Large-scale truck platoon scheduling method and device based on depth map learning

Abstract

The application discloses a large-scale truck platoon scheduling method and device based on depth map learning, belongs to the technical field of intelligent traffic scheduling, and solves the problems that the truck platoon scheduling algorithm in the prior art often has high calculation complexity and unstable solution quality when facing large-scale scheduling. In the application, the interaction information of a single truck at a hub with potential platoon objects is represented as a node-edge model, the coupling relationship between time and space in the platoon strategy is captured, compared with a traditional method, the quality of the platoon scheduling strategy is improved, more trucks are provided with platoon opportunities, and the total system cost is effectively reduced; meanwhile, based on an end-to-end graph neural network model for the large-scale truck platoon problem, the generation efficiency of the platoon scheduling strategy is effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent traffic scheduling, and in particular relates to a large-scale truck platoon scheduling method and device based on deep map learning. BACKGROUND

[0002] Highway freight transport, i.e. highway truck freight transport, is a transport mode that realizes spatial displacement of goods through highways by taking trucks as the main tool, covering ordinary freight transport, special transport, large object transport, and dangerous goods transport. Today, highway freight transport accounts for more than 32% of global transport energy consumption and more than one-third of current carbon emissions in the transportation industry, and its sustainable development faces severe challenges. Due to the wind resistance design problem of trucks, nearly 60% of the energy consumption of a single high-speed heavy truck is used to resist wind resistance, causing great energy waste. How to effectively reduce truck energy consumption while ensuring transport efficiency has become a real bottleneck problem that needs to be solved. Therefore, in the actual transport production process, truck platoon technology is used to form a vacuum zone by controlling the constant speed and small spacing of trucks in the queue, reducing the wind resistance of the following vehicles, and significantly reducing energy consumption. Experimental results show that the air resistance of the following vehicles in the truck platoon can be reduced by 15%-22%, and the overall platoon operation can reduce 8%-15% of the energy consumption.

[0003] Currently, the platoon scheduling problem is usually transformed into an integer linear programming problem (MILP) or an integer nonlinear programming problem (MINLP). Traditional solving schemes mainly involve exact solving algorithms and heuristic solving algorithms. Exact solving algorithms can provide optimality guarantees, but as the size of the platoon increases, the computational complexity of the problem will grow exponentially, and there is a "dimension disaster" problem, making it difficult to apply to actual large-scale scheduling. To improve the solving efficiency, problem decomposition and state space discretization are usually performed, and then dynamic programming and other exact algorithms are used to solve the sub-problems, but at the same time of problem decomposition, the global optimality is sacrificed. Heuristic solving algorithms can also improve the solving efficiency to some extent, but the quality of the solution cannot be guaranteed. The trade-off between computational efficiency and solution quality has become a key bottleneck in real-time and large-scale coordination.

[0004] Therefore, the current truck platoon scheduling algorithm is only suitable for small-scale scenarios, such as truck platoons with less than 100 trucks, and when facing large-scale truck platoons with more than 100 trucks, the scheduling often has high computational complexity and unstable solution quality problems. SUMMARY

[0005] In view of the above analysis, the embodiment of the present application aims to provide a large-scale truck platoon scheduling method and device based on deep graph learning, to solve the problem that the existing truck platoon scheduling algorithm often has high computational complexity and unstable solution quality when facing large-scale scheduling.

[0006] The purpose of the present application is achieved as follows:

[0007] A large-scale truck platoon scheduling method based on deep graph learning, comprising the steps of:

[0008] Defining a set of passing hubs and a set of passing road sections according to the vehicles, driving trajectories and hubs of the trucks in the road network;

[0009] Finding the set of potential platoon objects and the set of waiting times of each truck at each passing hub based on the set of passing hubs and the set of passing road sections of each truck;

[0010] Defining the interaction information graph of the potential platoon objects at the current hub for each truck according to the set of potential platoon objects and the set of waiting times, and training a graph neural network model based on the interaction information graph, and optimizing the internal parameters of the graph neural network model by learning the optimal solution dataset;

[0011] Inputting the passing hubs and departure time of each truck into the graph neural network model to generate the optimal waiting time and predicted platoon object at each passing hub, and to generate a platoon scheduling strategy.

[0012] The large-scale truck platoon scheduling method based on deep graph learning of the present application represents the interaction information between a single truck and potential platoon objects at a hub as a node-edge model, captures the coupling relationship between time and space in the platoon strategy, and improves the quality of the platoon scheduling strategy compared to traditional methods, providing platoon opportunities for more trucks and effectively reducing the total system cost. At the same time, based on the end-to-end graph neural network model for large-scale truck platoon problems, the generation efficiency of the platoon scheduling strategy is effectively improved.

[0013] As one of the optional embodiments, the process of defining the set of passing hubs and the set of passing road sections according to the vehicles, driving trajectories and hubs of the trucks in the road network is as follows:

[0014] The set of passing hubs of the truck is ; wherein, is the th hub passed by the truck , the set of trucks is , , and represents the hub where the truck ends;

[0015] The set of passing road sections is ;in, Indicates truck The first one along the way The and the first The road section between the hubs.

