Autonomous driving network car operation profit prediction and vehicle distribution optimization method and system

Abstract

This invention relates to the field of intelligent operation and scheduling technology for autonomous ride-hailing services, specifically to a method and system for predicting operational profits and optimizing vehicle distribution for autonomous ride-hailing services. The invention first constructs a multi-dimensional demand distribution vector and continuously updates the prototype demand distribution vector for holidays. It then calculates the date coupling coefficient using cosine similarity, and uses a three-branch MLP to output profit predictions for each scenario, which are then weighted and fused to obtain a comprehensive profit. The Big-M method is used to convert the nonlinear mapping of the MLP into equivalent linear constraints, building a mixed-integer programming optimization model. The date coupling coefficient is then substituted to solve for the optimal vehicle distribution and dispatch instructions are issued. Simultaneously, the network is iteratively updated by expanding the sample based on the error between predicted and actual profits. This invention adapts the coupling coefficient to the travel characteristics of transitional dates, eliminating the prediction jump error caused by binary partitioning, and ultimately achieving high-precision profit prediction for all scenarios, including weekdays, holidays, and transitional semi-holidays.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and scheduling technology for autonomous ride-hailing services, specifically to a method for predicting the operating profit of autonomous ride-hailing services, a method for optimizing the initial vehicle distribution of autonomous ride-hailing services, and an operation and scheduling system for autonomous ride-hailing services. Background Technology

[0002] As autonomous driving technology continues to integrate with the ride-sharing industry, autonomous ride-hailing vehicles are gradually being deployed on a large scale and have become an important component of urban smart transportation construction. For fleet operation and management, a crucial step in rationally planning the initial deployment distribution of vehicles before they go online is obtaining realistic operating profit forecasts.

[0003] Early profit forecasting methods commonly used in the industry often adopted a unified overall forecasting approach, without breaking down the operational space or time period, or distinguishing the differences in travel patterns between weekdays and holidays. While these methods were simple to operate and had low computational costs, enabling quick and rough estimates of overall profits, their standardized calculation logic made it difficult to accurately depict the characteristics of passenger flow changes across different times and dates, resulting in limitations in the precision of the forecasts.

[0004] To improve forecasting accuracy, forecasting schemes based on region, time period, and date type are gradually gaining widespread application. These methods break down the city's operational area into multiple sub-regions, divide peak hours into sub-time periods, and then use calendar labels for statutory holidays and weekdays to model samples separately. Finally, they integrate the profits from each module, mitigating the overall forecasting accuracy to some extent. However, most of these schemes rely on binary calendar information for classification, simply categorizing dates into weekdays and holidays. For "semi-holidays"—such as adjusted workdays, the day before a short holiday, days of major city events, and post-holiday recovery days—calendar labels struggle to accurately reflect unique travel patterns. Forcibly classifying them using a binary standard can cause significant fluctuations in forecast values. During profit aggregation and sample learning, the model may interpret these fluctuations as abnormal interference, reducing the learning weight of such sample features and negatively impacting forecasting accuracy for semi-holiday scenarios.

[0005] Overall, the two existing prediction approaches each have their own applicable scenarios, but the former lacks sufficient precision, while the latter is difficult to adapt to special dates when travel characteristics are in a transitional state. There is still room for improvement in terms of consistently outputting high-precision profit prediction results across all scenarios. Summary of the Invention

[0006] To address the technical problem of slightly lower accuracy in predicting overall profits during special dates when travel patterns are in transitional phases, this invention provides a method for predicting the operating profits of autonomous ride-hailing vehicles. Based on this profit prediction method, this invention also provides a method for optimizing the initial vehicle distribution of autonomous ride-hailing vehicles and an autonomous ride-hailing operation scheduling system.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting the operating profit of autonomous ride-hailing vehicles includes the following steps: Multiple multidimensional demand distribution vectors are collected from historical operational data. These vectors include the corresponding operational region, operational time period, total order demand for the region and time period, and the initial vehicle distribution status for the region and time period. From all multidimensional demand distribution vectors, select the multidimensional demand distribution vectors corresponding to all holidays, and use these to construct the prototype distribution vector of holiday demand. n H This represents the total number of multidimensional demand distribution vectors during holidays. Let j be the normalized multidimensional demand distribution vector for the j-th holiday. Obtain the normalized multidimensional demand distribution vector of the target region within the target time period, including weekdays, public holidays, and transitional semi-holidays in between. ,calculate and The cosine similarity is used to obtain the date coupling coefficient. , where ε is a positive smoothing term; The L2 norm is used, and the superscript T indicates the vector transpose operation. Obtain the basic profit forecast value B and the working day profit forecast value R for the target area within the target time period. N and holiday profit forecast R H ; According to α on R N and R H The weighted average is then overlaid with B to obtain the overall operating profit forecast for the target region within the target time period. .

[0008] As a further improvement to the above scheme: obtaining B, R N and R H The specific steps are as follows: Build a three-branch MLP forecasting model that includes a public backbone network, a basic profit forecasting branch, a weekday profit forecasting branch, and a holiday profit forecasting branch; Will Input the public backbone network, extract general travel features through the public backbone network, and output shared implicit features; The shared implicit features are input into the basic profit forecasting branch, the weekday profit forecasting branch, and the holiday profit forecasting branch, respectively. The basic profit forecast branch calculates and outputs a basic profit forecast value B that is not date-sensitive. The weekday profit forecast branch calculates and outputs a weekday profit forecast value R that adapts to weekday travel characteristics. N ; The holiday profit forecast branch calculates and outputs a holiday profit forecast value R that adapts to holiday travel characteristics. H .

[0009] As a further improvement to the above scheme, the multidimensional demand distribution vector is generated as follows: For the operating area, a fixed number of operating sub-areas and a fixed number of operating sub-time periods are pre-defined, and a unified area-time period sorting rule is set. For daily operational data, the total order demand and initial vehicle distribution status for each sub-region and sub-time period are calculated sequentially according to the sorting rules. By sequentially concatenating the total order demand and initial vehicle distribution status of all regional sub-time periods, the original multidimensional demand distribution vector of the corresponding regional time period is obtained. The original multidimensional demand distribution vector is normalized according to its dimensions to obtain the normalized multidimensional demand distribution vector.

