Shared electric vehicle dispatching method and system based on big data analysis

Abstract

This invention discloses a shared electric vehicle scheduling method and system based on big data analysis, relating to the field of resource scheduling technology. The method includes obtaining raw data records of deployment points from a shared electric vehicle operation platform for validity verification, forming a time-series sample set of multiple deployment points; based on the time-series sample set of multiple deployment points, calculating the actual road distance, two-way vehicle flow intensity, and Pearson correlation coefficient of rental demand changes between any two different deployment points, constructing a deployment point association graph, performing spatial aggregation, feature transformation, and prediction to obtain the predicted number of rented vehicles and the predicted number of returned vehicles; based on the predicted number of rented vehicles and the predicted number of returned vehicles, calculating the safe inventory and expected number of vehicles at each deployment point, determining the deployment point's pending addition, candidate removal, and supply-demand balance scheduling status, and then calculating the unit service value and marginal service loss increment coefficient of the deployment point; significantly improving the effectiveness of supply-demand matching and overall scheduling efficiency.

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology, and in particular to a shared electric vehicle scheduling method and system based on big data analysis. Background Technology

[0002] With the rapid popularization of shared mobility, shared electric bikes have become an important part of urban short-distance travel. Shared electric bike operation platforms deploy fixed drop-off points in various urban areas to provide centralized parking, charging, and rental services. However, in actual operation, due to the significant uneven distribution of user travel demand in both time and space, the supply and demand status of vehicles at each drop-off point changes dynamically in real time. This often results in some drop-off points having excessive vehicle accumulation while others have no bikes available for rent. Therefore, how to accurately predict the short-term rental demand at each drop-off point based on real-time monitoring data and historical operational records, and formulate reasonable vehicle dispatch strategies accordingly, has become a core research topic in the field of shared electric bike operation and management.

[0003] However, existing technologies do not structurally deconstruct or systematically compensate for prediction errors, leading to a high risk of underestimating demand in scheduling decisions. Furthermore, they ignore the differences in service value of the same vehicle at different deployment points and in different spatiotemporal contexts, and fail to establish a mechanism for quantifying the opportunity cost after a vehicle is dispatched. This makes it impossible to dynamically compare and weigh vehicle retention value against scheduling and transportation costs during the scheduling plan generation process. In addition, the lack of comprehensive modeling of scheduling execution costs, route timeliness risks, and incentive compatibility makes it difficult to achieve economic optimization of total scheduling costs while ensuring supply and demand balance. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a shared electric vehicle scheduling method and system based on big data analysis, which solves the problems of existing technologies that fail to structurally deconstruct and systematically compensate for prediction errors, are unable to dynamically compare and weigh vehicle retention value and scheduling transportation costs during the scheduling scheme generation process, and are difficult to achieve economic optimization of total scheduling costs while ensuring supply and demand balance.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a shared electric vehicle scheduling method based on big data analysis, which includes, The original data records of the deployment points are obtained from the shared electric vehicle operation platform and their validity is verified to form a time-series sample set of multiple deployment points. Based on a time series sample set of multiple delivery points, after calculating the actual road distance, two-way vehicle flow intensity, and Pearson correlation coefficient of rental demand changes between any two different delivery points, a delivery point association graph is constructed to perform spatial aggregation, feature transformation, and prediction to obtain the predicted number of rented vehicles and the predicted number of returned vehicles. Based on the predicted number of rented vehicles and the predicted number of returned vehicles, the safety stock and expected number of vehicles at each deployment point are calculated. After determining the deployment point's status of pending transfer in, candidate transfer out, and supply-demand balance, the unit service value and marginal service loss increment coefficient of the deployment point are then calculated. Construct a set of comprehensive path costs and candidate outbound incentive values. Combine the unit service value and marginal service loss increment coefficient to obtain the minimum total scheduling cost of the outbound volume of the delivery point. Then, generate a scheduling task work order for execution and distribution.

[0007] As a preferred embodiment of the shared electric vehicle scheduling method based on big data analysis described in this invention, the following steps are taken: The original data records of deployment points obtained from the shared electric vehicle operation platform are validated to form a time-series sample set of multiple deployment points, as detailed below: The shared electric vehicle operation platform establishes a basic information table of deployment points and collects data according to a unified sampling period to obtain the original data records of each deployment point; Based on the original data records, records with empty placement point identifiers or collection times are removed. Then, the total number of vehicles available for rent in a given period is obtained by summing the number of available vehicles and the number of returned vehicles. After correction, the corrected data is sorted to generate a multi-placement point time-series sample set.

[0008] As a preferred embodiment of the shared electric vehicle scheduling method based on big data analysis described in this invention, the step of calculating the Pearson correlation coefficient of the actual road distance, two-way vehicle flow intensity, and rental demand changes between any two different deployment points based on a multi-deployment point time-series sample set is as follows: Call the road path calculation service to obtain the two-way road distance between any two delivery points; The intensity of two-way vehicle flow is obtained by statistically analyzing the number of two-way rentals and returns between each pair of shared electric vehicle order placement points throughout the history. After performing alignment operations on a time series sample set of multiple delivery points, the Pearson correlation coefficient between the changes in rental demand at each delivery point is calculated.

[0009] As a preferred embodiment of the shared electric vehicle scheduling method based on big data analysis described in this invention, the step of constructing a deployment point association graph involves spatial aggregation, feature transformation, and prediction to obtain the predicted number of rented vehicles and the predicted number of returned vehicles, as detailed below: The actual road distance, two-way vehicle flow intensity, and Pearson correlation coefficient are compared with the corresponding preset thresholds. If any two different delivery points meet at least one of the road distance condition, vehicle flow condition, or coefficient correlation condition, then an association edge is established between the current delivery points; otherwise, no association edge is established. The actual road distance, two-way vehicle flow intensity, and Pearson correlation coefficient are weighted and integrated to form the weights of the associated edges, and then an adjacency matrix is ​​constructed. The input feature vector of the delivery point is constructed and fed into the feature transformation layer to obtain the spatial correlation features of the delivery point; Based on spatial association features, a gated recurrent unit network is used to obtain the current hidden state, and the predicted number of rented vehicles and the predicted number of returned vehicles at the drop-off point are obtained through the Softplus function.

[0010] As a preferred embodiment of the shared electric vehicle scheduling method based on big data analysis described in this invention, the step of calculating the safe inventory and expected number of vehicles at each deployment point to determine the scheduling status of the deployment points (pending addition, candidate removal, and supply-demand balance) is as follows: The root mean square error between the actual rental volume and the predicted rental volume is used as the error compensation benchmark. Set the guarantee coefficient corresponding to the attribute matching of the current time period of the deployment point, combine the current error compensation benchmark and the cumulative predicted rental volume to calculate the number of vehicles required for guarantee, and combine the maximum vehicle capacity of the deployment point to calculate the current required safe inventory. Starting with the current number of available vehicles at the drop-off point, the predicted return volume is added and the predicted rental volume is subtracted. After non-negative constraint correction, the predicted number of vehicles after non-negative constraint correction is generated. The projected number of vehicles is compared with the safe inventory. Under the conditions that the projected number of vehicles is less than and greater than the safe inventory, the quantity to be transferred in and the maximum quantity to be transferred out are generated respectively. The scheduling status of the delivery point is generated based on the amount to be transferred in and the maximum amount that can be transferred out.