[0016] As one optional embodiment, the process of finding the potential platooning set and waiting time set for each truck at each platooning hub, based on the set of hubs and road segments traversed by each truck, includes the following steps:

[0017] Based on the set of routes taken by each truck, two trucks sharing a common route are potential platooning partners; among them, trucks with common routes are potential platooning partners. The first one it passed through A set of potential formation objects with common paths at a hub As shown in the following formula:

[0018] ;

[0019] in, Indicates truck Arrive at the first of its routes After the hub, the next section of the road to be traversed; For trucks The trucks gather at the section of the road they are traveling through. The driving trajectory also includes road segments At that time, it represented the truck. with truck With public routes, trucks That is, a truck The first one it passed through Potential formation targets for each hub;

[0020] ;

[0021] in, Indicates truck Arrive at the first of its routes A pivotal moment, For trucks From the Departure times of each hub; Indicates truck The first one along the way The and the first The distance between the hubs For trucks driving speed; truck From the Departure times of each hub = equal to the time it arrives at the hub the sum of the waiting times of the truck at the hubs it passes through

[0022] A set of waiting times is determined according to the waiting times of any truck and each potential platoon object; wherein the truck passes through the first hub of the set of waiting times is as follows:

[0023]

[0024] wherein, denotes the time at which the truck arrives at the first hub it passes through, denotes the time at which the truck arrives at the first hub, the time difference between the truck arriving at the first hub and the truck arriving at the first hub is i.e. the time at which the truck waits at the current hub for the truck to arrive at the first hub it passes through.

[0025] As one of the optional embodiments, the process of defining the interaction information graph of the potential platoon objects of each truck at the current hub according to the set of potential platoon objects and the set of waiting times comprises the steps of:

[0026] determining the set of nodes and the set of edges in the interaction information graph according to the set of potential platoon objects; wherein the set of nodes comprises the truck and all the potential platoon objects of the truck passing through the first hub, as follows:

[0027]

[0028] wherein, is the set of potential platoon objects passing through the first hub; in the set of nodes , the truck is the central node and the rest of the potential platoon objects are edge nodes; each node is associated with a feature vector x i,k , as follows:

[0029] ;​​​​​

[0030] where x i,k represents the feature vector of the node in the interaction information graph ; is the initial feature of the node; represents the number of the truck, represented by a one-hot vector with a length of ; represents the position information of the th hub that the truck passes through, and the order of the hubs that the truck passes through is represented by a one-hot vector with a length of ;

[0031] determining the edge set of the interaction information graph according to the waiting time set ; wherein the waiting time between the center node truck and the edge node truck is taken as the input feature of the edge, and the feature vector of the edge is represented as , represented by a one-hot vector with a length of , is the upper limit of the single waiting time of the truck;

[0032] fuse and transform the original feature vectors of the nodes and edges through the ReLU activation function, as follows:

[0033] ;

[0034] where x i,k represents the feature vector of the node in the interaction information graph ; represents the feature vector of the edge; represents the activation function; [;] represents the vector splicing operation; represents the waiting time between the center node truck and the edge node truck ; , , and , , , , are all trainable parameters.

[0035] As one of the optional embodiments, the graph neural network model is trained based on the interaction information graph, and the process of optimizing the internal parameters of the graph neural network model by learning the optimal solution data set includes the following steps:

[0036] obtaining the initial embedding of the graph convolution layer through the trainable parameters and​​ , the initial embedding of each node in the graph is the feature vector of the input layer, which is the original feature representation of the model before any graph convolution operation, to project the features of the nodes and edges of the interaction information graph into the graph neural network model as follows:

[0037] ;

[0038] wherein, is the node initial embedding of the graph convolution layer, denotes the edge initial embedding of the graph convolution layer; x i,k denotes the interaction information graph the feature vector of the node ; denotes the feature vector of the edge; , , and denote trainable parameters;

[0039] updating the node aggregation information and the node combination information of the graph neural network model; wherein the node aggregation information is derived from the neighborhood of the center node, which integrates the embeddings of all neighbor nodes; the node combination information integrates the feature of the center node itself and the aggregation feature of its neighbor nodes, as follows:

[0040] ;

[0041] ;

[0042] wherein, denotes the node neighbor node of the node in the layer graph convolution layer, neighbor node set of the center node represents activation function; denotes the first aggregation function; denotes the combination of the aggregation feature of the neighbor node of the center node in the layer graph convolution layer and its own feature, i.e. the node combination information, and similarly, denotes the layer node combination information, denotes the layer node combination information; and are both trainable parameters;

[0043] Update edge aggregation information of the graph neural network model Information on edge combinations As shown in the following formula:

[0044] ;

[0045] ;

[0046] ;

[0047] in, Indicates the first Edges in layered graph convolutional layers Edge aggregation information of neighboring nodes, Indicates the first Edges in layered graph convolutional layers The combination of the aggregation features of neighboring nodes and its own features, i.e., edge combination information; Indicates the first Edges in layered graph convolutional layers The combination of the aggregation features of the neighboring nodes and the features of the edge itself; the neighbors of an edge are the nodes at both ends of the edge, i.e. ; represent Activation function; Indicates the second aggregation function; , , , as well as All of these are trainable parameters.

[0048] As one optional embodiment, the process of training a graph neural network model based on the interactive information graph and optimizing the internal parameters of the graph neural network model by learning the optimal solution dataset further includes the following steps:

[0049] When it exists When using a layered graph convolutional layer, the first layer and the second layer... The feature embedding of a fully connected layer is as follows:

[0050] ;

[0051] in, For the feature embedding of edges in the first fully connected layer, For the first Feature embedding of edges in fully connected layers No. Feature embedding of edges in fully connected layers For the first Feature embedding of edges in fully connected layers; , , and are trainable parameters; since the prediction task of the method of the present application is only related to the output of the final edge, only the feature update of the edge needs to be considered here;

[0052] When there are layer graph convolution layers, the final output of the encoder part is which is converted into a prediction result related to the task by the decoder as follows:

[0053] ;

[0054] wherein, and are trainable parameters; is a function, which realizes the normalization processing of the output result; is a function, which realizes the conversion of the normalized probability output into a hard-assignment one-hot encoding vector;

[0055] The gap between the predicted value and the true value is calculated using the cross-entropy loss function as follows:

[0056] ;

[0057] When the predicted value tends to the true value , the value will gradually decrease, and each trainable parameter in the graph neural network model is gradually optimized based on this benchmark, so that the graph neural network model gradually improves the ability to find the formation scheduling strategy, and the converged model is used as the final applied graph neural network model.

