Vehicle dispatching method, vehicle dispatching server and computer readable storage medium

Abstract

The present application relates to the technical field of vehicle scheduling, in particular to a vehicle scheduling method, a vehicle scheduling server and a computer readable storage medium, comprising differentiating original data into grid data and abstracting the state of each grid as input to obtain individual action; taking the joint action set or individual state as new input into the double evaluation network to obtain the reward value; then through the operation of the double evaluation network and the uncertainty weighting, the target value and the loss function of the evaluation network are obtained, and the two functions are compared to obtain the best scheduling scheme. The present application can realize the large-scale coordination and scheduling problem of multiple agents, and the uncertainty weighting module can more conservatively estimate the out-of-distribution data set to avoid the problem of excessive estimation of the value function. It performs excellently in complex scenarios and can capture complex dynamic supply and demand changes, thereby realizing a better scheduling mode.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle scheduling, in particular to a vehicle scheduling method, a vehicle scheduling server and a computer readable storage medium. BACKGROUND

[0002] In recent years, with the rapid development of the Internet, people's life has changed a lot. The "online car" has entered people's life. The online car platform optimizes the allocation of traffic resources by sending orders to idle drivers, and also improves the situation of traffic congestion.

[0003] A key challenge of reasonable scheduling of vehicles is to balance supply and demand, that is, to balance the number of drivers and passenger orders. Large cities will generate millions of orders every day, and among these large-scale orders, there will be many orders canceled because there is no driver nearby, or there is no suitable order for the driver to receive nearby, resulting in idle traffic resources. If these idle drivers are dispatched to areas with high demand, it will greatly increase the response rate of orders, improve the utility of transportation capacity, increase the income of drivers and the satisfaction of passengers. Vehicle scheduling is a complex dynamic process, and the current vehicle scheduling decision will affect the gap between future traffic supply and demand. Supervised learning methods are difficult to capture and simulate these real-time dynamic changes, but reinforcement learning excels in this area and can capture complex dynamic supply and demand changes, thereby making good vehicle scheduling to ensure the full use of traffic resources. In the classification of deep reinforcement learning, it is mainly divided into two ways based on value and policy. The Actor-Critic algorithm combines the two ways.

[0004] In related technologies, context deep Q learning and context multi-agent Actor-Critic algorithm are used, which has achieved excellent results in vehicle scheduling. However, in the value-based reinforcement learning method (such as deep Q learning), function approximation errors can lead to overestimation of value estimates and suboptimal strategies. SUMMARY

[0005] The present application aims to at least solve the problem of large-scale imbalance between supply and demand of vehicles in the prior art, and to reduce the influence of OOD (out-of-distribution data, OOD data: data outside the distribution of the offline data set) data pairs on value estimation by weighting the uncertainty of historical data, thereby limiting the influence of overestimation of the contrast. To this end, the present application proposes a vehicle scheduling method, a vehicle scheduling server and a computer readable storage medium.

[0006] The following explanations are made for the process of the present application, and the corresponding application flowchart is as follows Figure 1As shown:

[0007] The original vehicle order scheduling data is differentiated into grid data to obtain initial vehicle distribution and order information, and global state and local state of the corresponding grid are obtained from the information. The local state is then input into a strategy network (Actor network) to obtain the corresponding action, and the corresponding Q value is input into a double evaluation network (Critic network) to obtain the reward value that can be obtained by selecting the corresponding action in the current state and The target value is obtained by calculating the two different values , so as to obtain the vehicle scheduling timetable corresponding to each grid data. The corresponding matching order is scheduled to achieve the best vehicle scheduling operation. After completing a set of vehicle scheduling operations, the current state is returned to the Actor network to update the corresponding global state and local state, and then the next scheduling operation is performed.

