Online car-hailing vehicle dispatch optimization method, device and storage medium based on end-to-end prediction

Abstract

The present invention relates to an optimization method, device and storage medium for online car-hailing scheduling based on end-to-end prediction, and the method steps include: according to the sub-region and time slice division, a supply and demand prediction model based on historical supply and demand characteristics and spatiotemporal characteristics is constructed to obtain the number of online car-hailing orders and the number of drivers in each time slice of each sub-region; with the number of orders and the initial number of drivers in each sub-region and time slice as unknown parameters, the vehicle scheduling optimization model is constructed to maximize the total transaction amount of the platform as the optimization goal, and the vehicle scheduling plan is output; and a loss function with the minimum decision regret value as the goal is adopted, and a penalty term is set according to the degree to which the optimal solution violates the constraint, and the gradient of the loss function with respect to the prediction parameter is estimated by using zero-order estimation, and the prediction model is returned to realize end-to-end training. Compared with the prior art, the present invention realizes end-to-end modeling of the prediction model and the optimization model by coupling the training of the prediction model with the decision-making goal of the scheduling, thereby improving the benefits of the scheduling decision.

Description

Technical Field

[0001] The present invention relates to the field of intelligent traffic control, and in particular to a data-driven end-to-end intelligent prediction and post-optimization online-hailing vehicle scheduling optimization method. Background Art

[0002] With the acceleration of urbanization and the growth of mobile travel demand, online ride-hailing has become an indispensable part of the modern transportation system. However, online ride-hailing faces the dual challenges of supply-demand mismatch and supply-demand uncertainty. On the one hand, there is a significant time-space mismatch between idle vehicles and passenger demand. For example, during the morning rush hour, there may be a situation where supply exceeds demand in the suburbs and supply exceeds demand in the city center; on the other hand, due to the bilateral nature of the online ride-hailing market, both demand and supply have strong uncertainty, and it is difficult for the platform to accurately grasp the real-time dynamic supply and demand situation, which seriously affects the efficiency of the platform's dispatch of vehicles and has a serious negative impact on both the platform's revenue and the passenger travel experience.

[0003] At present, the data-driven approach to solving the dispatch optimization problem of online ride-hailing usually adopts a two-stage prediction-post-optimization framework, in which the prediction model and the optimization model are independent of each other: first, the machine learning model is trained to predict the relevant parameters, and then the prediction results are unidirectionally input into the optimization model to solve the dispatch method. However, this separate framework has inherent defects, that is, the improvement of prediction accuracy does not necessarily lead to an improvement in decision quality, especially when the area with low prediction accuracy happens to have a greater impact on the system revenue, even if the overall prediction accuracy is high, it may lead to suboptimal dispatch decisions. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide an online car-hailing scheduling optimization method based on end-to-end prediction.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] As a first aspect of the present invention, a method for optimizing online car-hailing vehicle scheduling based on end-to-end prediction is provided, the steps comprising:

[0007] According to the pre-divided city sub-regions and time slices, based on historical order data and spatiotemporal information characteristics, a supply and demand forecasting model based on historical supply and demand information is constructed to output the number of online ride-hailing orders and drivers in each time slice of each sub-region on the next day;

[0008] Based on the number of orders in the sub-region, the number of drivers, and the travel time between regions, an online car-hailing vehicle scheduling optimization model is constructed with vehicle scheduling as the decision variable, the number of orders issued in each sub-region and time slice and the initial number of drivers as unknown parameters, and maximizing the total transaction amount of the platform within the time range as the optimization goal. The vehicle scheduling optimization model outputs the vehicle scheduling plan.

[0009] Among them, the loss function of the supply and demand forecasting model adopts the decision regret value that measures the difference between the optimal target value and the target value obtained by the actual decision and includes the penalty for violating the constraints. The gradient of the loss function with respect to the predicted value is derived using zero-order gradient estimation and passed back to the supply and demand forecasting model for training.