[0058] As one of the optional embodiments, the large-scale truck formation scheduling method based on deep graph learning further comprises the steps of:

[0059] calculating the total cost of a single truck in the formation scheduling strategy; wherein the total cost includes fuel cost, waiting time cost and transportation task delay cost;

[0060] comparing the difference between the total cost and the initial cost to determine whether the formation scheduling strategy is executed; wherein the initial cost represents the fuel cost consumed by a single truck without waiting at the hub and without accepting formation scheduling.

[0061] As one of the optional embodiments, the process of calculating the total cost of a single truck in the formation scheduling strategy is as follows:

[0062] the truck is on a road segment fuel costs As shown in the following formula:

[0063] ;

[0064] in, The fuel cost per second for the following vehicle when trucks are traveling in platoons. This represents the fuel cost per second for a truck traveling alone and as a lead vehicle in a platoon. When trucks travel in platoons, the cost is reduced by 10% compared to traveling alone. For example, when a truck travels at a constant speed of 80 km / h, and considering that a fully loaded heavy truck consumes approximately 40 liters of fuel per 100 kilometers, and given that diesel, currently the primary fuel for road freight, is priced at around 6.5 yuan per liter, we can assume the fuel cost per second for a truck traveling alone or as a lead vehicle in a platoon. The fuel cost per second for all following vehicles when traveling in a convoy is 0.057 yuan. It is 0.051 yuan; For trucks On the road section Travel time on the road; For trucks The first one it passed through The predicted fleet objects of each hub, namely, trucks It has a public route and is close to trucks. At the same time, the collection of vehicles that depart from the starting point of this public route; For set The number of trucks in the middle, That is, a truck On the road section The size of the convoy; all trucks in the convoy share the fuel cost for that section of the road;

[0065] truck The first one it passed through Waiting time cost at each hub As shown in the following formula:

[0066] ;

[0067] in, For trucks The first one it passed through The waiting time cost at each hub; For trucks From the starting hub to the first point it passes through When using a hub, the cumulative waiting time; For trucks The first one it passed through Waiting time at each hub; For trucks The maximum waiting time for trucks Without formation scheduling, 10% of the total travel time; when At that time, the waiting time cost for trucks is Based on the current wage level of truck drivers in my country, we can assume it to be 0.016 yuan per second. At that time, the waiting time cost of trucks per second Relatively low, which can be assumed to be 0.011 yuan per second;

[0068] truck The cost of delays in transportation tasks is as follows:

[0069] ;

[0070] in, For trucks The difference between the actual arrival time at the destination and the arrival time stipulated in the transportation task. The cost per second of delay can be flexibly adjusted according to the urgency of freight transportation. For example, a higher cost per second of delay can be set for fresh goods, while a relatively lower cost per second of delay can be set for goods such as coal.

[0071] A large-scale truck platooning scheduling device based on deep graph learning includes:

[0072] The data definition module is used to define the set of transit hubs and the set of transit road segments based on the vehicles, driving trajectories, and hubs of trucks in the road network;

[0073] The collection processing module is used to find the potential platooning object set and waiting time set for each truck at each platooning hub based on the set of hubs and road segments it passes through.

[0074] The model training module is used to define the interaction information graph of each truck at the current hub based on the set of potential platooning objects and the set of waiting times, and to train the graph neural network model based on the interaction information graph. The internal parameters of the graph neural network model are optimized by learning the optimal solution dataset.

[0075] The strategy output module is used to take the route hubs and departure times of each truck as input into the graph neural network model, generate its optimal waiting time at each route hub and predict the platooning objects, so as to generate the platooning scheduling strategy.

[0076] At least one embodiment of this application also provides a data control device, including:

[0077] One or more memories that store computer-executable instructions non-transitory;

[0078] One or more processors are configured to run computer-executable instructions, wherein the computer-executable instructions are executed by the one or more processors to implement a large-scale truck platooning scheduling method based on deep graph learning according to any embodiment of the present application.

[0079] At least one embodiment of this application also provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement a large-scale truck platooning scheduling method based on deep graph learning according to any embodiment of this application.

[0080] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0081] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings.

[0082] Figure 1 A flowchart illustrating a large-scale truck platooning scheduling method based on deep graph learning, according to an embodiment of the application.

[0083] Figure 2 This is a schematic diagram illustrating a preferred embodiment of a truck's ability to select platooning targets and waiting times at a hub;

[0084] Figure 3 Information graph construction for another preferred embodiment of the large-scale truck platooning scheduling method based on deep graph learning;

[0085] Figure 4 This is a model training framework for a large-scale truck platooning scheduling method based on deep graph learning, as described in one embodiment of the application.

[0086] Figure 5 This is a structural diagram of a large-scale truck platooning scheduling device based on deep graph learning, according to an embodiment of the application.