[0008] According to the vehicle scheduling method of the first aspect of the embodiment of the application, the method comprises:

[0009] The initial vehicle scheduling information and order information are obtained as original data, the original data is differentiated into grid data, the vehicle scheduling information and order information in the data set are divided into the grid according to the latitude and longitude according to the map data distribution, the map is divided into a hexagonal grid, the time information of a day is divided into a plurality of time segments, the scheduling is performed once every time segment, and one day in the data set is extracted every day to fit the actual data, so as to extract the global state;

[0010] The individual state of each grid is obtained from the global state as input, wherein the individual state includes the information of itself and six neighbor nodes around it;

[0011] The individual state is obtained through a fully connected network layer to obtain an individual action (Action): the agent (Agent) in each grid is homogeneous and has the same scheduling strategy;

[0012] The joint action set of all agents / individual state is input into a double evaluation network (double Critic network) to obtain the reward value that can be obtained by selecting the corresponding action in the current status 、 The reward value and obtained is input into a harmonic network module to obtain the target value ;

[0013] The obtained The input is fed into the uncertainty weighting module to obtain the loss function (Critic Loss) of the evaluation network. The loss function (Actor Loss) of the policy network and the loss function (Critic Loss) of the evaluation network are compared to obtain the final scheduling scheme.

[0014] According to the vehicle scheduling method of this invention, the order delivery problem is modeled as a Markov decision process. Based on this, an uncertain weighted harmonic dual-Critic algorithm is proposed to solve the large-scale coordinated scheduling problem of multiple agents. The uncertainty weighting module can more conservatively estimate outlier datasets to avoid overestimation of the value function. A harmonic dual-Critic architecture is designed to weight the uncertainty. The goal of vehicle scheduling is to determine in advance how many idle vehicles will be scheduled to areas with higher demand to serve more orders. This invention performs well in complex scenarios, capturing complex dynamic supply and demand changes, thereby achieving a better scheduling method.

[0015] According to some embodiments of the present invention, the joint action set / individual state of all agents is used as a new input and fed into a dual-critic network to obtain the reward value that may be obtained by selecting the corresponding action in the current state. and The reward value will be obtained and The target value is obtained by inputting it into the harmonic network module. This involves inputting the same action and state information into two Critic networks with the same structural layers. Due to different parameters during training, the two Critic networks will output different values. These two different values ​​are then compared, and the larger value is multiplied by the weight coefficient α, and the smaller value is multiplied by the weight coefficient β. The two values ​​are then added together to obtain the target value. .

[0016] By designing a dual-Critic harmonic network, the two Critic networks output different values. These two different values ​​are then compared, and the larger value is multiplied by a weight coefficient α, while the smaller value is multiplied by a weight coefficient β. The sum of these two values ​​yields the target value. This avoids overestimating the value while not limiting the generalization ability of the Q function.

[0017] According to some embodiments of the present invention, the obtained reward value and The target value is obtained by inputting it into the harmonic network module. The specific formula is as follows:

[0018]

[0019] in All are weighting coefficients. , .

[0020] According to some embodiments of the present invention, the action space of an agent It contains seven discrete action values, pointing to the i-th network itself and six neighboring grids respectively; the obtained action is input into the uncertainty weighting module to obtain the corresponding Actor Loss (that is, the weight is obtained by the uncertainty of the obtained action to reduce the impact of out-of-distribution data on value estimation).

[0021] By designing an uncertainty-weighted module, a constant quantity is maintained for different input instances of the same module. When the homoscedasticity uncertainty of the module is high, the task has a smaller impact on the network weight update, and this is used to balance the magnitude of loss and the corresponding gradient.

[0022] According to some embodiments of the present invention, the uncertainty weighting module includes the following:

[0023] The uncertainty of a module can be approximated by the variance estimated from the Q-value:

[0024] V a r [ Q ( s , a ) ] 2 + 1 T t = 1 T <m> Q t < / m> ( s , a ) ? E [ <m> Q t < / m> ( s , a ) ]

[0025] in: V a r [ Q ( s , a ) ] These are the uncertainty weights for the module, used to estimate the weights of different state-action pairs. Due to inherent noise in the data, The degree of uncertainty of the model's predictions. This is for predicting the mean. This refers to the uncertainty of the model regarding out-of-distribution (OOD) data samples from the offline dataset.

