTGCN, which uses the attention Laplacian matrix; (2) DGCN_Mask, which uses the Mask Laplacian matrix; (3) GCN_Res, which uses the residual Laplacian matrix; (4) DGCN_GAT, a method that replaces the spatial feature layer GTCL of the model with GAT, thus making it comparable to other methods. The performance of all methods is measured by three metrics, mean absolute percentage error (MAPE), mean absolute error (MAE), and mean square error (RMSE), as follows:

The average one-hour traffic prediction accuracy on the dataset is shown in the following table. The results show that the DGCN_Res model has the best performance in all indicators. The Laplace latent matrix can better extract the dynamic spatial relationship of the road network. At the same time, the accuracy of DGCN_Res is better than other models because the empirical Laplace matrix is used instead of the globally optimized residual Laplace matrix. The results of predicting the next 12 periods, the prediction accuracy results

Results of road network partitioning and regional status recognition using the weighted GN algorithm

The experimental data set comes from the checkpoint GIS-T distribution data, which includes 39 checkpoints. Combined with the travel time data, 60 road sections are selected as edges. The traditional GN algorithm and the weighted GN algorithm are used to divide the urban road network. The number of partitions and the corresponding modularity are shown in the following table. For the traditional GN algorithm and the weighted GN algorithm, as the number of partitions increases, the modularity gradually decreases, indicating that in the actual road network, the correlation between the edges is also decreasing. Combined with the conditions for avoiding congested areas and the correlation of road sections in the actual road network, the number of partitions is selected as 15, and the partition results are: [['11','1','2','12','10','26'],['35 ','31','3'],['7','4','13','14','9'],['5','36','34'],['39','24','8','6'],['30','16','15'],['18','17'],['21','28','29','19'],['20'],['22'],['27','23'],['25','33'],['32'],['37'],['38']], and according to the partition results, the area is divided into G1-G15 in sequence to represent the congestion area identification results in the next section.

According to the results of travel time prediction in the previous section, the travel time data of 60 road sections can be obtained. Combined with the results of road network partitioning and the steps of regional status identification, the traffic status of some regions is shown in the following table:

Route recommendation results for avoiding congestion areas based on Floyd algorithm

The road network adjacency matrix based on travel time in the initial state is shown in Figure 3 below. The weight between two points represents the travel time of the road section predicted by DGCN in the next 15 minutes, which is recorded as the value of v(i,j). v(i,j) is recorded as ∞, and its weight is set to 0. Traverse v(1,j) and v(i,1) in turn, compare the sizes of v(1,j), v(i,1) and v(i,j), and replace if v(1,j)+v(i,1)<v(i,j). Traverse the matrix 39 times, record the matrix results each time, and finally obtain the adjacency matrix of the shortest travel time and the corresponding routing matrix as shown in Figures 4(a), (b) and 5.

The shortest path is not necessarily the best. Only considering the shortest path for path recommendation will cause congestion in the recommended path for the global road network architecture. If the path recommendation for avoiding congested areas is performed for some vehicles, the traffic pressure in the congested areas can be alleviated to a certain extent. Taking the road network of Xuancheng City, Anhui Province as an example, the traffic feature of flow is used as the weight, and the edge weight ratio is formed by combining the edge betweenness, and the weighted GN algorithm is used to divide the area according to the actual road network situation; the historical checkpoint data is applied to the model, and the Laplace potential matrix of the nearest neighbor data is introduced for sampling the nearest neighbor 15min, 30min, and 45min every day, so as to better learn the dynamic structure of the road network, and the prediction of travel time is 18.7% and 30% higher than the MAE and RMSE indicators of Pems04; at the same time, the shortest path between each checkpoint is obtained and the shortest path result after avoiding the congested area is compared, and it is concluded that the path recommendation for avoiding the congested area is less in the travel time required for travel. The content not described in detail in this specification belongs to the prior art known to professional and technical personnel in this field. The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims (2)

1. The route recommendation method for avoiding the congestion area based on travel time prediction is characterized by comprising three steps of a road section travel time prediction model based on DGCN, a road network area state identification model based on travel time prediction and a route recommendation model based on Floyd algorithm, and specifically comprising the following steps of:

the road section travel time prediction model based on the DGCN comprises a Laplacian matrix submarine network module and a traffic prediction module based on a graph rolling network, wherein: the laplacian matrix prediction unit includes three parts:

(1) Feature sampling: sampling data of 15min,30min and 60min nearest to each other every day, and sampling the nearest traffic while reducing the data dimension of the features;

(2) Spatial attention mechanism: in order to establish a space relation of a dynamic traffic network, estimating an adjacent matrix of the current neighbor time of the road network by adopting an attention mechanism;