[0087] Figure 6 A schematic block diagram of a data control device provided by the present invention;

[0088] Figure 7This is a schematic diagram of a non-transitory computer-readable storage medium provided by the present invention. Detailed Implementation

[0089] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0090] To facilitate understanding of the embodiments of this application, further explanation and description will be provided below with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of this application. In the drawings, the dimensions and relative dimensions of components may be exaggerated for clarity and / or descriptive purposes. When exemplary embodiments can be implemented differently, a specific process sequence may be performed in a different order than that described. For example, two consecutively described processes may be performed substantially simultaneously or in the reverse order of their description.

[0091] The terminology used herein is for the purpose of describing particular embodiments and is not intended to be limiting. As used herein, unless the context clearly indicates otherwise, the singular forms “a” and “(the)” are also intended to include the plural forms. Furthermore, when the terms “comprising” and / or “including” and variations thereof are used in this specification, it indicates the presence of the stated features, integrals, steps, operations, parts, components, and / or groups thereof, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, parts, components, and / or groups thereof. It should also be noted that, as used herein, the terms “substantially,” “about,” and other similar terms are used as approximate terms rather than as terms of degree, thus explaining the inherent biases in measurements, calculated values, and / or provided values ​​that would be recognized by one of ordinary skill in the art.

[0092] This application provides a method for large-scale truck platooning scheduling based on deep graph learning.

[0093] Figure 1 Here is a flowchart of a large-scale truck platooning scheduling method based on deep graph learning, as an embodiment of the application. Figure 1 As shown, a method for large-scale truck platooning scheduling based on deep graph learning according to one embodiment of the application includes steps S100 to S103:

[0094] S100 defines the set of hubs and the set of road segments traversed based on the vehicles, driving trajectories, and hubs of freight trucks in the road network.

[0095] S101, based on the set of hubs and road segments that each truck passes through, find the set of potential platooning objects and the set of waiting times for each truck at each hub.

[0096] S102, define the interaction information graph of each truck at the current hub based on the potential formation object set and the waiting time set, and train the graph neural network model based on the interaction information graph, and optimize the internal parameters of the graph neural network model by learning the optimal solution dataset;

[0097] S103. Input the transit hubs and departure times of each truck into the graph neural network model to generate its optimal waiting time at each transit hub and predict the platooning objects, so as to generate a platooning scheduling strategy.

[0098] In this embodiment, the truck's vehicle, driving trajectory, and hub are all represented by codes, with each code corresponding one-to-one with the entity, establishing the set and determining the symbol of the marker. There are N trucks in the road network, each with the same driving speed, but their origin, destination, and departure times are not exactly the same. By adjusting the waiting time of trucks at hubs along the route, trucks with common routes can form platoons and travel together. Based on this, step S100 defines the hub set and the road segment set according to the truck's vehicle, driving trajectory, and hub in the road network, as follows:

[0099] truck The hubs along the route are ;in, For trucks The first one along the way A hub, where trucks are grouped together , Indicates truck The hub where the destination is located;

[0100] The route segments are grouped as follows ;in, Indicates truck The first one along the way The and the first The road section between the hubs.

[0101] In the embodiments of this application, such as Figure 2 As shown, step S101, which involves finding the potential platooning set and waiting time set for each truck at each platooning hub based on the set of hubs and road segments it passes through, includes the following steps:

[0102] Based on the set of routes taken by each truck, two trucks sharing a common route are potential platooning partners; among them, trucks with common routes are potential platooning partners. The first one it passed through A set of potential formation objects with common paths at a hub As shown in the following formula:

[0103] ;

[0104] in, Indicates truck Arrive at the first of its routes After the hub, the next section of the road to be traversed; For trucks The trucks gather at the section of the road they are traveling through. The driving trajectory also includes road segments At that time, it represented the truck. with truck With public routes, trucks That is, a truck The first one it passed through Potential formation targets for each hub;

[0105] ;

[0106] in, Indicates truck Arrive at the first of its routes A pivotal moment, For trucks From the Departure times of each hub; Indicates truck The first one along the way The and the first The distance between the hubs For trucks driving speed; truck From the Departure times of each hub = equal to the time it arrives at the hub Rather than the waiting time at this node sum;

[0107] Based on the waiting time of any truck and each potential platoon member, a set of waiting times is determined; where, trucks The first one it passed through Hub wait time set As shown in the following formula:

[0108] ;

[0109] in, Indicates truck Arrive at the first of its routes A pivotal moment, Indicates truck Reaching the At the hub, trucks with truck Reaching the Time difference of hub trucks Waiting for trucks at the current hub The time when the truck Choose the first one along the way When no stop is made at a hub, the waiting time is 0. In the embodiments of this application, such as Figure 3 As shown, this displays the platooning arrangement of truck number 154 at hub 81. Figure 3 The numbers 65, 176, 116, and 91 are all truck numbers in the potential platoon; based on this, step S102, which defines the interaction information graph of each truck in the current hub according to the potential platoon object set and the waiting time set, includes the following steps:

[0110] Determine the interaction information graph based on the set of potential formation objects. The set of nodes in With edge set The node set includes trucks. And all trucks The first one it passed through The potential formation objects for each hub are as follows:

[0111] ;

[0112] in, for The first one it passed through A set of potential formation objects for each hub; in the set of nodes In the middle, with trucks The central node is the node in the formation, and the remaining potential formation objects are edge nodes; each node is associated with a feature vector x. i,k Related, as follows:

[0113] ;