[0010] As a further improvement to the above scheme: the holiday demand prototype distribution vector adopts a rolling update mechanism: a fixed update cycle is preset, and after the end of each update cycle, the newly added holiday operation data within the update cycle is collected to generate the corresponding normalized holiday multidimensional demand distribution vector; the newly added holiday multidimensional demand distribution vector is merged with the previous holiday multidimensional demand distribution vector, and the updated holiday demand prototype distribution vector is recalculated according to the holiday demand prototype distribution vector calculation formula, replacing the original holiday demand prototype distribution vector, which is used for the subsequent calculation of the date coupling coefficient.

[0011] A method for optimizing the initial vehicle distribution of autonomous ride-hailing vehicles includes the following optimization steps: Obtain the normalized multidimensional demand distribution vector of the target operating area within the target operating period, and calculate the corresponding date coupling coefficient α using the autonomous driving ride-hailing operation profit prediction method. The three-branch MLP prediction model used in the profit prediction method for autonomous ride-hailing operations is called, and its internal nonlinear profit prediction relationship is transformed into equivalent linear constraints. The equivalent linear constraints are transformed into: determining all linear operation layers and ReLU nonlinear activation layers in the common backbone network, basic profit prediction branch, weekday profit prediction branch, and holiday profit prediction branch within the three-branch MLP prediction model; determining the upper and lower bounds of the pre-activation values ​​layer by layer based on the input range of the multidimensional demand distribution vector, the value range of network weights and bias parameters; introducing binary 0-1 auxiliary variables and using the Big-M method to convert the ReLU functions of each layer into equivalent linear constraint groups; replacing all ReLU nonlinear operations with linear constraint groups; and transforming the nonlinear profit prediction operation logic of the three-branch MLP prediction model into linear equality constraints and linear inequality constraints, thus completing the equivalent linear transformation of the overall nonlinear profit prediction relationship. Based on the profit forecast relationship after linear transformation, a mixed integer programming optimization model is built in combination with operational constraints; The pre-calculated date coupling coefficient α is substituted into the mixed integer programming optimization model as a fixed constant, and different initial vehicle distribution schemes are input for solution to obtain the optimal initial vehicle distribution scheme. Based on the optimal initial vehicle distribution scheme, vehicle pre-scheduling instructions are generated and sent to the autonomous vehicle terminal.

[0012] As a further improvement to the above scheme, the linear constraint group is specifically represented as follows: ; Among them, h r z is the activation output value of the r-th layer network; r Let be the pre-activation value of the r-th layer network; The absolute value of the negative lower bound constant. δ is the absolute value of the positive upper bound constant. r Let be a binary auxiliary variable of the r-th layer network.

[0013] As a further improvement to the above scheme, the process of obtaining the optimal initial vehicle distribution scheme is as follows: Let v be the initial number of vehicles deployed and their distribution locations in each sub-region during each operating sub-period. t,i , where i=1,2,...,I, I is the total number of operating sub-regions; t=1,2,...,N, N is the total number of operating sub-time periods; The pre-calculated date coupling coefficient α is used as a fixed constant and substituted into the linearized profit objective function expression: max represents the maximization operation; Simultaneously, a constraint on the conservation of the total number of vehicles in the fleet is introduced: ; Among them, V total This refers to the total number of vehicles that the fleet can deploy within its operating area. Introduce a non-negative constraint on the number of vehicles deployed in each sub-area within each operating sub-time period: ; By simultaneously establishing the profit objective function and the two constraints, a complete mixed-integer programming optimization model is constructed. This mixed-integer programming optimization model is then solved using a solver to obtain the result that... The largest group of v t,i This is the optimal initial vehicle distribution scheme corresponding to the target operating period.

[0014] As a further improvement to the above scheme, the training steps of the three-branch MLP prediction model are as follows: Using the normalized multidimensional demand distribution vector corresponding to historical operational data as training samples, calculate the date coupling coefficient α corresponding to each training sample. The training samples are fed into the public backbone network to extract shared latent features, which are then input into the basic profit prediction branch, weekday profit prediction branch, and holiday profit prediction branch to obtain the corresponding comprehensive operating profit prediction values. The overall loss function is calculated based on the comprehensive operating profit forecast and the actual operating profit, and all network parameters of the public backbone network are updated synchronously in reverse. For the weekday profit prediction branch and the holiday profit prediction branch, the loss gradient is amortized with weight 1-α and weight α respectively to complete the iterative update of the network parameters of the two prediction branches.

[0015] As a further improvement to the above scheme: the optimal initial vehicle distribution scheme obtained by solving the problem and the corresponding multidimensional demand distribution vector are input into the real operation simulator, and the actual operating profit value is output. ; Calculate the corresponding ones and The relative error value between them, when the relative error value exceeds the preset threshold, will affect the multidimensional demand distribution vector, the optimal initial vehicle distribution scheme, and As new training samples, they are added to the training dataset; The three-branch MLP prediction model is retrained using the updated training dataset. The relative error value is calculated iteratively until the relative error value of a single sample is no greater than the preset threshold, or the relative error values ​​of multiple consecutive rounds of sampling meet the threshold requirement.

[0016] An autonomous ride-hailing operation and dispatch system includes a data processing module, a profit prediction module, and a distribution optimization module; The data processing module is used to generate a normalized multidimensional demand distribution vector, construct a prototype distribution of holiday demand, and calculate the date coupling coefficient. The profit forecasting module executes the autonomous ride-hailing operation profit forecasting method and outputs the comprehensive operation profit forecast value. The distribution optimization module executes the initial vehicle distribution optimization method for autonomous ride-hailing vehicles, solves for the optimal initial vehicle distribution, and generates pre-schedule instructions.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. In the autonomous driving ride-hailing operation profit prediction method of the present invention, a multi-dimensional demand distribution vector including operating area, operating time period, order demand, and vehicle distribution status is first constructed. The travel characteristics of the date are quantified in the form of high-dimensional spatiotemporal data, eliminating the dependence on calendar labels and more accurately identifying the transition state. Then, by taking the average of all holiday samples, a holiday demand prototype distribution vector is constructed to form a standardized holiday travel characteristic reference system. Cosine similarity calculation is introduced to convert the similarity between the target time period demand vector and the prototype vector into a date coupling coefficient with continuous values ​​in the range of 0 to 1. This effectively quantifies the degree of transition bias of travel characteristics, replacing the traditional binary hard division of either 0 or 1. Finally, a weighted fusion formula is adopted, with the basic profit prediction value as a general benchmark. The contribution weight of the weekday profit prediction value and the holiday profit prediction value is dynamically allocated through the date coupling coefficient. For transitional semi-holidays, the prediction results of the two scenarios will be smoothly superimposed according to the similarity ratio, rather than forcibly using a single model output. This not only preserves the prediction adaptability of the two scenarios, but also eliminates the prediction jump error caused by binary division. Finally, high-precision profit prediction for all scenarios of weekdays, holidays, and transitional semi-holidays is achieved.