[0011] As a preferred embodiment of the shared electric vehicle scheduling method based on big data analysis described in this invention, the calculation of the unit service value and marginal service loss increment coefficient of the deployment point is as follows: Based on the scheduling status of the delivery points, establish a set of candidate delivery points to be transferred out and a set of delivery points to be transferred in; Extract the cumulative predicted rental volume and maximum vehicle capacity of candidate drop-off points, calculate the predicted demand intensity and impose an upper limit constraint, and then calculate the basic service value of a single vehicle remaining at the drop-off point by combining the preset minimum basic value and value conversion coefficient. Extract the error compensation benchmark and cumulative predicted rental volume, calculate the demand uncertainty and impose an upper limit constraint, and then multiply it by the prediction error value conversion coefficient to obtain the added service value. Sum the added service value with the basic service value to obtain the unit service value of a single vehicle remaining at the deployment point. The marginal service loss increment coefficient is determined based on the maximum available capacity and the safe holding capacity.

[0012] As a preferred embodiment of the shared electric vehicle scheduling method based on big data analysis described in this invention, before obtaining the feedback dispatch volume from the deployment point, the comprehensive path cost should be obtained first, specifically including: For any delivery point and any delivery point The road route calculation service obtains the shortest time path that dispatched vehicles can take and the estimated total dispatch time. Query delivery points The predicted sequence of steps where the number of vehicles first falls below the safe stock level within the scheduling time window is determined, and the latest arrival time is calculated. Based on the estimated total scheduling time, a combination of delivery points is selected to obtain the retained combination of delivery points; For the retained deployment point combination, the difference between the maximum scheduling duration and the expected total scheduling time will be used as the scheduling time margin, starting from the most recent Extract the actual scheduling time from historical scheduling tasks that pass through the same road area and are in the same time period, calculate the standard deviation of the scheduling time, and combine it with the scheduling time margin to calculate the delay risk. The shortest path time, the estimated total scheduling time, and the delay risk are weighted and integrated to generate a comprehensive path cost.

[0013] As a preferred embodiment of the shared electric vehicle dispatching method based on big data analysis described in this invention, the step of obtaining the feedback dispatch volume of the deployment points is specifically as follows: Pre-set the lower limit of incentives, the upper limit of incentives, and the fixed increment, and generate a set of candidate incentive values ​​to be removed according to the fixed increment; For any candidate outgoing delivery point, query the delivery points to be transferred in that have a time-feasible candidate scheduling path, form a set of delivery points to be transferred in, and use the number of delivery points to be transferred in as a weight, and combine it with the set of delivery points to be transferred in to perform a weighted average of the scheduling cost, so as to obtain the average generalized execution cost corresponding to the delivery point. The net incentive per unit that a released vehicle can obtain is obtained by subtracting the average generalized execution cost from the candidate release incentive value in the candidate release incentive value set, and then subtracting the unit service value. If the net incentive per unit is greater than zero, the feedback release amount is calculated based on the marginal service loss increment coefficient; otherwise, the feedback release amount is zero.

[0014] As a preferred embodiment of the shared electric vehicle scheduling method based on big data analysis described in this invention, the step of generating and issuing a scheduling task work order after performing minimum total scheduling cost modeling and solving specifically includes: With the goal of minimizing the total expenditure of scheduling incentives and covering the total demand for transfers, the smallest candidate transfer incentive value that makes the total feedback transfer amount not less than the total number of transfers is selected as the optimal transfer incentive. Substituting this value into the formula for the feedback transfer amount, the optimal feedback transfer amount under the optimal transfer incentive is obtained as the final number of vehicles to be transferred. The final number of vehicles dispatched is taken as the supply node capacity, and the number of vehicles to be dispatched to the destination points is taken as the demand node capacity. After using the comprehensive path cost as the edge weight, the minimum total scheduling cost model is performed and constraints are set. The minimum total scheduling cost model is solved by the operations research algorithm, and the specific vehicle allocation quantity and scheduling path from each candidate dispatch point to each destination point are output. Based on the specific number of vehicles allocated and the dispatch route, a dispatch task work order is generated, which includes the origin and destination points, the number of dispatched vehicles, the estimated arrival time, and the execution incentive amount. The dispatch task work order is then pushed to the handheld smart terminal of the dispatch staff through the Internet of Things communication network, and the staff executes the dispatch according to the dispatch task work order.

[0015] Secondly, this invention provides a shared electric vehicle dispatching system based on big data analysis, including: The sample generation module is used to obtain the original data records of the deployment points from the shared electric vehicle operation platform, perform validity verification, and form a time-series sample set of multiple deployment points. The return prediction module is composed of a multi-deployment point time series sample set. It calculates the actual road distance between any two different deployment points, the two-way vehicle flow intensity, and the Pearson correlation coefficient of rental demand changes. Then, it constructs a deployment point association map, performs spatial aggregation, feature transformation, and prediction to obtain the predicted number of rented vehicles and the predicted number of returned vehicles. The status determination and parameter calculation module is used to calculate the safe inventory and expected number of vehicles at each deployment point based on the predicted number of rented vehicles and the predicted number of returned vehicles. After determining the scheduling status of the deployment point, such as pending transfer, candidate transfer, and supply-demand balance, it then calculates the unit service value and marginal service loss increment coefficient of the deployment point. The scheduling task generation module is used to construct a set of comprehensive path costs and candidate dispatch incentive values. After combining the unit service value and the marginal service loss increment coefficient, it obtains the feedback dispatch volume of the delivery point, performs the minimum total scheduling cost modeling and solution, and generates a scheduling task work order for execution and distribution.

[0016] The beneficial effects of this invention are as follows: By constructing a multi-source information association graph that integrates actual road distance, two-way vehicle flow intensity, and rental demand changes, and using a spatiotemporal graph neural network for multi-step rental demand prediction, the prediction model can effectively capture the spatial dependence and temporal evolution patterns between deployment points. Furthermore, by extracting the root mean square of historical positive prediction errors as a compensation benchmark and combining it with a time-period attribute guarantee coefficient to establish an asymmetric error compensation mechanism, the negative impact of prediction deviations on scheduling decisions is effectively reduced. Secondly, by constructing a vehicle retention unit service value model based on predicted demand intensity and demand uncertainty, cumulative retention service loss, a set of candidate withdrawal incentive values, and solutions for optimal withdrawal incentives and minimum total scheduling costs, the invention achieves synergistic optimization of supply and demand matching, path reliability, and incentive economy. Therefore, this invention significantly improves the supply and demand matching effectiveness and comprehensive scheduling efficiency of the shared electric vehicle scheduling method. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the shared electric vehicle scheduling method based on big data analysis in Example 1.

[0019] Figure 2 This is a structural diagram of the shared electric vehicle dispatching system based on big data analysis in Example 1.