[0010] As a preferred technical solution, the supply and demand forecasting model specifically includes: an embedding layer, which converts the sub-area ID and time slice TID into vectors and concatenates them with the historical data sequence as input, and the dimension of the input layer depends on the number of sub-areas and the number of time slices; an input layer, which concatenates the input historical supply and demand sequence, time features and the embedded vectors of the sub-area ID and time slice TID; a hidden layer, which includes two fully connected hidden layers; and an output layer, which outputs the predicted values ​​of the number of orders and the number of drivers corresponding to each sub-area and time slice combination on the next day.

[0011] As a preferred technical solution, the scheduling optimization model is a mixed integer linear programming problem, and its optimization objective function is expressed as:

[0012]

[0013] In the formula, is a decision variable, representing the sub-region To sub-area in Number of vehicle dispatches ; Indicates sub-area In time slice The number of orders issued, Indicates sub-area In time slice The initial number of drivers, these two items are unknown parameters in the optimization problem and are predicted by the supply and demand forecasting model; is a known parameter, representing the sub-region In time slice Average platform transaction volume GMV; The transaction rate.

[0014] As a preferred technical solution, the transaction rate Affected by the supply-demand ratio, it is expressed as:

[0015]

[0016] In the formula, is a monotonically increasing function; is the supply-demand ratio, defined as:

[0017]

[0018] In the formula, Sub-area after scheduling In time slice The number of drivers is calculated as follows:

[0019]

[0020] In the formula, Indicates sub-area and The travel time between Indicates at time From sub-area Departure at time Arrival Sub-Area The number of vehicles; Indicates at time Previously from the sub-area Number of vehicles transferred to other sub-areas; Indicates in time slice Dispatch vehicles to sub-areas The number of vehicles.

[0021] As a preferred technical solution, the scheduling optimization model is set with a constraint on the number of dispatched vehicles:

[0022] By limiting the minimum supply-demand ratio of the starting sub-area, a constraint is imposed on the maximum number of vehicles that can be dispatched from each sub-area:

[0023]

[0024] In the formula, is the supply-demand ratio; Indicates whether there is a vehicle from the sub-area call out; The minimum value of the supply-demand ratio that is set;

[0025] And constrain the total dispatch cost of dispatching vehicles not to exceed the set upper limit .

[0026] As a preferred technical solution, the scheduling optimization model adopts the loss function of the decision regret value, and substitutes the optimal solution obtained under the real parameters and the optimal solution obtained under the estimated parameters of the supply and demand forecasting model into the real scene to calculate GMV, and the difference between the two is expressed as:

[0027]

[0028] In the formula, Indicates the number of orders with unknown parameters and the initial number of drivers The truth value of Indicates the number of orders with unknown parameters and the initial number of drivers The predicted value of and They respectively represent the optimal solutions obtained through the scheduling optimization model based on the true value and predicted value of the unknown parameters.

[0029] As a preferred technical solution, a penalty term is added to the loss function based on the regret value, and the loss function including the penalty is:

[0030]

[0031] In the formula, is the coefficient of the penalty term, Indicates the degree of constraint violation:

[0032]

[0033] When a sub-region is a starting sub-region, if the supply-demand ratio of the sub-region is within the threshold If the person fails to do so, he will be punished.

[0034] As a preferred technical solution, the method uses the chain rule to calculate the loss function Relative to unknown parameters The gradient of the decision variable with respect to the unknown parameter is estimated based on the zero-order estimation method. The gradient of the loss function with respect to the unknown parameter is:

[0035]

[0036]

[0037] In the formula, It follows a normal distribution; is a positive parameter.

[0038] As a second aspect of the present invention, there is provided an electronic device, comprising:

[0039] one or more processors;

[0040] A memory for storing one or more programs;

[0041] When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described above.