[0018] 2. By adopting a periodic rolling update mechanism, it can continuously absorb the newly added holiday operation data in each update cycle. After merging the newly added holiday multi-dimensional demand distribution vector with all historical holiday samples, it recalculates the holiday demand prototype distribution vector and replaces the old prototype. It can dynamically adapt to the long-term changes in urban holiday travel demand with seasons, urban activities, and residents' travel habits. It avoids the benchmark shift problem caused by the lag of samples and the iteration of travel patterns in a single static prototype. It ensures that the date coupling coefficients calculated in different periods can truly and accurately reflect the matching degree between the travel distribution on the target date and the current holiday travel pattern. It continuously and stably improves the long-term reliability and accuracy of the operating profit prediction results under weekdays, holidays, and various transitional semi-holidays. At the same time, it eliminates the need for manual sample expansion and prototype recalculation, reducing manual maintenance costs and improving the automation and adaptability of the entire prediction method.

[0019] 3. In the initial vehicle distribution optimization method for autonomous ride-hailing vehicles, the date coupling coefficient α, which can quantify and distinguish the travel characteristics of various dates, is first obtained based on the autonomous ride-hailing operation profit prediction method. Then, the equivalent linear constraint transformation is completed by the Big-M method for the nonlinear operation inside the three-branch MLP. The neural network prediction relationship that could not be directly involved in the planning solution is transformed into a linear constraint. Based on this, a mixed integer programming optimization model is built. The pre-calculated α is substituted into the model as a fixed constant to solve the optimal vehicle deployment scheme. Finally, the scheduling instructions are automatically generated and issued to autonomous vehicles. This opens up the complete link from travel demand prediction to vehicle resource optimization and allocation, effectively solving the defects of traditional optimization models that are difficult to integrate refined time-segmented and scenario-segmented profit prediction results and can only rely on simple profit function modeling, resulting in low profit of the scheduling scheme.

[0020] 4. A closed-loop calibration process for the simulator is set up to feed the obtained optimal initial vehicle distribution scheme into the operation simulator to obtain the real profit. The relative error is calculated by comparing it with the predicted profit. When the error exceeds the standard, the training set is updated by automatically supplementing the demand, vehicle distribution, and real profit samples of that group and retraining the three-branch MLP. The process is iterated until the error reaches the standard. This can continuously correct the prediction bias caused by the lack of optimal scheduling samples in the neural network, and continuously reduce the deviation between the profit prediction and the actual operating profit. This allows the linearized optimized model benchmark to continuously fit the real operating scenario, taking into account both the theoretical optimality of vehicle scheduling optimization and the long-term practical accuracy. This significantly improves the profit improvement effect and the adaptive iteration capability of the model in the pre-scheduling of autonomous ride-hailing vehicles. Attached Figure Description

[0021] Figure 1 This is a flowchart of the initial vehicle distribution optimization method.

[0022] Figure 2 This is a flowchart illustrating the linear transformation from a ReLU multilayer perceptron based on the Big-M method to a mixed-integer programming model.

[0023] Figure 3 A flowchart of the closed-loop correction iteration process for optimizing the initial vehicle distribution of autonomous ride-hailing vehicles.

[0024] Figure 4 The figure shows the prediction error distribution curves of the two models in the example on all 21 samples.

[0025] Figure 5 The figure shows the absolute error distribution curves of the two models in the example on all 21 samples. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] The overall solution of this invention is divided into two core modules: prediction of operating profits for autonomous ride-hailing vehicles and optimization of the initial vehicle distribution for autonomous ride-hailing vehicles. For example... Figure 1 As shown, a standardized holiday demand prototype distribution vector is first constructed by collecting multi-dimensional demand distribution vectors containing operating areas, time periods, order volumes, and vehicle distribution. A rolling update mechanism is then used to continuously iterate this holiday demand prototype distribution vector. Based on cosine similarity, a date coupling coefficient with continuous values ​​is calculated. This coefficient is then used to weight and fuse the basic profit, weekday profit, and holiday profit output from the three-branch MLP, accurately adapting to the differentiated travel characteristics of weekdays, holidays, and transitional semi-holidays, thus solving the prediction jumps and insufficient accuracy problems caused by traditional binary date division. Next, based on the Big-M method, the ReLU nonlinear layer in the three-branch MLP is transformed into an equivalent linear constraint, converting the prediction logic into linear constraints to construct a mixed-integer programming optimization model. A fixed date coupling coefficient is substituted to solve for the optimal vehicle distribution in each region and time period, and dispatch instructions are issued. Simultaneously, an error correction closed loop from an operation simulator is used to continuously expand the sample and retrain the neural network, dynamically correcting prediction biases. This achieves full-process adaptive optimization from travel demand prediction to autonomous ride-hailing vehicle pre-positioning. The technical solution of this invention will be fully and meticulously described below with specific embodiments.

[0028] I. Profit Forecasting Methods for Autonomous Ride-Hailing Operations

[0029] This method for predicting the operating profit of autonomous ride-hailing services is used to accurately predict the overall operating profit of a target area within a target time period, targeting weekdays, holidays, and transitional semi-holiday scenarios. The specific details are as follows: (a) Generating a multidimensional demand distribution vector First, a standardized multidimensional demand distribution vector is constructed as the basic input data for subsequent prediction and optimization. The generation process is as follows: A fixed number of operational sub-regions and operational sub-time periods are pre-defined, and a unified region-time period sorting rule is set. In this embodiment, the target city's operational scope is divided into I operational sub-regions according to the traffic network grid, and the entire operational cycle is equally divided into N operational sub-time periods (for example, if the operational cycle is one week, and each day is equally divided into 12 operational sub-time periods, then the entire cycle has a total of 84 operational sub-time periods); the sorting rule adopts an ascending order of region number, and then an ascending order of time period number within the same region, ensuring that the meaning of all vector dimensions corresponds one-to-one.

[0030] Based on historical operational data within the operational cycle, the total order demand within each operational sub-region and operational sub-period is statistically analyzed according to the above sorting rules, along with the initial vehicle distribution status within each operational sub-region. The initial vehicle distribution status within each operational sub-period refers to the number and location distribution of autonomous ride-hailing vehicles in each operational sub-region before the start of the current operational sub-period.