[0026] Once the uncertainty of the data pairs outside the offline dataset distribution is obtained, a policy distribution based on uncertainty weighting can be defined. :

[0027] ' ( a | s ) = V a r [ Q 0 ' ( s , a ) ] ( a | s ) / Z ( s )

[0028]

[0029] in: This represents a policy distribution that incorporates uncertainty weights on top of the original policy. For uncertain constants, V a r [ Q 0 ' ( s | a ) ] This indicates that agent 0 is in the policy Uncertainty under different distribution conditions To select an action in state s The probability, To select all different actions in state s The probability distribution;

[0030] By incorporating the uncertainty of prediction into the Actor-Critic framework, we can obtain new loss functions for the evaluation network (Critic) and the policy network (Actor), as shown below:

[0031]

[0032]

[0033]

[0034] in: For the new evaluation network loss function, The loss function for the new policy network, , , This means that the uncertainty of estimating the data is used as a regularization term to achieve the purpose of conservative estimation, and the uncertainty of the prediction is incorporated into the framework to obtain a new loss value. Specifically, it is represented as a regularization operation under the policy conditions of the corresponding state or action. For uncertain constants, V a r [ Q ' ( s ' , a ' ) ] For the uncertainty weight of the next state-action value pair, The error in each time difference is referred to as TD Error. The function represents the loss between the current value of the standard and the target value; The value generated by choosing an action for the current state s. The reward value obtained after performing an action in state s. As a discount factor, This represents an estimate of the value of choosing the action to be performed in the next state; V a r [ Q ( s | a ) ] This indicates that the agent is in the policy Uncertainty-weighted weights for different distributions;

[0035] The Critic loss function adds a weight before the TD Error, while the Actor loss function adds an uncertainty weight before the Q value. In this invention, the... 0.8 is preferred.

[0036] According to some embodiments of the present application, the initial vehicle scheduling information and order information are obtained as raw data, and the raw data is differentiated into grid data, including: dividing a map into a hexagonal grid, specifically, dividing the map into a hexagonal grid of 17*15, a total of 255 hexagonal grids, and dividing the vehicle information and order information in the data set into the grid according to the latitude and longitude; dividing the time information of a day into 144 time segments, and performing scheduling once every 10 minutes for each segment, and extracting a day from the data set every day to fit the actual data, so as to extract the global state.

[0037] In addition, the evaluation index of the present application is the average response rate of the order and the GMV, the average response rate of the order refers to the proportion of the order responded by the vehicle; the GMV is the total price of all orders, and each order will get the corresponding order price according to the time length and the service quality of the order and other factors.

[0038] The vehicle scheduling server according to the second aspect of the embodiments of the present application comprises a processor, a memory and a computer program stored on the memory and executable on the processor, wherein the processor executes the above-mentioned method when running the computer program.

[0039] The computer-readable storage medium according to the third aspect of the embodiments of the present application has a computer program stored thereon, wherein the computer program is run by the processor to realize the above-mentioned method.

[0040] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0042] Figure 1 is a flow chart of the vehicle scheduling method according to the embodiments of the present application;

[0043] Figure 2 is a framework diagram of the vehicle scheduling method according to the embodiments of the present application;

[0044] Figure 3 is a reconciliation network model diagram according to the embodiments of the present application;

[0045] Figure 4 is a super parameter analysis diagram of the total price of orders in the reconciliation network according to the embodiments of the present application;

[0046] Figure 5 is a super parameter analysis diagram for harmonizing average response rate in a network according to an embodiment of the present application. DETAILED DESCRIPTION

[0047] Embodiments of the present application are described in detail below with reference to the attached drawing figures, which are incorporated in the description for further understanding wherefore, it should be made clear that the embodiments described herein are merely exemplary of the present application and are intended to be illustrative only and are not intended to limit the present application.

[0048] The terms "first", "second", "third", and the like in the description and in the claims of the present specification and the accompanying drawings are used for distinguishing between similar objects talking to different objects and are not necessarily used to describe a particular sequential order. Also, the terms "comprises", "comprising", "includes", "including", "has", "having" and the like are used synonymously to denote a non-exclusive inclusion, such that anything including any recited steps or units is also encompassed when optionally comprising additional steps or units not recited.