[0042] As a third aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the online-hailing vehicle scheduling optimization method based on end-to-end prediction are implemented as described above.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1) In order to meet the challenge of spatiotemporal uncertainty in the supply and demand of online ride-hailing vehicles, this paper designs an end-to-end prediction optimization framework. In the supply and demand prediction model part, a supply and demand prediction model based on historical supply and demand characteristics and spatiotemporal characteristics is constructed according to the sub-region and time slice division. In the scheduling optimization model part, a vehicle scheduling optimization model is established to maximize the total transaction volume of the platform. By coupling the training of the supply and demand prediction model with the decision-making goal of scheduling, end-to-end modeling of the prediction model and the optimization model is realized, thereby improving the decision-making benefits.

[0045] 2) The present invention designs a loss function for the proposed end-to-end intelligent prediction post-optimization framework with the goal of minimizing the decision regret value, and sets a penalty term according to the degree to which the optimal solution violates the constraint. In the gradient estimation part, the gradient of the loss function with respect to the prediction parameter is estimated using zero-order estimation. End-to-end training is achieved by modifying the loss function and gradient estimation method in the prediction model. By coupling the training of the prediction model with the decision-making goal of scheduling, end-to-end modeling of the prediction model and the optimization model is achieved, thereby improving the benefits of scheduling decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a structural schematic diagram of the online car-hailing vehicle scheduling optimization method based on end-to-end prediction of the present invention;

[0047] Figure 2 A schematic diagram of the division of sub-areas for online car-hailing dispatch in the present invention;

[0048] Figure 3 Schematic diagram of the transaction rate model constructed by the present invention. DETAILED DESCRIPTION

[0049] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0050] Example 1

[0051] The present invention provides an end-to-end intelligent prediction and post-optimization online car-hailing dispatch optimization method, which realizes joint modeling of prediction and optimization by modifying the loss function of the prediction model to the decision error, thereby improving the dispatch effect of online car-hailing. Figure 1 As shown, this method mainly includes the following four stages:

[0052] Supply and demand forecasting model: Based on historical order data, spatiotemporal information and other features, a supply and demand forecasting model based on historical supply and demand information is constructed to output the number of online car-hailing orders and the number of drivers in each time slot in each region on the next day.

[0053] Scheduling optimization model: Based on the pre-divided urban sub-regions and time slices, and based on information such as the number of orders in the sub-regions, the number of drivers, and the travel time between regions, an online car-hailing vehicle scheduling optimization model with unknown parameters is established to output a vehicle scheduling plan.

[0054] Loss function design: Based on the prediction model and optimization model, a decision-oriented loss function, namely the regret value, is designed to measure the difference between the optimal target value and the target value obtained by the actual decision.

[0055] Gradient estimation method: According to the designed loss function, a zero-order gradient estimation method is designed to derive the gradient of the loss function with respect to the predicted value, and the gradient is returned in combination with the chain rule.

[0056] 2. Modeling of supply and demand forecasting model:

[0057] 2.1. A large city is taken as the research object, such as Figure 2 As shown, the city is divided into several sub-areas, and the day is divided into A supply and demand forecasting model based on historical supply and demand information and spatiotemporal characteristics is constructed. In this embodiment, the entire city is divided into a cellular network with a hexagonal side length of 7 km, and a day is divided into 144 time slices, each time slice is 10 minutes long.

[0058] 2.2. The historical number of orders issued in each sub-area and time slice in the past 14 days and the number of drivers are known quantities and can be obtained through observation.

[0059] 2.3. Construct a prediction model. The features include the supply and demand data of the past 14 days, time features (such as date, hour, day of the week), and the embedding of the sub-region ID and time slice TID. The input layer includes the historical supply and demand series, time features, sub-region ID and time slice TID, which are converted into vectors through the embedding layer and concatenated with the historical data series as input. The dimension of the input layer depends on the number of sub-regions and time slices. The hidden layer contains two fully connected hidden layers, each with 128 neurons, and the activation function uses ReLU. The output layer is the predicted value of the number of orders and drivers corresponding to the sub-region and time slice combination on the second day.