[0031] By sequentially piecing together the order data and initial vehicle distribution status corresponding to all regional time periods in a fixed order, the original multidimensional demand distribution vector within the operating period is obtained.

[0032] In this embodiment, the regional time-segment order demand vector is denoted as:

[0033] Where, d i,t This represents the total order demand within the i-th operating sub-region and the t-th operating sub-period. The subscript i is the sub-region number, with a value range of i=1,2,...,I; the subscript t is the operating sub-period number, with a value range of t=1,2,...,N.

[0034] The initial vehicle distribution state vector is denoted as: ; Among them, v t,i This represents the initial vehicle distribution status of the ith operating sub-region within the t-th operating sub-period, i.e., the number and location distribution of autonomous ride-hailing vehicles.

[0035] By concatenating the regional time-segment order demand vector with the initial vehicle distribution state vector vertically, the complete original multidimensional demand distribution vector is obtained: ; In the formula, [D;V] represents the vertical concatenation operation of vectors.

[0036] The original multidimensional demand distribution vector is normalized by dimension, and the normalization formula is as follows: ; Where, x k This represents the value of the k-th dimension of the original multidimensional demand distribution vector. x is the normalized value of the k-th dimension of the original multidimensional demand distribution vector. k,min x k,max These are the minimum and maximum values ​​of the k-th dimension in the full historical operational data, respectively. After processing, a normalized multidimensional demand distribution vector is obtained, eliminating the interference of fluctuations in total passenger flow and total vehicle scale in different periods, and ensuring the effectiveness of subsequent similarity calculations.

[0037] (ii) Constructing the prototype distribution vector of holiday demand

[0038] From all historically normalized multidimensional demand distribution vectors, the multidimensional demand distribution vectors corresponding to all statutory holidays are selected. These are then used to construct a prototype distribution vector of holiday demand representing standard holiday travel patterns. The construction formula is as follows: ; Where, n H This represents the total number of multidimensional demand distribution vectors during holidays. Let be the normalized multidimensional demand distribution vector for the j-th holiday.

[0039] Each dimension of the holiday demand prototype distribution vector corresponds to the average relative demand level of that region during a standard holiday period.

[0040] Furthermore, in this embodiment, the holiday demand prototype distribution vector adopts a rolling update mechanism: a fixed update cycle (such as a calendar month) is preset, and after each update cycle, newly added holiday operation data within the update cycle is collected to generate a corresponding normalized holiday multidimensional demand distribution vector; the newly added holiday multidimensional demand distribution vector is merged with all historical holiday multidimensional demand distribution vectors, and the updated holiday demand prototype distribution vector is recalculated according to the above prototype construction formula, replacing the original holiday demand prototype distribution vector for subsequent date coupling coefficient calculation, ensuring that the prototype continuously adapts to the dynamic changes in urban travel patterns.

[0041] (iii) Calculate the date coupling coefficient

[0042] Currently, dates are typically binary-coded into weekdays and holidays based on calendar labels, and each is then assigned a separate prediction model to output profit results. However, in actual operational scenarios, the travel demand characteristics of transitional semi-holidays, such as adjusted workdays, days before and after short holidays, and days of large-scale urban events, differ from both regular workdays and standard statutory holidays. Forcibly classifying them into either category leads to abrupt deviations in prediction results and is easily identified as noise by the model, reducing prediction accuracy. To address this, this solution introduces a continuously valued date coupling coefficient. Cosine similarity is used to quantify the approximation of the target day's travel distribution pattern to the standard holiday prototype. Continuous weights are used to smoothly weight and fuse the prediction branches for workdays and holidays, adapting to the gradual changes in travel patterns during transitional days and fundamentally solving the problem of insufficient prediction accuracy for transitional days caused by binary hard-classification. For the target date to be predicted, a normalized multidimensional demand distribution vector for the target region within the target time period is obtained. Calculate its relationship with the holiday demand prototype distribution vector μ. H The cosine similarity is used to obtain the date coupling coefficient α, which is calculated using the following formula: ; Where the superscript T represents the vector transpose operation; The L2 norm is used; ε is a positive smoothing term, which is 10 in this embodiment. -8 This is used to avoid calculation errors caused by a denominator of zero.

[0043] The date coupling coefficient α ranges from 0 to 1. A value closer to 1 indicates a higher similarity between the travel pattern on that day and that of a standard holiday; a value closer to 0 indicates a higher similarity between the travel pattern on that day and that of a regular workday. For transitional semi-holidays such as adjusted workdays, days before and after short holidays, and days of large events, where travel patterns fall between workdays and holidays, the date coupling coefficient will exhibit a continuous value between 0 and 1, eliminating the need to rely on calendar labels to forcibly classify the target date into a binary system of workday or holiday.

[0044] The type can accurately quantify the degree of travel pattern bias during transitional semi-holidays, solving the prediction jump problem caused by binary hard division.

[0045] (iv) Three-branch MLP prediction model and profit output

[0046] This embodiment uses a three-branch MLP prediction model to output basic profit and scenario-specific profit prediction values. The three-branch MLP prediction model includes a common backbone network, a basic profit prediction branch, a weekday profit prediction branch, and a holiday profit prediction branch. The specific execution steps are as follows: The normalized multidimensional demand distribution vector of the target region within the target time period. Input a common backbone network; in this embodiment, the common backbone network consists of alternating layers of fully connected linear layers and ReLU activation layers, used to extract general travel latent features from the input vector and output a shared latent feature z with uniform dimension, which serves as the unified input for the subsequent three prediction branches.

[0047] The shared latent feature z is input into the basic profit prediction branch, the weekday profit prediction branch, and the holiday profit prediction branch, respectively. All three branches adopt a network structure of "multi-layer fully connected + ReLU activation + final output layer" and output the corresponding scalar prediction value, respectively.

[0048] The step-by-step calculation process for the weekday profit forecast branch is as follows: Branch Level 0 Input To share the latent feature z, that is: ; For the r-th layer (r=1,2,...,L) N L N (Total number of layers in the weekday profit forecasting branch), first perform a linear weighted transformation to obtain the pre-activation value of that layer: ; in, The pre-activation value of the r-th layer network in the weekday profit forecasting branch; The weight matrix of the r-th layer network for the weekday profit prediction branch; The activation output value of the (r-1)th layer network in the weekday profit prediction branch; Let be the bias vector of the r-th layer network in the weekday profit prediction branch.