[0049] Only parts of the apparatuses are shown in the drawings and not all of the parts of the apparatuses are shown. Before explaining example embodiments in more detail, it is to be understood that the example embodiments are not limited in application to the details of construction and the arrangement of components set forth in the following description or illustrated in the accompanying drawings. The example embodiments are described as processes depicted as flow diagrams or as methods, functions, procedures, subroutines, subprograms, etc. Although the processes are described in a particular sequential order, many of the processes can be performed concurrently, in parallel, or concurrently, without a particular bias. In addition, the illustrated processes can be implemented in programs of instructions or software, firmware, hardware, or a combination thereof, which can be stored in any type of computer-readable medium or other storage device(s) or memory that can be accessed by a computer, processor, or other implementation device. The processes described can correspond to methods, functions, procedures, subroutines, subprograms, etc.

[0050] The terms "component," "module," "system," "unit," and the like as used herein are used to represent a computer-related entity, hardware, firmware, a combination of hardware and software, software, or a software in execution. For example, a unit can be, but is not limited to, a process running on a processor, a processor, an object, an executable, a thread of execution, a program, and / or a computer. Also, the units can be executed from a variety of computer-readable media having various data structures stored thereon. The units can communicate over a local and / or remote processes, for example, by signals having one or more data packets (e.g., from a second unit data from a local system, a distributed system, and / or a network with another unit interacting with the other system. For example, the Internet interacts with other systems through signals with other systems.

[0051] Embodiment 1

[0052] Reference Figure 1 and Figure 2This embodiment proposes a vehicle scheduling method based on an uncertainty-weighted harmonic dual-Critic algorithm. The original data is divided into grid-like data, and a global state is extracted from it. The individual state of each grid is obtained from the global state as input, where the individual state contains information about itself and its six neighboring nodes. Actions are obtained: the individual state is passed through a fully connected layer to obtain the individual's action, which is then input into an uncertainty-weighted module to obtain the Actor Loss. The joint action set of all agents, consisting of individual states, is used as new input and fed into the dual-Critic network to obtain the reward value. and The reward value obtained and The target value is obtained by inputting it into the harmonic network module. The target value obtained The input is fed into the uncertainty weighting module to obtain the CriticLoss. The Actor Loss and Critic Loss are compared to obtain the optimal scheduling scheme. Specifically, this includes:

[0053] Initial vehicle dispatch and order information were obtained as raw data. This raw data was then divided into gridded data: Following the map data distribution, the map was divided into 17*15 hexagonal grids, resulting in 255 hexagonal grids. The vehicle dispatch and order information in the dataset were then assigned to these grids based on latitude and longitude. The time information for one day was divided into 144 time segments, with dispatch executed every 10 minutes in each segment. One day was extracted from the dataset each day to fit the actual data and extract the global state. Each 144 segments constituted one epoch. One round of experiments included 15 epochs for the training set and 15 epochs for the test set, with the average value of the 15 epochs from the test set being used.

[0054] The individual state of each grid is obtained from the global state as input, where the individual state contains information about itself and its six neighboring nodes;

[0055] Individual states are transformed into individual actions through fully connected network layers: agents within each grid are homogeneous and possess the same scheduling policy; the agent's action space... It contains seven discrete action values, pointing to the i-th network itself and six neighboring grids respectively; the obtained action is input into the uncertainty weighting module to obtain the corresponding Actor Loss (that is, the weight is obtained by the uncertainty of the obtained action to reduce the impact of out-of-distribution data on value estimation).

[0056] The uncertainty of a module can be approximated by the variance estimated from the Q-value:

[0057] V a r [ Q ( s , a ) ] 2 + 1 T t = 1 T <m> Q t < / m> ( s , a ) ? E [ <m> Q t < / m> ( s , a ) ]

[0058] in: V a r [ Q ( s , a ) ] The uncertainty weights of the module are used to estimate the weights of different state-action pairs. Due to inherent noise in the data, The degree of uncertainty of the model's predictions. This is for predicting the mean. This refers to the uncertainty of the model regarding out-of-distribution (OOD) data samples in the offline dataset;

[0059] Once the uncertainty of the data pairs outside the offline dataset distribution is obtained, a policy distribution based on uncertainty weighting can be defined. :