[0114] Where, x i,k Represents interactive information graphs Middle node The feature vectors are the initial features of the nodes. The truck's number is represented by a string of length 1000. The one-hot vector representation; Indicates truck The first one to pass The location information of the hubs, and the order in which the routes pass through the hubs, are obtained through a length of [length missing]. The one-hot vector representation;

[0115] Determine the edge set of the interactive information graph based on the set of waiting times. Among them, the central node trucks trucks with edge nodes waiting time between As input features of an edge, the feature vector of an edge is represented as: Through a length of The one-hot vector representation (including the 0-second option, i.e., no waiting). The maximum waiting time for a single truck (all candidate waiting times are divided in increments of 1 second);

[0116] The feature vectors of nodes and edges are fused and transformed using the ReLU activation function, as shown in the following equation:

[0117] ;

[0118] Where, x i,k Represents interactive information graphs Middle node eigenvectors; The eigenvectors representing the edges; represent Activation function; [;] denotes vector concatenation operation; Indicates the central node truck trucks with edge nodes The waiting time between; , , as well as , , All of these are trainable parameters.

[0119] In the embodiments of this application, such as Figure 4 As shown, step S102, which involves training a graph neural network model based on an interactive information graph and optimizing the internal parameters of the graph neural network model by learning the optimal solution dataset, includes the following steps:

[0120] The initial embedding of the graph convolutional layer is obtained through trainable parameters. and The initial embedding is the feature vector of each node in the graph at the input layer. It is the original feature representation of the model before any graph convolution operation, which projects the features of the nodes and edges of the interactive information graph onto the graph neural network model, as shown in the following equation:

[0121] ;

[0122] in, It is a node In the initial embedding of the graph convolutional layer Representing an edge Initial embedding in the graph convolutional layer; x i,k Represents interactive information graphs Middle node eigenvectors; The eigenvectors representing the edges; , , as well as Indicates trainable parameters;

[0123] The node aggregation information and node combination information of the graph neural network model are updated; where the node aggregation information comes from the neighborhood of the central node and gathers the embeddings of all neighboring nodes; the node combination information integrates the self-features of the central node and the aggregation features of its neighboring nodes, as shown in the following formula:

[0124] ;

[0125] ;

[0126] in, Indicates the first Nodes in a layered graph convolutional layer The node aggregation information of neighboring nodes, As the central node The set of neighboring nodes; represent Activation function; Represents the first aggregate function; Indicates the first Central node in a layered graph convolutional layer The aggregation characteristics of a node's neighboring nodes are combined with its own characteristics; that is, node combination information. Similarly, Indicates the first Layer nodes Combined information, Indicates the first Layer nodes Combined information; and All of these are trainable parameters;

[0127] Update edge aggregation information of the graph neural network model Information on edge combinations As shown in the following formula:

[0128] ;

[0129] ;

[0130] ;

[0131] in, Indicates the first Edges in layered graph convolutional layers Edge aggregation information of neighboring nodes, Indicates the first Edges in layered graph convolutional layers The combination of the aggregation features of neighboring nodes and its own features, i.e., edge combination information; Indicates the first Edges in layered graph convolutional layers The combination of the aggregation features of the neighboring nodes and the features of the edge itself; the neighbors of an edge are the nodes at both ends of the edge, i.e. ; represent Activation function; Indicates the second aggregation function; , , , as well as All of these are trainable parameters;

[0132] When it exists When using a layered graph convolutional layer, the first layer and the second layer... The feature embedding of a fully connected layer is as follows:

[0133] ;

[0134] in, For the feature embedding of edges in the first fully connected layer, For the first Feature embedding of edges in fully connected layers No. Feature embedding of edges in fully connected layers For the first Feature embedding of edges in fully connected layers; , , as well as All of these are trainable parameters; since the prediction task of the method in this application is only related to the final edge output, we only need to consider the edge feature update here.

[0135] When it exists In a layer graph convolutional layer, the final output of the encoder part is: The decoder then transforms it into task-related prediction results. As shown in the following formula:

[0136] ;

[0137] in, and All of these are trainable parameters; for The function performs normalization on the output results; for The function converts the normalized probability output into a hard-assigned one-hot encoded vector.

[0138] The predicted value is calculated using the cross-entropy loss function. Compared with the true value The difference between them is as follows:

[0139] ;

[0140] When the predicted value Approaching the true value hour, The value will gradually decrease, and based on this, the trainable parameters in the graph neural network model will be gradually optimized to improve the graph neural network model's ability to find formation scheduling strategies. The converged model will then be used as the final graph neural network model for application.

[0141] Input the truck's transit hubs and departure time into the trained graph neural network to generate its optimal waiting time and predicted platooning objects at each transit hub.

[0142] As one of the preferred embodiments, such as Figure 1 As shown, it also includes steps S200 and S201:

[0143] S200 calculates the total cost of a single truck in the platooning scheduling strategy; the total cost includes fuel cost, waiting time cost, and transportation task delay cost.

[0144] S201. Compare the difference between the total cost and the initial cost to determine whether the platooning scheduling strategy has been implemented. The initial cost represents the fuel cost consumed by a single truck that does not wait at the hub and does not accept platooning scheduling.

[0145] Cost comparison is performed based on the generated formation scheduling strategy to determine whether the formation scheduling strategy should be executed.