[0049] After further processing using the ReLU activation function, the activation output of this layer is obtained: ; In the formula, σ(·) is the ReLU activation function. max(·,·) represents the maximum value operation, meaning that when the pre-activated value is greater than 0, the original value is output; when it is less than or equal to 0, 0 is output. Finally, after passing through the Lth branch of the working day profit prediction... N The linear transformation of the output layer of the layer network yields the predicted profit value for the workday: ; Among them, w N b is the weight vector for the output layer of the weekday profit prediction branch. N R is the bias scalar for the output layer of the weekday profit forecast branch. N (z) refers to the weekday profit forecast R, which is adapted to weekday travel characteristics. N .

[0050] The profit forecast branch for holidays is structurally symmetrical with the profit forecast branch for weekdays, and the calculation process layer by layer is as follows: Branch Level 0 Input To share the latent feature z, that is: ; For the r-th layer (r=1,2,...,L) H L H (Total number of network layers for the holiday profit prediction branch), pre-activation value and activation output are as follows: ; ; In the formula, The pre-activation value of the r-th layer network for the holiday profit prediction branch; The weight matrix of the r-th layer network for the holiday profit prediction branch; The activation output value of the (r-1)th layer network of the holiday profit prediction branch; This is the bias vector of the r-th layer network in the holiday profit prediction branch.

[0051] Finally, after linear transformation of the output layer, the predicted profit value for the holidays is obtained: ; in, The weight vector for the output layer of the holiday profit prediction branch. R is the bias scalar of the output layer for the holiday profit forecasting branch. H (z) refers to the holiday profit forecast value R, which is adapted to the travel characteristics of holidays. H .

[0052] The basic profit prediction branch and the two profit branches mentioned above adopt the same multi-layer fully connected network structure, with the input of the 0th layer of the branch... To share the latent feature z, that is: ; For the r-th layer (r=1,2,...,L) B L B (Total number of layers in the basic profit prediction branch), the pre-activation value and the activation output are respectively: ; ; In the formula, The pre-activation value of the r-th layer network in the basic profit prediction branch; The weight matrix of the r-th layer network in the basic profit prediction branch; The activation output value of the (r-1)th layer network of the basic profit prediction branch; It is the bias vector of the r-th layer network of the basic profit prediction branch.

[0053] Finally, after linear transformation of the output layer, the predicted profit value for the holidays is obtained: ; in, The weight vector of the output layer of the basic profit prediction branch. The bias scalar of the output layer of the basic profit prediction branch, R B (z) is the basic profit forecast value B.

[0054] Finally, the predicted profit value R for working days is calculated based on the date coupling coefficient α. N and holiday profit forecast R H The weighted average is then overlaid with the baseline profit forecast B to obtain the overall operating profit forecast for the target region within the target period. .

[0055] (V) Training Methods for Three-Branch MLP Prediction Models

[0056] The three-branch MLP prediction model is trained using a weighted gradient update method, and the specific steps are as follows: The normalized multidimensional demand distribution vector corresponding to the historical operation data is used as the training sample, and the date coupling coefficient α corresponding to each training sample is calculated according to the method in (III) above.

[0057] The training samples are fed into the public backbone network to extract shared latent features, which are then input into the basic profit prediction branch, weekday profit prediction branch, and holiday profit prediction branch to obtain the corresponding comprehensive operating profit prediction values.

[0058] Based on the comprehensive operating profit forecast and the actual operating profit, the overall loss function is calculated. In this embodiment, the mean squared error loss function is adopted. Backpropagation is performed based on the loss function to synchronously update all network parameters of the public backbone network. For the weekday profit forecast branch and the holiday profit forecast branch, the loss gradient is amortized with weight 1-α and weight α respectively to complete the iterative update of the network parameters of the two branches.

[0059] This training method allows samples from transitional semi-holiday periods to participate in the optimization of two profit branches simultaneously, avoiding branch overfitting caused by single-label samples and improving the prediction accuracy of transition dates.

[0060] II. Optimization Method for Initial Vehicle Distribution of Autonomous Ride-Hailing Vehicles

[0061] Based on the above profit prediction method, this embodiment uses mixed integer programming to solve for the optimal initial vehicle distribution and maximize operating profit. The specific steps are as follows: (I) Overall process optimization Obtain the normalized multidimensional demand distribution vector of the target scheduling region within the target scheduling period, and calculate the corresponding date coupling coefficient α using the above-mentioned autonomous ride-hailing operation profit prediction method.

[0062] The three-branch MLP prediction model is invoked, and its internal nonlinear profit prediction relationship is transformed into an equivalent linear constraint.

[0063] Based on the profit forecast relationship after linear transformation, a mixed integer programming optimization model is built in combination with operational constraints.

[0064] The pre-calculated date coupling coefficient α is substituted into the mixed integer programming optimization model as a fixed constant, and different initial vehicle distribution schemes are input to solve the problem, thus obtaining the optimal initial vehicle distribution scheme.

[0065] Based on the optimal initial vehicle distribution scheme, vehicle pre-schedule instructions are generated and sent to the autonomous vehicle terminal to guide vehicle parking and scheduling before operation in each time period.

[0066] (ii) Equivalent linear constraint transformation

[0067] like Figure 2 As shown, the Big-M method is used to perform an equivalent linear transformation on the ReLU nonlinear activation in the three-branch MLP prediction model, so that the entire prediction model can be embedded into a mixed integer programming model for solution. The specific steps are as follows: Identify all linear computation layers and ReLU nonlinear activation layers in the common backbone network, basic profit prediction branch, weekday profit prediction branch, and holiday profit prediction branch within the three-branch MLP prediction model; let the preactivation calculation formula for the r-th layer be: ; The ReLU activation output is: ; Among them, z r W is the pre-activation value of the r-th layer network. r b r These are the weight matrix and bias vector of the r-th layer network, respectively; h r-1 h r These are the activation output values ​​of the (r-1)th and rth layers of the network, respectively; max(·,·) represents the maximum value operation.

[0068] Based on the input range of the multidimensional demand distribution vector and the value range of network weights and bias parameters, the positive upper bound of the neuron preactivation value is determined layer by layer using the interval propagation method. absolute value of negative lower bound ,Right now .

[0069] Introducing a binary 0-1 auxiliary variable δ r The Big-M method is used to convert the ReLU function into the following set of linear constraints: ; in, It is the absolute value of the negative lower bound constant. δ is the absolute value of the positive upper bound constant. r is a binary auxiliary variable in the r-th layer of the network, used to identify the activation state of neurons.