[0060] 3. Modeling of scheduling optimization model:

[0061] The topological relationship between each sub-region is known. is the set of sub-regions, is the set of time slices studied. The scheduling optimization model is a mixed integer linear programming problem. The goal is to maximize the total transaction volume (Gross Merchandise Volume, GMV) of the platform within the study time range, which can be expressed as:

[0062]

[0063] In the formula, is a decision variable, representing the sub-region To sub-area in Number of vehicle dispatches , Indicates sub-area In time slice The number of orders issued, Indicates sub-area In time slice The initial number of drivers is unknown in the optimization problem and needs to be predicted by the prediction model. is a known parameter, representing the sub-region In time slice The average platform GMV can be estimated based on the historical order data of online ride-hailing vehicles. The calculation method is as follows: In time slice The historical total GMV divided by the historical total number of completed orders. is the transaction rate, which is affected by the supply-demand ratio and is expressed as:

[0064]

[0065] In the formula, is a monotonically increasing function, is the supply-demand ratio. According to the properties of the monotonically increasing function, the larger the supply-demand ratio, the higher the transaction rate. The supply-demand ratio is defined as:

[0066]

[0067] The transaction rate function can be fitted based on the historical order data of online ride-hailing. Figure 3 As shown, taking the piecewise function as an example, the observation point contains the historical total transaction rate and historical supply-demand ratio information of each sub-region, and the parameters of the piecewise function are obtained by fitting based on the least squares method. Sub-area after scheduling In time slice The number of drivers can be calculated as:

[0068]

[0069] In the formula, Indicates sub-area and The travel time between the two is set to the average travel time of the past 7 days in this embodiment; Indicates at time From sub-area Departure at time Arrival Sub-Area The formula represents the number of vehicles in the sub-area In time The number of vehicles is affected by three parts: the initial number of drivers , at time Previously from sub-area Number of vehicles transferred to other sub-areas and in time slices Dispatch vehicles to sub-areas Number of vehicles .

[0070] Set up two 0-1 auxiliary variables and , indicating whether there is a vehicle from the sub-area Call in or call out, they and decision variables There are the following relations:

[0071]

[0072]

[0073] In the formula, is a very large positive number, which is set to 100000 in this embodiment; Indicates whether there is a vehicle from the sub-area Call out, Indicates whether there is a vehicle transferred into the sub-area , when the value is 1 it means yes, and when the value is 0 it means no.

[0074] In order to prevent excessive dispatch of vehicles, a constraint on the number of dispatched vehicles is designed. In order to prevent excessive impact on the transaction rate of the dispatch starting sub-area, a constraint is imposed on the maximum number of vehicles that can be dispatched from each sub-area, which is achieved by limiting the minimum supply-demand ratio:

[0075]

[0076] In the formula, The minimum value of the supply-demand ratio can be set as the segmentation point of the segmentation function. The slope on the right side of the segmentation point is smaller than the slope on the left side. When the supply-demand ratio of the sub-region is on the right side of the segmentation point, the transaction rate of the transferred vehicles to the sub-region is small. The constraint restricts the minimum supply-demand ratio of the starting sub-region to not be less than In addition, since dispatching vehicles requires costs, the total dispatch cost is set to not exceed the upper limit , in this embodiment, it is set to 5000 yuan:

[0077]

[0078] In the formula, is the scheduling cost per minute, which is set to 1 yuan / minute in this embodiment.

[0079] 4. Design of loss function

[0080] The present invention designs a loss function of the decision regret value, which is the optimal solution obtained under the real parameters and the optimal solution obtained under the estimated parameters of the prediction model. They are substituted into the real scene to calculate the GMV, and the difference between the two is expressed as:

[0081]

[0082] In the formula, Indicates the number of orders with unknown parameters and the initial number of drivers The true value of , corresponding to, Represents the predicted value of the parameter. and They represent the optimal solution obtained by the optimization model based on the true value and predicted value of the unknown parameter, which can be expressed as:

[0083]

[0084]

[0085] In the formula, and They represent the feasible domain of the optimization problem consisting of the predicted values ​​and true values ​​of the parameters respectively.