[0020] Figure 3 This is a flowchart of establishing associated edges in Example 1. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Example 1, referring to Figures 1-3 This is the first embodiment of the present invention, which provides a shared electric vehicle scheduling method based on big data analysis, including the following steps: S1. Obtain the original data records of the deployment points from the shared electric vehicle operation platform, perform validity verification and operational logic verification, and form a time-series sample set of multiple deployment points; S1.1 The shared electric vehicle operation platform establishes a basic information table of deployment points and collects data according to a unified sampling period to obtain the original data records of each deployment point.

[0025] Specifically, the shared electric vehicle operation platform establishes a basic information table for deployment points, assigns a unique and fixed deployment point identifier to each deployment point, and establishes a unique mapping relationship between the deployment point identifier and the deployment point's latitude and longitude, administrative region, deployment capacity, and the operating area to which it belongs. Data is collected according to a unified sampling period (e.g., 5 minutes) to obtain the original data records of each deployment point.

[0026] It should be noted that the data collected during the unified sampling period includes the number of currently available shared electric vehicles, the number of vehicles rented out in the current time period, the number of vehicles returned in the current time period, and the maximum vehicle capacity of the deployment point.

[0027] S1.2. Based on the original data records, remove records with empty placement point identifiers or collection times. Then, correct the total number of vehicles available for rent in a period based on the sum of the number of available vehicles and the number of returned vehicles. Sort the corrected data to generate a multi-placement point time series sample set.

[0028] Specifically, based on the original data records, it is checked whether each data record contains both the delivery point identifier and the collection time. If either of these items is empty, it is considered that the current data cannot correspond to the specific delivery point and time sequence, and the current original data record is directly deleted; otherwise, the current original data record is retained. Based on the retained original data records, the number of currently available shared electric vehicles and the number of vehicles returned within the current time period are extracted and added together to obtain the total number of vehicles that the deployment point can provide for rent during the collection period. Extract the number of vehicles rented out in the current time period from the retained original data records and compare it with the total number of vehicles that the deployment point can provide for rent during the collection period; If the number of vehicles rented out in the current time period is less than or equal to the total number of vehicles that the deployment point can provide for rent during the collection period, then the original data records are considered to be in line with the operational logic and the original data is retained. If the number of vehicles rented out in the current time period exceeds the total number of vehicles that the deployment point can provide for rent during the data collection period, the data is considered abnormal, and a correction method is used to adjust the number of rented vehicles to the theoretical maximum number of vehicles that can be rented.

[0029] The number of rented vehicles is adjusted to the theoretical maximum number of rentable vehicles using a correction method, expressed as: In the formula, Indicates the first Each collection point during the collection cycle The theoretical maximum number of vehicles that can be rented. This indicates the operation of finding the minimum value. Indicates the first Each collection point during the collection cycle The number of vehicles rented out, Indicates the first Each collection point during the collection cycle The total number of vehicles available for rent at that time; The corrected original data are arranged in chronological order, and the data at each sampling time are combined into a data group and sorted according to the delivery point identifier to form a multi-delivery point time series sample set.

[0030] It should be noted that for each delivery point with data records at the same sampling time, they are placed into the data group sequentially according to the fixed sorting rules of the delivery point identifiers. This ensures that data groups generated at different sampling times have a consistent delivery point arrangement order. For example, when the operating area includes three delivery points numbered first, second, and third, the data at each sampling time are arranged in the order of first, second, and third delivery points, and this order remains unchanged regardless of the order in which the data from each delivery point arrives at the platform.

[0031] S2. Based on a time series sample set of multiple delivery points, calculate the actual road distance, two-way vehicle flow intensity, and Pearson correlation coefficient of rental demand changes between any two different delivery points. Then, construct a delivery point association graph, perform spatial aggregation, feature transformation, and prediction to obtain the predicted number of rented vehicles and the predicted number of returned vehicles.

[0032] S2.1 Call the road path calculation service to obtain the bidirectional road distance between any two delivery points.

[0033] Specifically, the latitude and longitude coordinates of each delivery point are read from the multi-delivery-point time-series sample set, and any two different delivery points are... and The coordinates are sent as the start and end points to a road path calculation service (such as Gaode or Baidu Maps). The road path calculation service then searches for the shortest feasible path and delivery point based on the actual road network that allows dispatched vehicles to pass. to the drop-off point The actual road distance.

[0034] It should be noted that, because the actual driving direction of shared electric vehicles on the road may be affected by one-way streets, no-left-turn, no-U-turn, and traffic organization, the route from the origin to the destination and the route from the destination to the origin are considered two different dispatch directions, and road path calculations are performed separately, rather than using the same distance data directly. For example, when the operating area includes a first, second, and third drop-off point, multiple dispatch directions are established, such as from the first drop-off point to the second drop-off point, from the second drop-off point to the first drop-off point, from the first drop-off point to the third drop-off point, from the third drop-off point to the first drop-off point, from the second drop-off point to the third drop-off point, and from the third drop-off point to the second drop-off point. The corresponding road path information is queried independently for each dispatch direction.

[0035] S2.2. Based on the historical complete shared electric vehicle order statistics, the number of two-way vehicle rentals and returns between each pair of delivery points is used to obtain the two-way vehicle flow intensity.

[0036] Specifically, the system retrieves complete historical shared electric vehicle orders from the shared electric vehicle operation platform; each order includes at least the vehicle identifier, rental point identifier, rental time, return point identifier, and return time. Based on complete historical shared electric vehicle orders, rental and return locations are extracted and grouped for statistical analysis, resulting in a breakdown of rental and return locations by location. Rent out and at the drop-off point Number of times vehicles were returned; The intensity of two-way vehicle flow between drop-off points is obtained based on the number of times vehicles are returned.

[0037] The grouped statistics are expressed as follows: In the formula, Indicates by the first The first delivery point was rented out and in the [number]th [location]... The number of times vehicles are returned to each drop-off point. This represents the total number of orders within the historical period. Indicates the indicator function (when) The value is 1 if the condition is met, otherwise it is 0. Indicates the first The rental delivery point number corresponding to each order. This indicates that both conditions are true simultaneously. Indicates the first The return and delivery point number corresponding to each order; The expression for obtaining the bidirectional vehicle flow intensity between delivery points is: In the formula, Indicates the first The first delivery point and the first Two-way vehicle flow intensity between delivery points Indicates the period within a historical cycle from the first The first delivery point was rented out and in the [number]th [location]... The number of times vehicles were returned to each drop-off point.

[0038] It should be noted that data lacking rental / return point identification or order completion information, or where the order was not properly closed, will not be included in vehicle flow statistics. When the same shared electric scooter completes multiple valid orders within a historical period, each order will be counted separately. However, changes in vehicle location due to maintenance or manual dispatch will not be considered part of the rental / return flow and will not be included in the number of returned vehicles. .

[0039] S2.3 After performing alignment operations based on the time series sample set of multiple delivery points, calculate the Pearson correlation coefficient between the changes in rental demand at the delivery points.