[0146] The process of calculating the total cost of a single truck in the platooning scheduling strategy is as follows:

[0147] truck On the road section fuel costs As shown in the following formula:

[0148] ;

[0149] in, The fuel cost per second for the following vehicle when trucks are traveling in platoons. This represents the fuel cost per second for a truck traveling alone and as a lead vehicle in a platoon. When trucks travel in platoons, the cost is reduced by 10% compared to traveling alone. For example, when a truck travels at a constant speed of 80 km / h, and considering that a fully loaded heavy truck consumes approximately 40 liters of fuel per 100 kilometers, and given that diesel, currently the primary fuel for road freight, is priced at around 6.5 yuan per liter, we can assume the fuel cost per second for a truck traveling alone or as a lead vehicle in a platoon. The fuel cost per second for all following vehicles when traveling in a convoy is 0.057 yuan. It is 0.051 yuan; For trucks On the road section Travel time on the road; For trucks The first one it passed through The predicted fleet objects of each hub, namely, trucks It has a public route and is close to trucks. At the same time, the collection of vehicles that depart from the starting point of this public route; For set The number of trucks in the middle, That is, a truck On the road section The size of the convoy; all trucks in the convoy share the fuel cost for that section of the road;

[0150] truck The first one it passed through Waiting time cost at each hub As shown in the following formula:

[0151] ;

[0152] in, For trucks The first one it passed through The waiting time cost at each hub; For trucks From the starting hub to the first point it passes through When using a hub, the cumulative waiting time; For trucks The first one it passed through Waiting time at each hub; For trucks The maximum waiting time for trucks Without formation scheduling, 10% of the total travel time; when At that time, the waiting time cost for trucks is Based on the current wage level of truck drivers in my country, we can assume it to be 0.016 yuan per second. At that time, the waiting time cost of trucks per second Relatively low, which can be assumed to be 0.011 yuan per second;

[0153] truck The cost of delays in transportation tasks is as follows:

[0154] ;

[0155] in, For trucks The difference between the actual arrival time at the destination and the arrival time stipulated in the transportation task. The cost per second of delay can be flexibly adjusted according to the urgency of freight transportation. For example, a higher cost per second of delay can be set for fresh goods, while a relatively lower cost per second of delay can be set for goods such as coal.

[0156] As one preferred embodiment, different weighting coefficients are configured for each cost. .

[0157] That is, the total cost of a single truck from its starting point to the completion of its transportation task. As shown in the following formula:

[0158] ;

[0159] in, For trucks The total transportation cost includes total fuel cost, total waiting time cost, and transportation delay cost. Total fuel cost is the sum of fuel costs incurred by the truck across all routes it travels. , For trucks On the road section On fuel costs, For trucks The set of route segments; the total waiting time cost is the sum of the waiting time costs of trucks at all the junctions along the route. , For trucks The first one it passed through The waiting time cost at each hub For trucks The hubs along the route; For trucks The cost of delays in transportation tasks. Weighting factor. Adjustments can be made based on different truck routes and transportation tasks. Therefore, the cost of the platooning strategy is compared with the initial cost of the trucks (i.e., the fuel cost incurred if the trucks do not wait at hubs or accept platooning), and the lowest-cost option is chosen as the final transportation plan for that truck.

[0160] Meanwhile, the formation scheduling strategy of this application embodiment requires less time to solve compared with traditional algorithms, as shown in Table 1 below:

[0161] Table 1. Comparison of Solution Time

[0162]

[0163] The scheduling scheme of this application embodiment can provide more platooning opportunities compared with traditional algorithms. Table 2 below shows the comparison data of the scheduling results of truck number 89.

[0164] Table 2 Comparison of dispatch results for truck number 89

[0165]

[0166] The large-scale truck platooning scheduling method based on deep graph learning in this application represents the interaction information of a single truck with potential platooning objects at a hub as a node-edge model, capturing the temporal and spatial coupling relationship in the platooning strategy. Compared with traditional methods, this improves the quality of the platooning scheduling strategy, provides more platooning opportunities for trucks, and effectively reduces the total system cost. Simultaneously, based on an end-to-end graph neural network model for the large-scale truck platooning problem, it effectively improves the efficiency of platooning scheduling strategy generation.

[0167] This application also provides a large-scale truck platooning scheduling device based on deep graph learning.

[0168] Figure 5 This is a structural diagram of a large-scale truck platooning scheduling device based on deep graph learning, according to one embodiment of the application. Figure 5 As shown, one embodiment of a large-scale truck platooning scheduling device based on deep graph learning includes:

[0169] The data definition module 100 is used to define the set of transit hubs and the set of transit road segments based on the vehicles, driving trajectories and hubs of trucks in the road network;

[0170] The set processing module 101 is used to find the potential platooning object set and waiting time set of each truck at each platooning hub based on the set of hubs and road segments passed by each truck.

[0171] The model training module 102 is used to define the interaction information graph of each truck at the current hub based on the potential platooning object set and the waiting time set, and to train the graph neural network model based on the interaction information graph, and to optimize the internal parameters of the graph neural network model by learning the optimal solution dataset.

[0172] The strategy output module 103 is used to input the route hubs and departure times of each truck into the graph neural network model, generate its optimal waiting time and predicted platooning objects at each route hub, and generate a platooning scheduling strategy.

[0173] The large-scale truck platooning scheduling device based on deep graph learning in this application represents the interaction information of a single truck with potential platooning objects at a hub as a node-edge model, capturing the temporal and spatial coupling relationship in the platooning strategy. Compared with traditional methods, this improves the quality of the platooning scheduling strategy, provides more platooning opportunities for trucks, and effectively reduces the total system cost. Simultaneously, based on an end-to-end graph neural network model for the large-scale truck platooning problem, it effectively improves the efficiency of platooning scheduling strategy generation.