(3) LSTM unit: learning a time correlation by using LSTM, and exploring an internal correlation between sequences of the adjacency matrix Ld;

the GCN-based traffic prediction module includes two parts:

(1) Time convolution layer: comprising four two-dimensional convolution layers with a convolution kernel of 1X3, for extracting high-dimensional local time in the data, traffic data segment X (1:kχt) = (X1.. Xt...a. Xk X T) is:

TC＝Conv _1×t (X(1:k*T))(1)

TC＝Conv _1×t (X (1: kT)) is a time convolution

(2) Graph time convolution layer: stacking GCN and TCL as space-time modules, but there is a problem of large calculation amount, this problem can be solved well by integrating these two functions into one GTCL, and the GTCL after replacement can be expressed as:

road network area state identification model based on travel time prediction: the community discovery refers to directed connection of all nodes in the complex network by adopting a GN algorithm in a community division algorithm based on a hierarchical division complex network, all nodes in the same community are divided into communities according to division judgment, the nodes in the same community have similar characteristics, the GN algorithm is high in efficiency, the directed relationship between all nodes and edges is considered in network division, the division result of an experimental result has hierarchy, the judgment standard of the road network division result is modularity, the greater the modularity is, the more obvious the divided community structure is, and the definition of the modularity Q is as follows:

wherein: a, a _vw Is the weight of the edge between the nodes v and w, k _w The degree of node w;

according to the edge attribute in the complex network, namely the edge value in the GN algorithm, after weighting the flow attribute of the road section in the road network, when the maximum edge betweenness is removed, when the edge betweenness is equal, the weight of the corresponding edge is obtained according to the flow ratio between the edges in the current time period, so that the edge weight ratio is obtained, the relevance of each edge in the road network can be obtained, the obtained community division result is better, and the relevance between the community division result and the two edges in the actual road network is more reliable;

path recommendation model based on Floyd algorithm: the Floyd algorithm is also called an insertion point method, the shortest path algorithm between any two points can be solved, the problem of the shortest path of a directed graph or an undirected graph is solved, the core of the algorithm is that global optimization is solved through local optimization, dynamic planning is carried out, the shortest path length of a target OD is found out firstly, then the path of the length is recorded, a recommended route can be found, the road sections in the severe congestion state area in the path are removed, the algorithm is repeated, and a recommended path matrix and a route matrix based on congestion area avoidance can be found;

the idea of the Floyd algorithm based on congestion area avoidance is that:

(1) In the initial state, according to the travel time data predicted by DGCN, writing out an adjacent matrix W and an initial route matrix R based on the distance of the initial graph G ₀ ＝[r _ij ] _n×n For each pair of vertices v _i And v _j If v _i And v _j If there is an edge, the edge is weighted asRoad segment length; if no edge exists, setting the length to be infinite;

(2) k=0, i.e. for each pair of vertices v _i And v _j The subscript of the path vertex is not greater than k, where in practice only v can pass _o The path may be divided into two segments, i.e. (v) _i ，v _o ) Sum (v) _o ，v _j ) This length is the sum of the lengths of the two paths, and the new path is compared with the previous path (v _i ，v _j ) V can be determined _i To v _j A shortest path with a path index not greater than k;

(3) k=1, and the path is detachable (v) _i ,……，v _k ) Sum (v) _k ，..，v _j ) Two sections whose length is determined when k=0, and comparing the new path with the previously known path (v _i ，v _j ) The shortest path with a path index no greater than k can be determined, where k=1;

(4) Repeating the above steps until k=n-1, at which time the slave v has been determined _i And v _j All possible shortest travel time paths and corresponding routing matrices;

(5) Selecting a part of paths with fixed starting and ending points, removing edges of the paths containing the severely congested road sections to form a new road network graph G', and repeating the steps (1) - (4) to obtain a path distance matrix and a route matrix based on congestion area avoidance, so as to obtain all the shortest travel time paths after the avoidance.

2. The route recommendation method for congestion area avoidance based on travel time prediction according to claim 1, wherein the specific flow of the weighted GN algorithm is as follows:

the first step: in the initial state, each node is regarded as an independent community, namely the number of communities is the same as the number of nodes in the state;

and a second step of: neglecting the weight of the edges formed by two nodes in the data, and obtaining the edge betweenness of each edge according to an unauthorized network calculation method;

and a third step of: dividing the edge betweenness by the weight of the corresponding edge to obtain an edge weight ratio;

fourth step: removing the edge with the largest edge weight ratio, and calculating the modularity Q of the current network;

fifth step: repeating the second step to the fourth step for the rest edges, and calculating the edge weight ratio and the modularity of each step until all edges in the network are removed;

sixth step: and (5) after the operation is finished, taking the corresponding community division number and partition result when Q is maximum.

Images