[0040] Specifically, the number of rented vehicles at the same sampling time of each delivery point is extracted from the time series sample set of multiple delivery points, and an alignment operation is performed with the same time period of each day as the alignment unit to generate a rental demand sequence. Based on the rental demand sequence, the number of rental vehicles at adjacent sampling times is subtracted to obtain the change in rental demand at each delivery point. Then, the Pearson correlation coefficient between delivery points is calculated based on the change in rental demand.

[0041] The Pearson correlation coefficient between delivery points is calculated based on changes in rental demand, and its expression is as follows: In the formula, Indicates the first The first delivery point and the first The Pearson correlation coefficient between the delivery points Indicates the total number of sampling times. Indicates the first The first delivery point is at the... Changes in rental demand at each sampling time point Indicates the first The average change in rental demand for each location over a historical period. Indicates the first The first delivery point is at the... Changes in rental demand at each sampling time point Indicates the first The average change in rental demand for each location over a historical period; The specific calculation method for the change in average rental demand is as follows: It should be noted that the Pearson correlation coefficient is used to characterize the consistency of the rental demand trends between two rental locations, not to indicate whether the absolute number of vehicles rented out at the two locations is the same. If the increase and decrease in rental demand at two locations are basically synchronized within the historical statistical period, the corresponding Pearson correlation coefficient is relatively large, indicating that the two locations are affected by similar changes in travel demand, and their demand changes are highly synchronized. If the demand at one location increases while the demand at the other location usually decreases, the corresponding Pearson correlation coefficient is relatively small or even negative, indicating that there is a significant difference in the direction of demand changes between the two locations. If the Pearson correlation coefficient is close to zero, it indicates that there is no stable linear correspondence between the rental demand changes at the two locations.

[0042] S2.4. Compare the actual road distance, two-way vehicle flow intensity, and Pearson correlation coefficient with the corresponding preset thresholds. If any two different delivery points satisfy at least one of the road distance condition, vehicle flow condition, or correlation coefficient condition, then establish an association edge between the current delivery points; otherwise, do not establish one. The expression is: In the formula, Indicates the first The first delivery point and the first Whether there are related edges between the delivery points is indicated by a value of 1 (if any) and 0 (if any). Indicates the first From the first delivery point to the... The actual road distance to each delivery point This indicates the preset distance threshold. This indicates the preset vehicle flow threshold. This indicates a preset threshold related to changes in demand. Indicates or, This indicates otherwise.

[0043] It should be noted that the distance threshold is determined based on the actual road distance distribution between each delivery point within the operating area. In practice, the operating platform reads the actual road distance between any two different delivery points within the current operating area, excludes delivery point pairs with inaccessible roads, abnormal coordinates, or failed road path calculations, and arranges the remaining effective road distances in ascending order. The distance corresponding to a preset quantile position in the effective road distance distribution is selected as the candidate distance threshold. For example, the quantile distances within the top 20% to top 40% of all effective road distances are selected as candidate values, ensuring that the established road distance association edges primarily cover delivery points with relatively close geographical locations. Preferably, the actual road distance corresponding to the 30th quantile position is selected as the distance threshold. The vehicle flow threshold is determined based on the bidirectional vehicle flow intensity distribution of all deployment point pairs within the historical statistical period. In practice, the operations platform reads the bidirectional vehicle flow intensity between all different deployment points, excludes deployment point pairs that cannot be statistically analyzed due to missing order data, and arranges the remaining bidirectional vehicle flow intensity in ascending order. Since many deployment point pairs may not have vehicle flow, the operations platform preferentially excludes deployment point pairs with zero bidirectional vehicle flow intensity first, and then determines the vehicle flow threshold based on non-zero vehicle flow intensity. For example, the number of vehicles corresponding to the 60th to 80th positions of the non-zero vehicle flow intensity distribution is selected as a candidate vehicle flow threshold, with the 70th position being the preferred choice. The threshold for demand change is determined based on the Pearson correlation coefficient distribution of all valid delivery point pairs and the number of valid samples. The operations platform first excludes delivery point pairs with insufficient valid demand change samples, zero variance in the demand change sequence, or incalculable correlation coefficients. Threshold determination is only performed on delivery point pairs for which correlation coefficients can be calculated. For example, candidate thresholds are determined based on the distribution of valid Pearson correlation coefficients. Specifically, valid correlation coefficients greater than zero are arranged in ascending order, and the correlation coefficients corresponding to the 60th to 80th percentiles are selected as candidate values. The correlation coefficient corresponding to the 70th percentile is preferred as the threshold for demand change.

[0044] S2.5. After weighted fusion of actual road distance, two-way vehicle flow intensity, and Pearson correlation coefficient to form the weights of associated edges, an adjacency matrix is ​​constructed.

[0045] Specifically, based on the placement point pairs for establishing associated edges, the actual road distance, two-way vehicle flow intensity, and Pearson correlation coefficient are weighted and integrated to form the weight of the associated edges; For each delivery point, a self-association edge is set, and the weight of the associated edge is set to 1. For any two different delivery points that are not associated, the weight of their associated edge is set to 0. This is combined with the weight of the associated edge. Construct an adjacency matrix.

[0046] The actual road distance, two-way vehicle flow intensity, and Pearson correlation coefficient are weighted and fused together, and the expression is as follows: In the formula, Indicates the first The first delivery point and the first The weights of the edges connecting each delivery point The weights representing the actual road distances The weight representing the intensity of two-way vehicle flow. This indicates the weight of the Pearson correlation coefficient.

[0047] It should be noted that: three weights , , All are non-negative numbers, and the sum of the three fusion weights is set to one. Furthermore, the larger the fusion weight, the greater the influence of the corresponding type of association information on the final association edge weight; For example, in one implementation, the weight of road distance is set to 0.4, the weight of two-way vehicle flow intensity is set to 0.35, and the weight of Pearson correlation coefficient is set to 0.25. This weighting is based on the fact that the deployment point association of shared electric vehicles is primarily constrained by dispatch reachability distance, followed by historical user rental and return flows, with demand synchronization changes serving as a supplement. Therefore, the road distance indicator has the highest weight, the vehicle flow indicator has the second highest weight, and the demand-related indicators have relatively lower weights. For operation areas with densely packed delivery points and short dispatch distances, such as urban centers, campuses, or industrial parks, the distinguishing effect of road distance on the association of delivery points may decrease. In this case, the weight of road distance can be set to 0.3, the weight of vehicle flow intensity can be set to 0.4, and the weight of Pearson correlation coefficient can be set to 0.3.

[0048] S2.6 Construct the input feature vector of the delivery point and input it into the feature transformation layer to obtain the spatial correlation features of the delivery point.

[0049] Specifically, the adjacency matrix is ​​normalized according to the sum of the weights of the associated edges to obtain the normalized associated weights; The number of available vehicles and the number of rented and returned vehicles are extracted from the time series sample set of multiple deployment points, and then normalized using the minimum-maximum normalization method before being concatenated into the input feature vector. Based on the normalized association weights and the input feature vector, spatial aggregation is performed to obtain the spatial aggregation features of the delivery points. The spatial aggregation features are then input into the feature transformation layer for feature transformation to obtain the spatial association features of the delivery points.