[0060] ' ( a | s ) = V a r [ Q 0 ' ( s , a ) ] ( a | s ) / Z ( s )

[0061]

[0062] in: This represents a policy distribution that incorporates uncertainty weights on top of the original policy. For uncertain constants, V a r [ Q 0 ' ( s | a ) ] This indicates that agent 0 is in the policy Uncertainty under different distribution conditions To select an action in state s The probability, To select all different actions in state s The probability distribution;

[0063] By incorporating the uncertainty of prediction into the Actor-Critic framework, we can obtain new loss functions for the evaluation network and the policy network, as shown below:

[0064]

[0065]

[0066]

[0067] in: For the new evaluation network loss function, The loss function for the new policy network, , , The uncertainty of estimating data is represented as a regularization term to achieve conservative estimation, and the uncertainty of prediction is integrated into the framework to obtain a new loss value, which is represented as a regularization operation under the condition of the corresponding state or action policy. is an uncertainty constant, V a r [ Q ' ( s ' , a ' ) ] is the uncertainty weight of the next state-action value pair, is the error in each time difference, referred to as TD Error, The function represents the loss of the standard current value and the target value. is the value generated after selecting an action for the current state s, is the reward value obtained after performing an action in state s, is the discount factor, represents the estimate of the value of the corresponding action selected for the next state. V a r [ Q ( s | a ) ] represents the uncertainty of the agent in the policy distribution.

[0068] The loss function of the Critic is weighted in front of the TD Error, and the loss function of the Actor is weighted in front of the Q value. In the present application, is 0.8.

[0069] The joint action set / individual state is input into the double evaluation network (double Critic network) to obtain reward values and ; the obtained reward values and are input into the harmonic network module to obtain the target value , as shown in Figure 3 , which includes inputting the same action and state information into two Critic networks with the same structure. Due to the difference in parameters during training, the two Critic networks will output different values. Then, the larger value is multiplied by the weight coefficient α, and the smaller value is multiplied by the weight coefficient β to obtain the target value , which avoids overestimation of the value while not limiting the generalization ability of the Q function. The formula is as follows:

[0070]

[0071] Hyperparameters Analysis:

[0072] Take a 7-day data set to train in the case of 100% vehicles, Take a value every 0.1 interval between [0, 1], For . By Figure 4 And Figure 4 It can be obtained that when The effect is obviously better than taking the value of 0 or 1 when the value interval is [0.1, 0.9]. When The best effect can be achieved. In addition, when = 0, that is, the target value is equal to the minimum value in the double Critic network, the effect is significantly reduced.

[0073] The obtained is input into the uncertainty weighting module to obtain the loss function (CriticLoss) of the evaluation network. The loss function of the policy network and the loss function of the evaluation network are compared to finally obtain the best scheduling scheme.

[0074] Embodiment 2

[0075] The embodiment provides a vehicle scheduling server, comprising a processor, a memory and a computer program stored on the memory and executable on the processor, wherein the processor executes the above-mentioned method when executing the computer program.

[0076] Embodiment 3

[0077] The embodiment provides a computer-readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement the above-mentioned method.

[0078] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the application.

[0079] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "illustrative embodiment", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms does not necessarily refer to the same embodiment or example.

[0080] It is apparent that the described embodiments are only some, but not all, of the embodiments of the present application. Reference to "an embodiment" in this text means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present application. The appearances of the phrase "in an embodiment" in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive of one another. As those skilled in the art will appreciate, embodiments described herein can be combined with other embodiments in various ways. All other embodiments obtained by combining the embodiments described herein in various ways are within the scope of the present application.

[0081] Although the embodiments of the present application have been shown and described, it will be appreciated by those skilled in the art that changes can be made in these embodiments without departing from the principles and the scope of the application, which is defined by the claims and their equivalents.