[0174] At least one embodiment of this application also provides a data control device. Figure 6 This is a schematic block diagram of a data control device provided for at least one embodiment of this application. For example, such as... Figure 6 As shown, the data control device 20 may include one or more memories 200 and one or more processors 201. The memories 200 are used to store computer-executable instructions non-transitory; the processors 201 are used to run the computer-executable instructions, which, when run by the processors 201, can cause the processors 201 to perform one or more steps in the large-scale truck platooning scheduling method based on deep graph learning according to any embodiment of this application.

[0175] For the specific implementation and explanation of each step of the large-scale truck platooning scheduling method based on deep graph learning, please refer to the relevant content in the above-mentioned embodiment of the large-scale truck platooning scheduling method based on deep graph learning, which will not be repeated here. It should be noted that Figure 6 The components of the data control device 20 shown are merely exemplary and not limiting. The data control device 20 may have other components depending on the actual application requirements.

[0176] In one embodiment, the processor 201 and the memory 200 can communicate directly or indirectly with each other. For example, the processor 201 and the memory 200 can communicate via a network connection. The network can include wireless networks, wired networks, and / or any combination of wireless and wired networks; this application does not limit the type and function of the network. Alternatively, the processor 201 and the memory 200 can also communicate via a bus connection. The bus can be a Peripheral Component Interconnect Standard (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. For example, the processor 201 and the memory 200 can be located at a remote data server (cloud) or a distributed energy system (local), or at a client (e.g., a mobile device such as a mobile phone). For example, the processor 201 can be a central processing unit (CPU), a tensor processor (TPU), or a graphics processing unit (GPU), etc., with data processing and / or instruction execution capabilities, and can control other components in the data control device 20 to perform desired functions. The central processing unit (CPU) can be an x86 or ARM architecture, etc.

[0177] In one embodiment, memory 200 may include any combination of one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, erasable programmable read-only memory (EPROM), portable compact disc read-only memory (CD-ROM), USB memory, flash memory, etc. One or more computer-executable instructions may be stored on the computer-readable storage medium, and processor 201 may execute these computer-executable instructions to implement various functions of data control device 20. Various application programs and various data, as well as various data used and / or generated by the application programs, may also be stored in memory 200.

[0178] It should be noted that the data control device 20 can achieve similar technical effects to the aforementioned large-scale truck platooning scheduling method based on deep graph learning, and the repetitions will not be repeated.

[0179] At least one embodiment of this application also provides a non-transitory computer-readable storage medium. Figure 7 This is a schematic diagram of a non-transitory computer-readable storage medium provided for at least one embodiment of this application. For example, such as... Figure 7As shown, one or more computer-executable instructions 301 may be stored non-transitory on the non-transitory computer-readable storage medium 30. For example, when the computer-executable instructions 301 are executed by a computer, the computer may perform one or more steps in a large-scale truck platooning scheduling method based on deep graph learning according to any embodiment of this application.

[0180] In one embodiment, the non-transitory computer-readable storage medium 30 can be applied to the data control device 20 described above, for example, it can be the memory 200 in the data control device 20.

[0181] In one embodiment, the description of the non-transitory computer-readable storage medium 30 can be found in the description of the memory 200 in the embodiment of the data control device 20, and will not be repeated hereafter.

[0182] It should be noted that the memory 200 stores different non-transient computer-executable instructions, and the data control device 20 corresponds to the firmware upgrade device. When the computer-executable instructions are run by the processor 201, the processor 201 can perform one or more steps in the large-scale truck platooning scheduling method based on deep graph learning according to any embodiment of this application.

[0183] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A large-scale truck platooning scheduling method based on deep graph learning, characterized in that, Including the following steps: Based on the vehicles, travel trajectories, and hubs of freight trucks in the road network, define the set of hubs traversed and the set of road segments traversed, as follows: truck The hubs along the route are ;in, For trucks The first one along the way A hub, where trucks are grouped together , Indicates truck The hub where the destination is located; The route segments are grouped as follows ;in, Indicates truck The first one along the way The and the first The road section between the hubs; Based on the set of hubs and road segments traversed by each truck, the process involves finding the potential platooning set and waiting time set for each truck at each hub, including the following steps: with truck The first one it passed through A set of potential formation objects with common paths at a hub As shown in the following formula: ； Indicates truck After the first The route to be taken after the hub; For trucks The trucks gather at the section of the road they are traveling through. The driving trajectory also includes road segments At that time, it represented the truck. with truck With public routes, trucks For trucks The first one it passed through Potential formation targets for each hub; Based on the set of potential platooning objects and the set of waiting times, an interaction information graph of each truck at the current hub is defined, and a graph neural network model is trained based on the interaction information graph. The internal parameters of the graph neural network model are optimized by learning the optimal solution dataset. The graph neural network model is input with the transit hubs and departure times of each truck to generate its optimal waiting time and predicted platooning objects at each transit hub, thereby generating a platooning scheduling strategy.

2. The large-scale truck platooning scheduling method based on deep graph learning according to claim 1, characterized in that, The process of defining the interaction information graph of each truck at the current hub based on the set of potential platooning objects and the set of waiting times includes the following steps: The interaction information graph is determined based on the set of potential formation objects. The set of nodes in With edge set The node set includes trucks. And all trucks The first one it passed through The potential formation objects for each hub are as follows: ； In the node set In the middle, with trucks The central node is the node in the formation, and the other potential formation objects are edge nodes; each node is associated with a feature vector x. i,k Related, as follows: ； x i,k Represents interactive information graphs Middle node The feature vectors are the initial features of the nodes. Indicates the truck's serial number; Indicates truck The first one to pass Hub location information; The set of edges of the interactive information graph is determined based on the set of waiting times. ; the central node truck trucks with edge nodes waiting time between As input features of an edge, the feature vector of an edge is represented as: ; The original feature vectors of nodes and edges are fused and transformed using the ReLU activation function, as shown in the following equation: ； represent Activation function; [;] denotes vector concatenation operation; Indicates the central node truck trucks with edge nodes The waiting time between; , , as well as , , All of these are trainable parameters.