[0086] However, since some constraints contain unknown parameters, such as the constraint of the minimum supply-demand ratio, the predicted optimal solution may not be feasible in actual scenarios. In order to allow the model to learn information about violations of constraints during training, a penalty term is added to the loss function based on the regret value. The loss function including the penalty is:

[0087]

[0088] In the formula, is the coefficient of the penalty term, Indicates the degree of constraint violation:

[0089]

[0090] This formula measures the degree to which the starting sub-region violates the supply-demand ratio constraint. When a sub-region is the starting sub-region, if its supply-demand ratio state is within the threshold If this happens, punishment is necessary.

[0091] 5. Design of the Gradient Estimation Method

[0092] In order to pass the loss gradient back to the supply and demand prediction model, the loss function is calculated using the chain rule Relative to unknown parameters The gradient of is expressed as:

[0093]

[0094] In the formula, the three terms on the right side of the equal sign correspond to the gradients of the loss function with respect to the decision variable, the decision variable with respect to the unknown parameter, and the unknown parameter with respect to the hyperparameter of the prediction model. Since the optimization model is a linear model, the first term can be directly derived from the expression of the objective function, and the third term can be obtained using the automatic differentiation tool of the neural network. For the second term, since there are often multiple optimal solutions in mixed integer programming problems, a method based on zero-order estimation is proposed to approximate the estimate. , expressed as:

[0095]

[0096] In the formula, It follows a normal distribution. is a parameter, which is a positive number. Combined with the chain rule, the gradient of the loss function with respect to the unknown parameter is:

[0097]

[0098] Replace the loss function and gradient estimation method in the prediction model with the designed regret value containing penalty The gradient obtained by the zero-order estimate is used to realize the training of the end-to-end intelligent prediction and post-optimization online car-hailing scheduling optimization framework.

[0099] Finally, the trained end-to-end intelligent prediction and optimization framework for online car-hailing scheduling is adopted. The supply and demand prediction model predicts the number of online car-hailing orders and drivers in each time slot in each area on the next day, and inputs the scheduling optimization model to predict the output vehicle scheduling plan.

[0100] Example 2

[0101] As a second aspect of the present invention, the present application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned online car-hailing scheduling optimization method based on end-to-end prediction. In addition to the above-mentioned processor, memory and interface, any device with data processing capability in which the device in the embodiment is located may also include other hardware according to the actual function of the device with data processing capability, which will not be described in detail.

[0102] Example 3

[0103] As a third aspect of the present invention, the present application also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implements the above-mentioned online car-hailing scheduling optimization method based on end-to-end prediction. The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities as described in any of the foregoing embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be an external storage device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), an SD card, a flash card (FlashCard), etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and may also be used to temporarily store data that has been output or is to be output.