[0070] When δ r When =1, constraint enforces h r =z r This corresponds to the ReLU positive input activation state; when δ r When =0, constraint enforces h r =0, which corresponds to the ReLU negative input truncation state.

[0071] Perform the above transformation on all ReLU activation layers in the network, and solve the basic profit forecast B and the working day profit forecast R of the three-branch MLP prediction model using nonlinear mapping. N Holiday profit forecast R H The operational logic is uniformly transformed into linear equality constraints and linear inequality constraints, completing the equivalent linear transformation of the overall nonlinear profit prediction relationship.

[0072] Perform the above equivalent linear transformation on all ReLU activation layers in the network, and solve the basic profit forecast B and the working day profit forecast R of the three-branch MLP prediction model using nonlinear mapping. N Holiday profit forecast R H The operational logic is uniformly transformed into linear equality constraints and linear inequality constraints, completing the equivalent linear transformation of the overall nonlinear profit prediction relationship.

[0073] Due to the basic profit forecast value B and the working day profit forecast value R N Holiday profit forecast R H All three are obtained by alternating operations of linear layers and ReLU activation layers. All the network structures corresponding to these three can be equivalently decomposed into a set of linear constraints that can be identified by mixed integer programming using the Big-M method described above. Therefore, the complete profit prediction relationship can be directly embedded into the mixed integer programming optimization model as a constraint support for the objective function.

[0074] (III) Constructing and solving the mixed-integer programming model

[0075] After completing the linear transformation, a mixed-integer programming optimization model is constructed to solve for the optimal initial vehicle distribution. The specific process is as follows: Let v be the initial vehicle distribution state of each operating sub-area within each operating sub-time period. i,t v i,t The object to be optimized in the model.

[0076] The pre-calculated date coupling coefficient α is used as a fixed constant and substituted into the linearized profit objective function expression, with the optimization objective being to maximize overall operating profit: ; Here, `max` represents the maximization operation. Since the date coupling coefficient α is predetermined based on the target periodic demand distribution before optimization and remains constant during the optimization process, the objective function is about v. i,t A linear function.

[0077] Introducing the constraint of conservation of total number of vehicles in the fleet: ; Among them, V total This refers to the total number of vehicles that the fleet can deploy within its operating area.

[0078] Introduce a non-negative constraint on the number of vehicles deployed in each sub-area within each operating sub-time period: ; Ensure that the number of vehicles deployed in each area and at each time period has physical significance.

[0079] By combining the profit objective function with the aforementioned constraints, a complete mixed-integer programming optimization model is constructed. This model is then solved using commercial solvers (such as Gurobi or CPLEX) to obtain the result that satisfies the objective function. The largest group of v i,t This refers to the optimal initial vehicle distribution scheme corresponding to the target scheduling area within the target scheduling period.

[0080] (iv) Closed loop of adversarial training correction

[0081] like Figure 3 As shown, to address the distribution shift issue caused by optimization and improve the model's prediction accuracy in optimal scheduling scenarios, this embodiment sets up an adversarial training correction process, with the specific steps as follows: The optimal initial vehicle distribution scheme obtained from the solution and the corresponding multidimensional demand distribution vector are input into the real operation simulator, and the actual operating profit value is output. The realistic operations simulator is built upon massive amounts of historical operational data and can simulate the actual order acceptance and profit generation process under different vehicle distributions and demands.

[0082] The overall operating profit forecast predicted by the calculation model Compared with actual operating profit relative error values ​​between When the relative error value exceeds the preset threshold (set to 5% in this embodiment), the multidimensional demand distribution vector, the optimal initial vehicle distribution scheme, and the actual operating profit value are added as new training samples to the training dataset.

[0083] The three-branch MLP prediction model is retrained using the updated training dataset. The above relative error calculation and sample supplementation process is repeated until the relative error of a single sample is no greater than the preset threshold, or the relative error of multiple consecutive samples meets the threshold requirement, thus completing the model correction.

[0084] III. Verification of Examples

[0085] To further illustrate the feasibility and predictive effectiveness of this invention, this example constructs a verification environment based on real urban travel data to fully verify the accuracy of the date coupling coefficient and the comprehensive operating profit prediction of the three-branch MLP prediction model.

[0086] Regarding data sources and preprocessing, the dispatch server uses the travel data of yellow taxis in the target area as historical demand data and selects 21 days of operational data from October 1st to October 21st, 2025 to construct a verification sample set. In the data preprocessing stage, order records with zero trip distance, trip time less than 1 minute or more than 3 hours, origin and destination located outside the boundary of the target area, negative cost or significantly abnormal cost are deleted, thereby obtaining valid order data for demand statistics and simulator calls.

[0087] Regarding the division of operating areas and time periods, the scheduling server divides the operating sub-regions based on the target area map: first, the target area range is extracted, then the geometric centroid of the postal code area is used as a spatial feature for K-Means clustering, and the same cluster areas are merged into continuous analysis units; in this embodiment, a total of 20 operating sub-regions are divided; considering that the impact of the initial vehicle distribution status of the area time period on the operating profit is mainly reflected in the first hour after the start of operation, the morning peak operating period is selected as the core analysis window, and the first hour after the start of operation is divided into 6 operating sub-time periods; the total order demand of the area time period within the morning peak period consists of the order volume of the 20 operating sub-regions in the above 6 operating sub-time periods, corresponding to a 120-dimensional area time period order demand vector.

[0088] Regarding fleet size and input feature construction, this embodiment constructs a fleet system containing 2,000 autonomous ride-hailing vehicles. The initial vehicle distribution state for a given time period represents the number and location of vehicles in 20 operating sub-regions at the start of operation, forming a 20-dimensional initial vehicle distribution state vector that satisfies the constraint of conservation of the total number of vehicles in the fleet. The scheduling server vertically concatenates the 120-dimensional order demand vector for a given time period with the 20-dimensional initial vehicle distribution state vector to form a 140-dimensional multi-dimensional demand distribution vector, which serves as the input feature for the three-branch MLP prediction model.