Claims

1. A vehicle dispatching method characterized by comprising: The method comprises the following steps: obtaining initial vehicle scheduling information and order information as raw data, and differentiating the raw data into grid data: according to the distribution of map data, the map is divided into hexagonal grids, and the vehicle scheduling information and order information in the data set are divided into the grids according to the longitude and latitude; time information in a day is divided into a plurality of time segments, and scheduling is performed every other time segment; one day is extracted from the data set every day to fit the actual data, so as to extract the global state; obtaining the individual state of each grid from the global state as input, wherein the individual state includes the information of itself and six neighbor nodes around it; the individual state obtains individual action through a fully connected network layer: the agent in each grid is homogeneous, and has the same scheduling strategy; The joint action set of all agents or individual state is input into the double evaluation network as a new input to obtain a reward value obtained by selecting a corresponding action in the current state And The obtained reward value And Is input into the harmonic network module to obtain the target value Specifically, the same action and state information are input into two evaluation networks with the same structure, and due to different parameters in training, the two evaluation networks output different values, and the two different values are compared, the larger value is multiplied by a weight coefficient , the smaller value is multiplied by a weight coefficient And then added to obtain the target value , the formula is as follows: , wherein: , are weight coefficients, , ; The obtained The loss function of the evaluation network is obtained by inputting into the uncertainty weighting module, and the loss function of the policy network and the loss function of the evaluation network are compared to obtain a final scheduling scheme.

2. The vehicle dispatching method of claim 1, wherein Action space of the agent contains seven discrete action values, one pointing to the i-th network itself and six to the neighboring grids. the obtained action is input into an uncertainty weighting module to obtain the loss function of the corresponding policy network.

3. The vehicle dispatching method of claim 2, wherein The uncertainty weighting module comprises the following contents: the uncertainty of the module is obtained through the variance of the Q value estimation: , where: is the uncertainty weight of the module, used to estimate the weight of different state-action pairs, is the inherent noise in the data, is the uncertainty of the model in its prediction, is the predicted mean; is the uncertainty of the model about data outside the offline dataset distribution; after obtaining the uncertainty of data outside the offline dataset distribution, define the policy distribution based on uncertainty weighting : wherein: represents a policy distribution with uncertainty weight added on top of the original policy, is an uncertainty constant, represents the uncertainty of the 0th agent in the policy distribution, is the probability of choosing action in state s; is the probability distribution of choosing all different actions in state s; The predicted uncertainty is integrated into the Actor-Critic framework to obtain the loss function of the new evaluation network and the loss function of the policy network, which are as follows: in: For the new evaluation network loss function, The loss function for the new policy network, , , This means that the uncertainty of estimating the data is used as a regularization term to achieve the purpose of conservative estimation, and the uncertainty of the prediction is incorporated into the framework to obtain a new loss value. Specifically, it is represented as a regularization operation under the policy conditions of the corresponding state or action. For uncertain constants, For the uncertainty weight of the next state-action value pair, The error in each time difference is referred to as TD Error. The function represents the loss between the current value of the standard and the target value; The value generated by choosing an action for the current state s. The reward value obtained after performing an action in state s. As a discount factor, This represents an estimate of the value of choosing the action to be performed in the next state; This indicates that the agent is in the policy Uncertainty-weighted weights for distributed scenarios; the loss function for evaluating the network adds weights before the TD Error, while the loss function for the policy network adds uncertainty weights before the Q value.

4. The vehicle dispatching method of claim 3, wherein The is 0.

8.

5. The method of claim 1, wherein, The method for obtaining initial vehicle scheduling information and order information as raw data and differentiating the raw data into grid data comprises the following steps: the map is divided into hexagonal grids, specifically, the map is divided into 17*15 hexagonal grids, a total of 255 hexagonal grids, and the vehicle information and order information in the data set are divided into the grids according to the longitude and latitude.

6. The vehicle dispatching method of claim 5, wherein The method for obtaining initial vehicle scheduling information and order information as raw data and differentiating the raw date into grid data further comprises the following steps: time information in a day is divided into 144 time segments, and scheduling is performed every 10 minutes; one day is extracted from the data set every day to fit the actual data, so as the to extract the global state.

7. A vehicle dispatch server comprising a processor, a memory, and a computer program stored on the memory and executable on the processor, wherein, The processor executes the method of any one of claims 1-6 when running the computer program.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to implement the method of any one of claims 1-6.

Images