[0104] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for optimizing online car-hailing vehicle scheduling based on end-to-end prediction, characterized in that the steps include: According to the division of sub-regions and time slices, based on historical order data and spatiotemporal information characteristics, a supply and demand forecasting model based on historical supply and demand characteristics and spatiotemporal characteristics is constructed, and the number of online car-hailing orders and drivers in each time slice of each sub-region in the future is obtained; Based on the number of orders in the sub-region, the number of drivers, and the travel time between regions, an online car-hailing vehicle scheduling optimization model is constructed with vehicle scheduling as the decision variable, the number of orders issued in each sub-region and time slice and the initial number of drivers as unknown parameters, and maximizing the total transaction amount of the platform within the time range as the optimization goal. The vehicle scheduling optimization model outputs the vehicle scheduling plan. The optimization objective function of the vehicle scheduling optimization model is expressed as: ， In the formula, is a decision variable, representing the sub-region To sub-area in Number of vehicle dispatches ; Indicates sub-area In time slice The number of orders issued, Indicates sub-area In time slice The initial number of drivers, these two items are unknown parameters in the optimization problem and are predicted by the supply and demand forecasting model; is a known parameter, representing the sub-region In time slice Average platform transaction volume GMV per order; is the transaction rate, which is affected by the supply-demand ratio and is expressed as: ， In the formula, is a monotonically increasing function; is the supply-demand ratio, defined as: ， In the formula, Sub-area after scheduling In time slice The number of drivers is calculated as follows: ， In the formula, Indicates sub-area and The travel time between Indicates at time From sub-area Departure at time Arrival Sub-Area The number of vehicles; Indicates at time Previously from sub-area Number of vehicles transferred to other sub-areas; Indicates in time slice Dispatch vehicles to sub-areas The number of vehicles; Among them, the loss function of the supply and demand forecasting model aims to minimize the decision regret value and includes penalties for violating constraints; the gradient of the loss function with respect to the predicted value is derived using zero-order gradient estimation and passed back to the supply and demand forecasting model for training.

2. The method for optimizing online car-hailing vehicle scheduling based on end-to-end prediction according to claim 1, characterized in that: The supply and demand forecasting model specifically includes: an embedding layer, which converts the sub-region ID and time slice TID into a vector and concatenates it with the historical data sequence as input, and the dimension of the input layer depends on the number of sub-regions and the number of time slices; an input layer, which concatenates the input historical supply and demand sequence, time features and the embedding vector of the sub-region ID and time slice TID; a hidden layer, which includes two fully connected hidden layers; and an output layer, which outputs the predicted values ​​of the number of orders and the number of drivers corresponding to each sub-region and time slice combination on the next day.

3. The method for optimizing online car-hailing vehicle scheduling based on end-to-end prediction according to claim 1, characterized in that: The vehicle scheduling optimization model is set with the constraint of the number of scheduled vehicles: By limiting the minimum supply-demand ratio of the starting sub-area, a constraint is imposed on the maximum number of vehicles that can be dispatched from each sub-area: ， In the formula, is the supply-demand ratio; Indicates whether there is a vehicle from the sub-area call out; The minimum value of the supply-demand ratio that is set; And constrain the total dispatch cost of dispatching vehicles not to exceed the set upper limit .

4. The method for optimizing online car-hailing vehicle scheduling based on end-to-end prediction according to claim 1, characterized in that: The vehicle scheduling optimization model adopts the loss function of decision regret value, and substitutes the optimal solution obtained under the real parameters and the optimal solution obtained under the estimated parameters of the supply and demand forecasting model into the real scene to calculate GMV, and makes the difference between the two, which is expressed as: ， In the formula, Indicates the number of orders with unknown parameters and the initial number of drivers The truth value of Indicates the number of orders with unknown parameters and the initial number of drivers The predicted value of and They respectively represent the optimal solutions obtained by the vehicle scheduling optimization model based on the true value and predicted value of the unknown parameters.

5. The method for optimizing online car-hailing vehicle scheduling based on end-to-end prediction according to claim 4, characterized in that: Add a penalty term to the loss function based on the regret value. The loss function including the penalty is: ， In the formula, is the coefficient of the penalty term, Indicates the degree of constraint violation: ， When a sub-region is a starting sub-region, if the supply-demand ratio of the sub-region is within the threshold If the person fails to do so, he will be punished.

6. The method for optimizing online car-hailing vehicle scheduling based on end-to-end prediction according to claim 4 or 5, characterized in that: The method uses the chain rule to calculate the loss function Relative to unknown parameters The gradient of the decision variable with respect to the unknown parameter is estimated based on the zero-order estimation method. The gradient of the loss function with respect to the unknown parameter is: ， ， In the formula, Follows normal distribution; is a positive parameter.

7. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the online car-hailing scheduling optimization method based on end-to-end prediction are implemented as described in any one of claims 1 to 6.

Images