[0089] Training samples are generated using a real-world operations simulator. Specifically, the dispatch server randomly selects one day's data from preprocessed historical order data to form a corresponding regional time-segment order demand vector, and randomly generates an initial vehicle distribution state vector that satisfies the conservation of the total number of vehicles in the fleet. This set of regional time-segment order demand vectors and the initial vehicle distribution state vector are then input into the real-world operations simulator, which outputs the actual operating profit value under this combination. The real-world operations simulator can calculate operating profit based on order matching, vehicle empty runs, vehicle dispatching, service revenue, operating costs, and charging-related costs. In this embodiment, the rated power of the charging pile is 50kW, the average charging demand after the fleet's daily operation is 65kWh, the charging price is 0.39USD / kWh, and the vehicle dispatching cost is 0.45USD / km.

[0090] This example selects 21 multi-dimensional demand distribution vector samples corresponding to the aforementioned 21 days for verification, including 15 weekday multi-dimensional demand distribution vector samples and 6 holiday multi-dimensional demand distribution vector samples. A holiday demand prototype distribution vector is constructed using the average of the 6 normalized holiday multi-dimensional demand distribution vectors. The date coupling coefficient corresponding to the target sample is obtained using the cosine similarity calculation method consistent with the previous method. The prediction accuracy is compared between the conventional single-output MLP prediction model and the three-branch MLP prediction model based on date coupling coefficient weighted fusion of this solution. The network structure of the conventional single-output MLP prediction model is the same as that of the three-branch MLP prediction model of this invention, except that the date coupling coefficient α of this invention is not used for weighting; instead, a conventional proportional weighted summation method is used to predict the overall operating profit.

[0091] A single-sample validation was performed by extracting the 21st holiday multidimensional demand distribution vector sample from all 21 samples. The date coupling coefficient α for this sample was calculated to be 0.8008, the corresponding weekday profit prediction branch weight was 1-α=0.1992, and the holiday profit prediction branch weight was α=0.8008. This calculation indicates that the travel pattern of this sample has a high degree of similarity to the holiday demand prototype; therefore, the holiday profit prediction branch contributes a higher percentage to the overall operating profit prediction process.

[0092] The single-sample prediction results show that the actual operating profit for this sample is $40,198.50, while the predicted value of the conventional single-output MLP prediction model is $38,577.41, with an absolute error of $1,621.09. The predicted value of the three-branch MLP prediction model based on date coupling coefficient weighted fusion is $39,387.95, with an absolute error of $810.54, which is about 50.0% lower than the prediction error of the conventional single-output MLP prediction model.

[0093] The comparison of the above prediction accuracy can be visually presented through an error comparison line graph, such as... Figure 4 and Figure 5 As shown. Figure 4 The distribution of prediction errors for the two models across all 21 samples is shown. The vertical axis represents the difference between the predicted operating profit and the actual operating profit, and the black horizontal line represents the zero-error baseline. Figure 5 The absolute error distribution of the corresponding samples is displayed, with the vertical axis representing the absolute value of the prediction error; the horizontal axis represents the sample number, where samples 1 to 15 are weekday multi-dimensional demand distribution vector samples, and samples 16 to 21 are holiday multi-dimensional demand distribution vector samples; the orange line corresponds to the conventional single-output MLP prediction model, and the blue line corresponds to the three-branch MLP prediction model of this solution based on date coupling coefficient weighted fusion. Figure 4 and Figure 5As can be intuitively observed, the prediction error curve of this scheme is closer to the zero-error baseline, and the error fluctuation amplitude is significantly smaller than that of the conventional single-output MLP prediction model. In terms of absolute error, the absolute error values ​​of each sample in this scheme are lower than those of the conventional single-output MLP prediction model, and it shows a stable accuracy improvement effect in both weekday and holiday sample intervals.

[0094] Overall statistics on all 21 samples show that the average absolute error of the conventional single-output MLP prediction model is $1,949.73, while the average absolute error of the three-branch MLP prediction model based on date coupling coefficient weighted fusion is $974.86, representing a reduction of approximately 50.0% in average prediction error.

[0095] The above analysis shows that by introducing the date coupling coefficient, the three-branch MLP prediction model can dynamically adjust the output weights of the weekday profit prediction branch and the holiday profit prediction branch according to the similarity between the target sample and the holiday demand prototype. This reduces the prediction jump deviation caused by hard division and hard switching between weekday, holiday and transitional semi-holiday samples, making the comprehensive operating profit prediction result closer to the actual operating profit level.

[0096] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the operating profit of autonomous ride-hailing vehicles, characterized in that, Includes the following steps: Multiple multidimensional demand distribution vectors are collected from historical operational data. These vectors include the corresponding operational region, operational time period, total order demand for the region and time period, and the initial vehicle distribution status for the region and time period. From all multidimensional demand distribution vectors, select the multidimensional demand distribution vectors corresponding to all holidays, and use these to construct the prototype distribution vector of holiday demand. n H This represents the total number of multidimensional demand distribution vectors during holidays. Let j be the normalized multidimensional demand distribution vector for the j-th holiday. Obtain the normalized multidimensional demand distribution vector of the target region within the target time period, including weekdays, public holidays, and transitional semi-holidays in between. ,calculate and The cosine similarity is used to obtain the date coupling coefficient. , where ε is a positive smoothing term; The L2 norm is used, and the superscript T indicates the vector transpose operation. Obtain the basic profit forecast value B and the working day profit forecast value R for the target area within the target time period. N and holiday profit forecast R H ; According to α on R N and R H The weighted average is then overlaid with B to obtain the overall operating profit forecast for the target region within the target time period. .

2. The method for predicting the operating profit of autonomous ride-hailing vehicles according to claim 1, characterized in that, Get B, R N and R H The specific steps are as follows: Build a three-branch MLP forecasting model that includes a public backbone network, a basic profit forecasting branch, a weekday profit forecasting branch, and a holiday profit forecasting branch; Will Input the public backbone network, extract general travel features through the public backbone network, and output shared implicit features; The shared implicit features are input into the basic profit forecasting branch, the weekday profit forecasting branch, and the holiday profit forecasting branch, respectively. The basic profit forecast branch calculates and outputs a basic profit forecast value B that is not date-sensitive. The weekday profit forecast branch calculates and outputs a weekday profit forecast value R that adapts to weekday travel characteristics. N ; The holiday profit forecast branch calculates and outputs a holiday profit forecast value R that adapts to holiday travel characteristics. H .

3. The method for predicting the operating profit of autonomous ride-hailing vehicles according to claim 2, characterized in that, The multidimensional demand distribution vector is generated as follows: For the operating area, a fixed number of operating sub-areas and a fixed number of operating sub-time periods are pre-defined, and a unified area-time period sorting rule is set. For daily operational data, the total order demand and initial vehicle distribution status for each sub-region and sub-time period are calculated sequentially according to the sorting rules. By sequentially concatenating the total order demand and initial vehicle distribution status of all regional sub-time periods, the original multidimensional demand distribution vector of the corresponding regional time period is obtained. The original multidimensional demand distribution vector is normalized according to its dimensions to obtain the normalized multidimensional demand distribution vector.