[0050] The adjacency matrix is ​​normalized according to the sum of the weights of the associated edges, as expressed in the following expression: In the formula, Indicates the first The first delivery point and the first Normalized correlation weights between each delivery point Indicates the total number of collection points. This represents the index used to traverse all delivery points. Indicates the first The first delivery point and the first Weights of the associated edges between each delivery point; Based on the normalized association weights and the input feature vector, spatial aggregation is performed to obtain the spatial aggregation features of the delivery points. These spatial aggregation features are then input into a feature transformation layer for feature transformation, expressed as follows: In the formula, Indicates the first The first delivery point is at the... Spatial aggregation features at each sampling time, Indicates the first The first delivery point is at the... The input feature vector at each sampling time, Indicates the first The first delivery point is at the... Spatial correlation features at each sampling time, Represents a nonlinear activation function (specifically...) , Indicates when When greater than 0, output Otherwise, output 0). This represents the weight matrix of the feature transformation layer. This represents the bias vector of the feature transformation layer.

[0051] It should be noted that for delivery points with a normalized association weight of zero, they do not participate in spatial aggregation calculations, and no corresponding multiplication or accumulation operations are performed. Spatial aggregation is only performed within the same sampling time; that is, the current sampling time only uses the input feature vectors corresponding to the current sampling time of all delivery points to complete the aggregation, without mixing data across different sampling times. Spatial aggregation is performed independently at each sampling time, forming a spatially aggregated feature sequence for that sampling time. After spatial aggregation is completed, the feature transformation layer is implemented using a fully connected neural network, including an input layer, linear transformation units, bias terms, and nonlinear activation units. The linear transformation units are implemented using weight matrices, and the bias terms are implemented using bias vectors. Both the weight matrix and the bias vectors are randomly initialized during the model initialization phase.

[0052] S2.7 Based on spatial association features, a gated recurrent unit network is used to obtain the current hidden state, and the predicted number of rented vehicles and the predicted number of returned vehicles at the deployment point are obtained through the Softplus function.

[0053] Specifically, based on spatial association features, a gated recurrent unit network (GRU network) is used to obtain the current hidden state, and the predicted number of rented vehicles and the predicted number of returned vehicles at the deployment point are obtained through the Softplus function, expressed as follows: In the formula, Indicates the first Each delivery point in the future The One prediction parameter (predicting the number of rented vehicles or predicting the number of returned vehicles). This represents the Softplus smooth activation function. Indicates the first In the prediction step, with the first The prediction weight vector corresponding to each prediction parameter This indicates the transpose operation. Indicates time The hidden state, Indicates the first In the prediction step, with the first The prediction bias term corresponding to each prediction parameter; The Softplus smooth activation function is specifically expressed as follows: In the formula, Indicates the input value. Represents the natural logarithm function. Represents the natural constant.

[0054] It should be noted that a gated recurrent unit (GRU) network includes at least an update gate, a reset gate, and candidate hidden state computation units. Furthermore, before being used for actual prediction, the GRU network (i.e., a GRU network) constructs training samples using historical data. For any historical reference time, the training samples are constructed using continuous data prior to that time. The spatial correlation features of each time period are used as input, and then the subsequent continuous... The actual number of vehicles rented and returned in each time period is used as the supervision label; mean squared error or smoothed absolute error is preferably used as the prediction loss. For the number of rented and returned vehicles, the errors between the predicted values ​​and the actual values ​​for all deployment points and all prediction steps are calculated separately, and then weighted summation is performed to obtain the total training loss. The rental prediction loss and the return prediction loss can be given equal weights; for example, the rental vehicle prediction loss and the return vehicle prediction loss each account for half of the total loss.

[0055] S3. Based on the predicted number of rented vehicles and the predicted number of returned vehicles, calculate the safe inventory and expected number of vehicles at each deployment point, determine the deployment status of the deployment point as pending transfer, candidate transfer, and supply-demand balance, and then calculate the unit service value and marginal service loss increment coefficient of the deployment point.

[0056] S3.1. The root mean square of the positive error between the actual rental volume and the predicted rental volume is used as the error compensation benchmark.

[0057] Specifically, the predicted number of rented vehicles and the predicted number of returned vehicles are summed to obtain the cumulative predicted number of rented vehicles and the cumulative predicted number of returned vehicles. Retrieve the most recent delivery point (e.g., 30) completed units of the same length For each historical time window, the difference between the actual rental volume and the predicted rental volume is extracted. The positive error (i.e., the portion where the actual rental volume is greater than the predicted rental volume) is then extracted and the root mean square is calculated as the error compensation benchmark.

[0058] The positive error is truncated, and its expression is: In the formula, Indicates the first The first delivery point is at the... Positive error within a historical time window This indicates the operation of retrieving the maximum value. Indicates the first The first delivery point is at the... The difference between the actual number of rentals and the predicted number of rentals within a historical time window (this difference can be directly generated by subtracting the predicted number of rentals from the actual number of rentals). The root mean square (RMS) is used as the error compensation benchmark, and the expression is as follows: In the formula, Indicates the first The error compensation benchmark for each delivery point at the current time.

[0059] It should be noted that "recent" refers to selecting historical time windows from most recent to oldest, based on the current forecast time. The operating platform prioritizes historical time windows with end times earlier than the current forecast time and where actual rental data has been fully generated, and does not use time windows that have not yet ended or where actual order data has not yet been fully aggregated. Furthermore, the length of each historical time window is the same as the length H of the current forecast time window. For example, if the current forecast requires predicting the next 30 minutes, then each historical time window used for error statistics will also be 30 minutes; if the current forecast requires predicting the next 60 minutes, then each historical time window will also be 60 minutes. Window lengths of different lengths, such as 15 minutes, 30 minutes, and 60 minutes, must not be directly mixed to calculate the error compensation benchmark.

[0060] S3.2 Set the guarantee coefficient corresponding to the attribute matching of the current time period of the deployment point, calculate the number of vehicles required for guarantee based on the current error compensation benchmark and the cumulative predicted rental volume, and calculate the current required safe inventory based on the maximum vehicle capacity of the deployment point.

[0061] The expression for calculating the number of vehicles required to meet demand is as follows: In the formula, Indicates the first The number of vehicles required to meet the current demand at each delivery point. Indicates the first Cumulative projected rental volume for each location Indicates the first Each delivery point's current time period attributes are matched with a corresponding guarantee coefficient. Indicates rounding up; The required current security stock is calculated using the following expression: In the formula, Indicates the first The current required safety stock level for each delivery point This indicates the operation of finding the minimum value. Indicates the first The maximum number of vehicles that can be accommodated at each drop-off point.