3. The large-scale truck platooning scheduling method based on deep graph learning according to claim 2, characterized in that, The process of training a graph neural network model based on the interactive information graph and optimizing the internal parameters of the graph neural network model by learning the optimal solution dataset includes the following steps: The initial embedding of the graph convolutional layer is obtained through trainable parameters. and The features of the nodes and edges of the interactive information graph are projected onto the graph neural network model as follows: ； It is a node In the initial embedding of the graph convolutional layer Representing an edge Initial embedding in the graph convolutional layer; , , as well as Indicates trainable parameters; The node aggregation information and node combination information of the graph neural network model are updated as follows: ； ； Indicates the first Nodes in a layered graph convolutional layer The node aggregation information of neighboring nodes, As the central node The set of neighboring nodes; Represents the first aggregate function; Indicates the first Central node in a layered graph convolutional layer The combination of the aggregation features of neighboring nodes and their own features, i.e., the node combination information; Indicates the first Layer nodes Combined information, Indicates the first Layer nodes Combined information; and All of these are trainable parameters; Update the edge aggregation information of the graph neural network model. Information combined with edges As shown in the following formula: ； ； ； Indicates the first Edges in layered graph convolutional layers Edge aggregation information of neighboring nodes, Indicates the first Edges in layered graph convolutional layers The combination of the aggregation features of neighboring nodes and its own features. Indicates the first Edges in layered graph convolutional layers The combination of the aggregation features of the neighboring nodes and the features of the edge itself; the neighbors of an edge are the nodes at both ends of the edge, i.e. ; Indicates the second aggregation function; , , , as well as All of these are trainable parameters.

4. The large-scale truck platooning scheduling method based on deep graph learning according to claim 3, characterized in that, The process of training a graph neural network model based on the interactive information graph and optimizing the internal parameters of the graph neural network model by learning the optimal solution dataset further includes the following steps: When it exists When using a layered graph convolutional layer, the first layer and the second layer... The feature embedding of a fully connected layer is as follows: ； For the feature embedding of edges in the first fully connected layer, For the first Feature embedding of edges in fully connected layers No. Feature embedding of edges in fully connected layers For the first Feature embedding of edges in fully connected layers; , , as well as All of these are trainable parameters; When it exists In a layer graph convolutional layer, the final output of the encoder part is: The decoder then transforms it into task-related prediction results. As shown in the following formula: ； and All of these are trainable parameters; for function; for function; The predicted value is calculated using the cross-entropy loss function. Compared with the true value The difference between them is as follows: ； When the predicted value Approaching the true value In this process, each trainable parameter is gradually optimized, and the converged model is used as the final graph neural network model for application.

5. The large-scale truck platooning scheduling method based on deep graph learning according to any one of claims 1 to 4, characterized in that, It also includes the following steps: Calculate the total cost of a single truck in the platooning scheduling strategy; wherein the total cost includes fuel cost, waiting time cost, and transportation task delay cost; By comparing the difference between the total cost and the initial cost, it is determined whether the platooning scheduling strategy is executed; wherein, the initial cost represents the fuel cost consumed by a single truck that does not wait at the hub and does not accept platooning scheduling.

6. A large-scale truck platooning scheduling device based on deep graph learning, characterized in that, include: The data definition module is used to define the set of transit hubs and the set of transit road segments based on the vehicles, travel trajectories, and hubs of trucks in the road network, as follows: truck The hubs along the route are ;in, For trucks The first one along the way A hub, where trucks are grouped together , Indicates truck The hub where the destination is located; The route segments are grouped as follows ;in, Indicates truck The first one along the way The and the first The road section between the hubs; The set processing module is used to find the potential platooning object set and waiting time set for each truck at each platooning hub based on the set of hubs and road segments it passes through. The steps include: with truck The first one it passed through A set of potential formation objects with common paths at a hub As shown in the following formula: ； Indicates truck After the first The route to be taken after the hub; For trucks The trucks gather at the section of the road they are traveling through. The driving trajectory also includes road segments At that time, it represented the truck. with truck With public routes, trucks For trucks The first one it passed through Potential formation targets for each hub; The model training module is used to define an interaction information graph of each truck at the current hub based on the potential platooning object set and the waiting time set, and to train a graph neural network model based on the interaction information graph, and to optimize the internal parameters of the graph neural network model by learning the optimal solution dataset; The strategy output module is used to input the route hubs and departure times of each truck into the graph neural network model, generate its optimal waiting time and predicted platooning objects at each route hub, so as to generate a platooning scheduling strategy.

7. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer-executable instructions that, when executed by a processor, implement the large-scale truck platooning scheduling method based on deep graph learning as described in any one of claims 1 to 5.

8. A data control device, characterized in that, include: One or more memories that store computer-executable instructions non-transitory; One or more processors configured to run computer-executable instructions, wherein the computer-executable instructions are executed by the one or more processors to implement the large-scale truck platooning scheduling method based on deep graph learning as described in any one of claims 1 to 5.

Images