4. The method for predicting the operating profit of autonomous ride-hailing vehicles according to claim 1, characterized in that, The holiday demand prototype distribution vector adopts a rolling update mechanism: a fixed update cycle is preset, and after each update cycle, newly added holiday operation data within the update cycle is collected to generate a corresponding normalized holiday multidimensional demand distribution vector; the newly added holiday multidimensional demand distribution vector is merged with the previous holiday multidimensional demand distribution vector, and the updated holiday demand prototype distribution vector is recalculated according to the holiday demand prototype distribution vector calculation formula, replacing the original holiday demand prototype distribution vector for subsequent date coupling coefficient calculation.

5. A method for optimizing the initial vehicle distribution of autonomous ride-hailing vehicles, characterized in that, The optimization steps include the following: Obtain the normalized multidimensional demand distribution vector of the target operating area within the target operating period, and calculate the corresponding date coupling coefficient α using the autonomous ride-hailing operation profit prediction method described in claim 1. The three-branch MLP prediction model used in the autonomous ride-hailing operation profit prediction method of claim 2 is invoked, and its internal nonlinear profit prediction relationship is transformed into an equivalent linear constraint. The equivalent linear constraint is transformed into: determining all linear operation layers and ReLU nonlinear activation layers in the common backbone network, basic profit prediction branch, weekday profit prediction branch, and holiday profit prediction branch within the three-branch MLP prediction model; determining the upper and lower bounds of the pre-activation values ​​layer by layer according to the input range of the multidimensional demand distribution vector, the value range of network weights and bias parameters; introducing binary 0-1 auxiliary variables and using the Big-M method to convert the ReLU functions of each layer into equivalent linear constraint groups; replacing all ReLU nonlinear operations with linear constraint groups; and transforming the nonlinear profit prediction operation logic of the three-branch MLP prediction model into linear equality constraints and linear inequality constraints, thus completing the equivalent linear transformation of the overall nonlinear profit prediction relationship. Based on the profit forecast relationship after linear transformation, a mixed integer programming optimization model is built in combination with operational constraints; The pre-calculated date coupling coefficient α is substituted into the mixed integer programming optimization model as a fixed constant, and different initial vehicle distribution schemes are input for solution to obtain the optimal initial vehicle distribution scheme. Based on the optimal initial vehicle distribution scheme, vehicle pre-scheduling instructions are generated and sent to the autonomous vehicle terminal.

6. The method for optimizing the initial vehicle distribution of autonomous ride-hailing vehicles according to claim 5, characterized in that, The linear constraint group is specifically represented as follows: ； Among them, h r z is the activation output value of the r-th layer network; r Let be the pre-activation value of the r-th layer network; The absolute value of the negative lower bound constant. δ is the absolute value of the positive upper bound constant. r Let be a binary auxiliary variable of the r-th layer network.

7. The method for optimizing the initial vehicle distribution of autonomous ride-hailing vehicles according to claim 5, characterized in that, The process of obtaining the optimal initial vehicle distribution scheme is as follows: Let v be the initial number of vehicles deployed and their distribution locations in each sub-region during each operating sub-period. t,i , where i=1,2,...,I, I is the total number of operating sub-regions; t=1,2,...,N, N is the total number of operating sub-time periods; The pre-calculated date coupling coefficient α is used as a fixed constant and substituted into the linearized profit objective function expression: max represents the maximization operation; Simultaneously, a constraint on the conservation of the total number of vehicles in the fleet is introduced: ； Among them, V total This refers to the total number of vehicles that the fleet can deploy within its operating area. Introduce a non-negative constraint on the number of vehicles deployed in each sub-area within each operating sub-time period: ； By simultaneously establishing the profit objective function and the two constraints, a complete mixed-integer programming optimization model is constructed. This mixed-integer programming optimization model is then solved using a solver to obtain the result that... The largest group of v t,i This is the optimal initial vehicle distribution scheme corresponding to the target operating period.

8. The method for optimizing the initial vehicle distribution of autonomous ride-hailing vehicles according to claim 5, characterized in that, The specific training steps for the three-branch MLP prediction model are as follows: Using the normalized multidimensional demand distribution vector corresponding to historical operational data as training samples, calculate the date coupling coefficient α corresponding to each training sample. The training samples are fed into the public backbone network to extract shared latent features, which are then input into the basic profit prediction branch, weekday profit prediction branch, and holiday profit prediction branch to obtain the corresponding comprehensive operating profit prediction values. The overall loss function is calculated based on the comprehensive operating profit forecast and the actual operating profit, and all network parameters of the public backbone network are updated synchronously in reverse. For the weekday profit prediction branch and the holiday profit prediction branch, the loss gradient is amortized with weight 1-α and weight α respectively to complete the iterative update of the network parameters of the two prediction branches.

9. The method for optimizing the initial vehicle distribution of autonomous ride-hailing vehicles according to claim 8, characterized in that, The optimal initial vehicle distribution scheme obtained from the solution and the corresponding multidimensional demand distribution vector are input into the real operation simulator, and the actual operating profit value is output. ; Calculate the corresponding ones and The relative error value between them, when the relative error value exceeds the preset threshold, will affect the multidimensional demand distribution vector, the optimal initial vehicle distribution scheme, and As new training samples, they are added to the training dataset; The three-branch MLP prediction model is retrained using the updated training dataset. The relative error value is calculated iteratively until the relative error value of a single sample is no greater than the preset threshold, or the relative error values ​​of multiple consecutive rounds of sampling meet the threshold requirement.

10. An autonomous driving ride-hailing operation and dispatch system, characterized in that, It includes a data processing module, a profit forecasting module, and a distribution optimization module; The data processing module is used to generate a normalized multidimensional demand distribution vector, construct a prototype distribution of holiday demand, and calculate the date coupling coefficient. The profit forecasting module executes the autonomous ride-hailing operation profit forecasting method according to any one of claims 1-4 and outputs a comprehensive operation profit forecast value. The distribution optimization module executes the initial vehicle distribution optimization method for autonomous ride-hailing vehicles as described in any one of claims 5-9, solves for the optimal initial vehicle distribution, and generates pre-schedule instructions.

Images