[0062] It should be noted that: guarantee factor It is used to indicate the tolerance for uncertainty in forecast demand under different operating scenarios, and is essentially used to adjust the impact of historical forecast error compensation on the safety vehicle reserve; while the time attribute includes at least the time period category to which the current time belongs; For example, in one implementation: based on the usage patterns of shared electric vehicle users, a day is divided into multiple time attribute intervals, such as morning peak hours, morning off-peak hours, midday hours, afternoon off-peak hours, evening peak hours, and nighttime low-demand hours. The operating platform determines the corresponding time attribute based on the current system time, and then queries a pre-established time attribute-guarantee coefficient mapping table to obtain the guarantee coefficient corresponding to the current time period. For example, for periods with high user commuting demand, such as 7:00-9:00 AM and 5:00-7:00 PM on weekdays, the risk of vehicle shortages due to concentrated user rental demand and prediction errors is significant, so a higher guarantee coefficient can be set; for periods with low demand, such as 12:00-6:00 AM, a lower guarantee coefficient can be set to avoid maintaining excessively high vehicle reserves for extended periods.

[0063] The pre-established time attribute-guarantee coefficient mapping table can be configured according to the following example values: During weekday morning rush hour, the guarantee factor is set at 1.2; during weekday evening rush hour, the guarantee factor is set at 1.2; during weekday morning off-peak hours, the guarantee factor is set at 1.0; during weekday afternoon off-peak hours, the guarantee factor is set at 1.0; and during nighttime low-demand hours, the guarantee factor is set at 0.8. With these protection coefficient values, the historical positive forecast error compensation benchmark can be amplified to varying degrees. For example, a protection coefficient of 1.2 means an additional 20% error compensation on top of the historical forecast underestimation of risk; a protection coefficient of 0.8 means a reduction in the reserve of standby vehicles during periods of low demand.

[0064] S3.3. Starting with the current number of available vehicles at the drop-off point, the predicted return volume is added and the predicted rental volume is subtracted. After non-negativity constraint correction, the predicted number of vehicles after non-negativity constraint correction is generated, and the expression is: In the formula, Indicates the first The estimated number of vehicles at each delivery point after non-negativity constraint correction. This means that if the initial estimated number of vehicles is less than 0, the value is set to 0; otherwise, the original value is retained. Indicates the first The number of currently available vehicles at each delivery point Indicates the first The cumulative predicted return amount for each delivery point.

[0065] It should be noted that: if the initial estimated number of vehicles is greater than or equal to zero, this initial estimated number of vehicles will be used as the estimated number of vehicles after correction for non-negativity constraints; if the initial estimated number of vehicles is less than zero, the estimated number of vehicles will be corrected to zero. The negative part represents unmet demand where the predicted rental demand exceeds the current number of vehicles and the predicted total number of returned vehicles, and does not indicate that the deployment point will actually have a negative number of vehicles. Suppose that a certain drop-off point currently has 20 available vehicles, and the cumulative predicted return volume for the next 30 minutes is 8 vehicles, and the cumulative predicted rental volume is 24 vehicles. Then the initial estimated number of vehicles is 4, which remains 4 after applying non-negativity constraints. If the current available number of vehicles is 10, the cumulative predicted return volume is 5, and the cumulative predicted rental volume is 22, then the initial estimated number of vehicles is -7, which is set to zero after applying non-negativity constraints.

[0066] S3.4. Compare the expected number of vehicles with the safe inventory. If the expected number of vehicles is less than or greater than the safe inventory, generate the quantity to be transferred in and the maximum quantity to be transferred out, respectively.

[0067] Specifically, the projected number of vehicles will be compared with the safe number of vehicles in operation; If the expected number of vehicles is less than the safe inventory, it means that the delivery point will have insufficient vehicle supply during the scheduling time window, and vehicles need to be transferred from other delivery points. The number of vehicles to be transferred to the delivery point is obtained by subtracting the expected number of vehicles from the safe inventory. If the expected number of vehicles exceeds the safe inventory, it means that the deployment point still has surplus vehicles after meeting the expected rental demand and error protection requirements within the scheduling time window. The number of vehicles exceeding the safe inventory can be treated as ordinary available vehicles. The initial available quantity is obtained by subtracting the safe inventory from the expected number of vehicles. Then, it is checked whether the expected number of vehicles exceeds the maximum vehicle capacity of the deployment point. If it does, the capacity overflow is generated by subtracting the maximum vehicle capacity from the expected number of vehicles. Subsequently, the maximum available quantity is calculated using the current number of available vehicles at the deployment point, the cumulative predicted return quantity, the initial available quantity, and the capacity overflow.

[0068] The maximum adjustable amount is calculated using the following expression: In the formula, Indicates the first The maximum amount that can be adjusted at each delivery point.

[0069] S3.5. Based on the amount to be transferred in and the maximum amount that can be transferred out at the delivery point, generate the scheduling status of the delivery point, expressed as: In the formula, Indicates the first The scheduling status of each delivery point This indicates that the drop-off point is a drop-off point to be transferred. This indicates that the drop-off point is a candidate drop-off point. This indicates that the distribution point is a supply-demand balanced distribution point. Indicates the first The amount to be transferred in at each delivery point This indicates that both conditions are true simultaneously.

[0070] It should be noted that the scheduling status can be recalculated in each prediction task execution cycle. For example, with a sampling period of five minutes, it is updated every five minutes based on the latest number of currently available vehicles and the latest prediction results. For scheduling tasks that have been generated but not yet executed, the number of vehicles to be transferred in at the transfer-in point and the real-time number of vehicles to be transferred out at the transfer-out point should be re-verified before execution.

[0071] S3.6. Based on the scheduling status of the delivery points, establish a set of candidate delivery points to be transferred out and a set of delivery points to be transferred in.

[0072] The set of candidate delivery points is established, expressed as follows: In the formula, This represents the set of candidate delivery points within the current scheduling time window; The set of delivery points to be added is expressed as: In the formula, This represents the set of delivery points to be added within the current scheduling time window.

[0073] It should be noted that the candidate set of delivery points to be dispatched and the set of delivery points to be dispatched can be stored using arrays, lists, hash tables, or temporary database tables. To facilitate subsequent matching calculations, a unique index can be established based on the delivery point identifier, and the prediction start time and scheduling time window length can be used as a joint version identifier.

[0074] S3.7 Extract the cumulative predicted rental volume and maximum vehicle capacity of candidate drop-off points, calculate the predicted demand intensity and impose an upper limit constraint, and then calculate the basic service value of a single vehicle remaining at the drop-off point by combining the preset minimum basic value and value conversion coefficient.

[0075] The calculation of predicted demand intensity and the application of upper limit constraints are expressed as follows: In the formula, Indicates the first The predicted demand intensity of each delivery point within the scheduling time window. This indicates a positive number that prevents the denominator from being zero (e.g., ), Represents the first [unit] after the upper limit constraint. The predicted demand intensity of each delivery point within the scheduling time window. This represents the upper limit of the predicted demand intensity. The basic service value of a single vehicle remaining at this deployment point is calculated using the following expression: In the formula, This indicates that a shared electric scooter remains in the first... The basic service value of each delivery point This represents the minimum basic value of the vehicle's remaining value. This represents the value conversion coefficient corresponding to the predicted demand intensity.

[0076] It should be noted that: the upper limit of the predicted demand intensity. It can be determined based on historical operational data. In specific implementation, the distribution of predicted demand intensity of all candidate deployment points in the training set or the most recent stable operating cycle can be statistically analyzed, and the 95th percentile or 99th percentile value can be selected as the upper limit of predicted demand intensity. The minimum basic value of vehicle retention This parameter represents the minimum service value that a shared electric scooter should retain even when the predicted rental demand at a deployment point is very low. This parameter prevents the vehicle retention value at deployment points with zero or low demand from dropping directly to zero. If the minimum base value for vehicle retention is set to zero, then when the predicted demand intensity is also zero, the retention value of a single vehicle at that deployment point will drop to zero. This could lead to all available vehicles at that deployment point being relocated later, when the transport distance is short or the demand at the receiving point is high. To avoid a situation where low-demand deployment points are completely without vehicles due to prediction errors, the arrival of temporary users, or new orders, it is preferable to set the minimum base value to one. Secondly, the value conversion factor corresponding to the predicted demand intensity. This coefficient controls the increase in the base service value of a vehicle for each unit increase in predicted demand intensity. It matches the range of predicted demand intensity after an upper limit constraint; for example, when the minimum base value for vehicle retention is set to one, the value conversion coefficient corresponding to the predicted demand intensity can preferably be set to three. Thus, if the predicted demand intensity is zero, the base service value of a vehicle is one; when the predicted demand intensity is 0.5, the predicted demand intensity generates a value increment of 1.5; when the predicted demand intensity is one, it generates a value increment of three; and when the predicted demand intensity reaches the upper limit of two, it generates a value increment of six. At this point, the base service value of a vehicle is between one and seven, thus creating a suitable difference that can be used for subsequent scheduling and ranking.

[0077] S3.8 Extract the error compensation benchmark and cumulative predicted rental volume, calculate the demand uncertainty and impose an upper limit constraint, and then multiply it by the prediction error value conversion coefficient to obtain the added service value. Sum the added service value with the basic service value to obtain the unit service value of a single vehicle remaining at the deployment point.

[0078] The calculation of demand uncertainty and the application of upper limit constraints are expressed as follows: In the formula, Indicates the first The demand uncertainty of each delivery point within the scheduling time window Represents the first [unit] after the upper limit constraint. The demand uncertainty of each delivery point within the scheduling time window This represents a preset upper limit for the uncertainty of demand.

[0079] It should be noted that for most normally operating deployment points, the error compensation benchmark is usually lower than or close to the cumulative predicted rental volume, and the smoothed demand uncertainty is usually around zero to one. Therefore, the upper limit of demand uncertainty can preferably be set to one. When the upper limit of demand uncertainty is set to one, it means that the impact of historical demand underestimation error relative to the current predicted demand is treated as a maximum of 100% uncertainty level when entering the value calculation, thereby limiting the amplified impact of low demand anomalies or historical sudden demand on subsequent value calculations. It should also be noted that the value conversion coefficient corresponding to the forecast error is used to convert the dimensionless uncertainty of demand into a value that can be added to the value of the basic services. The forecast error value conversion coefficient is preferably a positive number. For example, in one implementation: when the upper limit of demand uncertainty is set to one, the value conversion coefficient corresponding to the prediction error can preferably be set to two; therefore, if the constrained demand uncertainty is zero, the value of the additional service is zero; when the constrained demand uncertainty is 0.2, the value of the additional service is 0.4; when the constrained demand uncertainty is 0.5, the value of the additional service is one; when the constrained demand uncertainty reaches the upper limit of one, the value of the additional service is two, thereby ensuring that the higher the demand uncertainty, the greater the value of the additional service of the vehicle remaining at the current deployment point.

[0080] S3.9. Determine the marginal service loss increment coefficient based on the maximum available capacity and the safe reserve capacity. The expression is as follows: In the formula, Indicates the first The marginal service loss increasing coefficient at each delivery point This indicates a uniformly set incremental adjustment parameter; It should be noted that the uniformly set incremental adjustment parameters It should be set to a non-negative number, preferably 0.5, and the numerical unit of this parameter should be consistent with the numerical scale of the unit service value so that the marginal incremental part and the basic unit service value can be compared in the same objective function.

[0081] S4. Construct a set of comprehensive path costs and candidate outbound incentive values. Combine the unit service value and the marginal service loss increment coefficient to obtain the minimum total scheduling cost of the outbound volume of the delivery point. Then, generate a scheduling task work order for execution and distribution.

[0082] S4.1 For any delivery point and any delivery point The system obtains the shortest possible route and the estimated total scheduling time for dispatched vehicles through the road route calculation service.

[0083] It should be noted that: for any delivery point and any delivery point For any delivery point in the candidate delivery point set, and any delivery point in the set of delivery points to be added The road route calculation service can be set to Gaode or Baidu Maps.

[0084] S4.2, Query delivery points The predicted sequence of steps where the number of vehicles first falls below the safe stock level within the scheduling time window is determined, and the latest arrival time is calculated.

[0085] Drop-off point The first predicted step when the number of vehicles falls below the safe inventory within the scheduling time window is to query the drop-off points. .

[0086] The latest arrival time is determined by the following expression: In the formula, This indicates that the dispatched vehicle has arrived at the designated delivery point. The latest arrival time, Indicates the current scheduling decision time. Indicates the duration of a single forecast period (e.g., 5 minutes). This indicates the reserved buffer time (5 minutes is acceptable).

[0087] S4.3. Based on the estimated total scheduling time, perform a selection process for the deployment point combination to obtain the retained deployment point combination.

[0088] Specifically, for any candidate delivery point... and delivery points to be transferred If the total estimated scheduling time is greater than the maximum allowed scheduling time, the current deployment point combination will be excluded; if the total estimated scheduling time is less than or equal to the maximum allowed scheduling time, the current deployment point combination will be retained.

[0089] It should be noted that: a drop-off point combination refers to a directed dispatch combination formed by a candidate drop-off point and a drop-in point. The dispatch direction is to transport shared electric vehicles from the candidate drop-off point to the drop-in point. Reverse combinations should be judged as another independent candidate combination. The existence of a passable road between two drop-off points should not be used to assume that the two-way dispatch conditions are the same.

[0090] S4.4 For the retained deployment point combination, the difference between the maximum scheduling duration and the expected total scheduling time will be used as the scheduling time margin, starting from the nearest... Extract the actual scheduling time from historical scheduling tasks that pass through the same road area and are in the same time period, calculate the standard deviation of the scheduling time, and combine it with the scheduling time margin to calculate the delay risk.

[0091] The specific calculation expression for the maximum scheduling duration is as follows: In the formula, This indicates the maximum scheduling time allowed from the current time to the latest arrival time to complete the scheduling; The formula for calculating delay risk is: In the formula, Indicates that at the current moment, the delivery point to the drop-off point The risk of delays This represents an exponential function with the natural constant as its base. Indicates the delivery point at the current time. to the drop-off point Scheduling time margin Indicates the drop-off point to the drop-off point The corresponding historical scheduling time standard deviation, This indicates a positive number that prevents the denominator from being zero.

[0092] It should be noted that the maximum scheduling duration is the complete duration from the current scheduling decision time to the latest arrival time.

[0093] Secondly, the same road area refers to a road network that historical scheduling tasks and current candidate scheduling tasks mainly pass through or highly overlap with each other. In specific implementation, any of the following methods can be used for determination; The first method uses a fixed path composed of outgoing and incoming drop points as the road area identifier, and only counts historical scheduling tasks between the same outgoing drop point and the same incoming drop point. The second method involves establishing road area codes based on the main road numbers or main road combinations returned by the map service. When two paths primarily pass through the same set of road numbers, they are considered to belong to the same road area. This invention preferably adopts the second method. The "same time period type" refers to historical and current dispatch tasks occurring within time periods with similar traffic patterns; time period types can be divided using fixed time periods. For example, the entire day can be divided into morning peak, off-peak, midday, evening peak, and nighttime periods. In one implementation, it can be divided into five categories. For example, 7:00 to 9:00 a.m. is the morning peak, 9:00 to 11:30 a.m. is the morning off-peak, 11:30 a.m. to 1:30 p.m. is the midday period, 4:30 p.m. to 7:30 p.m. is the evening peak, and the rest of the time is the ordinary off-peak or nighttime period.

[0094] S4.5. The shortest path time, estimated total scheduling time, and delay risk are weighted and combined to generate a comprehensive path cost, expressed as: In the formula, Indicates that at the current moment, the delivery point to the drop-off point The overall path cost The weight represents the shortest time path. Indicates that at the current moment, the delivery point to the drop-off point The shortest time path (i.e., road distance). The weights representing the estimated total scheduling time Indicates that at the current moment, the delivery point to the drop-off point The estimated total scheduling time, The weight representing the risk of delay; The scheduling cost for a single vehicle is obtained by multiplying a preset value cost coefficient by the comprehensive path cost. .

[0095] It should be noted that before weighted fusion, the shortest time path, the estimated total scheduling time, and the delay risk are first normalized; while for each weight... , , The settings must all be non-negative numbers, and the sum of the three weights must equal one; For example, in one implementation: preferably, the road distance weight is set to 0.3, the estimated total scheduling time weight is set to 0.4, and the delay risk weight is set to 0.3. Under this weight setting, the estimated total scheduling time accounts for a relatively high proportion of the comprehensive path cost, while road distance and delay risk are evaluated as transportation resource consumption and task reliability factors, respectively. Secondly, the preset value cost coefficient is used to establish the correspondence between the overall route cost and the dispatch cost of a single shared electric vehicle. The value cost coefficient should be set to a positive number; its specific value should be consistent with the range of unit service value. For example, when the unit service value of a vehicle remaining at a deployment point is typically between one and seven, the value cost coefficient can be set to five to ten, ensuring that the dispatch cost of most vehicles is on a similar order of magnitude to the unit service value. However, if the value cost coefficient is too small, such as setting it to 0.1, the cost differences between different dispatch routes will have little impact on the dispatch optimization results. If the value cost coefficient is too large, for example, set to more than one hundred, the path cost may completely outweigh the service value of the delivery point. Therefore, the present invention preferably sets the preset value cost coefficient to 10.

[0096] S4.6. Pre-set the lower limit of incentives, the upper limit of incentives, and the fixed increment, and generate a set of candidate incentive values ​​to be removed according to the fixed increment.

[0097] The set of candidate stimulus values ​​to be retrieved is expressed as: In the formula, This represents the set of candidate activation values. Indicates the first Each candidate incentive value was removed. This indicates a pre-set lower limit for incentives. This indicates the sequence number of the candidate activation value. This represents the total number of candidate incentive values ​​to be removed. Among them, the total number of candidate incentive values ​​is determined by the lower limit of the incentive. Incentive cap and fixed increment To obtain, the specific expression is: It should be noted that: the incentive lower limit can be set to 0 yuan / vehicle for example, the incentive upper limit can be set to 30 yuan / vehicle for example, and the fixed increment can be set to 1 yuan / vehicle for example, thereby generating 31 candidate incentive values ​​ranging from 0 yuan / vehicle, 1 yuan / vehicle, 2 yuan / vehicle to 30 yuan / vehicle. Secondly, the incentive range and fixed increment can also be adjusted based on historical dispatch costs, average vehicle service revenue, or operating budget.

[0098] S4.7 For any candidate outgoing delivery point, query the delivery points to be transferred in that have a time-feasible candidate scheduling path, form a set of delivery points to be transferred in, and use the quantity to be transferred in as a weight, combined with the set of delivery points to be transferred in, to perform a weighted average of the scheduling cost, and obtain the average generalized execution cost corresponding to the delivery point.

[0099] The average generalized execution cost corresponding to the delivery point is obtained by the following expression: In the formula, Indicates candidate removal of delivery points The average generalized execution cost at the current moment. It indicates that it belongs to.

[0100] It should be noted that: the delivery point to be transferred in with a time-feasible candidate scheduling path does not refer to all delivery points with delivery needs, but rather to the delivery point to be transferred in that, for the candidate delivery point to be transferred out currently being processed, there is at least one candidate path that starts from the candidate delivery point to be transferred out, can be actually passed by the dispatched vehicle, and is expected to arrive before the latest arrival time of the corresponding delivery point to be transferred in. If there are multiple time-feasible paths from the same candidate outgoing delivery point to the same incoming delivery point, the path with the lowest overall path cost should be selected first, and the scheduling cost of a single vehicle corresponding to that path should be used as the scheduling cost between the candidate outgoing delivery point and the incoming delivery point.

[0101] S4.8. Subtract the average generalized execution cost from the candidate release incentive value in the candidate release incentive value set, and then subtract the unit service value to obtain the unit net incentive that the released vehicle can obtain. If the unit net incentive is greater than zero, calculate the feedback release amount based on the marginal service loss increment coefficient; otherwise, the feedback release amount is zero.

[0102] The feedback outflow amount is calculated based on the marginal service loss increment factor, and the expression is as follows: In the formula, Indicates candidate removal of delivery points In the Feedback on the number of vehicles to be removed under each candidate removal incentive value (this removal amount is the number of vehicles removed). This indicates that a vehicle remains in the dispatch time window. The unit service value of each delivery point.

[0103] It should be noted that net incentive per unit represents the remaining value after deducting the platform's average transportation execution cost and the loss of unit service value due to the loss of a shared electric vehicle at a current deployment point, after the platform pays the unit incentive, and is sufficient to compensate for the continued deployment of vehicles. Therefore, net incentive per unit reflects the remaining revenue that can truly be used to compensate for marginal retention service losses after releasing vehicles, not the total incentive paid by the platform. For example, if the current candidate deployment incentive value is twelve value units, the average generalized execution cost is three value units, and the unit service value is five value units, then the net incentive per unit is four value units.

[0104] Furthermore, the set of candidate call-out incentive values ​​is traversed, and the total feedback call-out amount of all candidate call-out delivery points under each incentive value is calculated; With the goal of minimizing the total expenditure on scheduling incentives and covering the total demand for transfers in, the smallest candidate transfer incentive value that ensures the total feedback transfer amount is not less than the total transfer amount to be transferred in is selected as the optimal transfer incentive. The objective, which is to minimize the total expenditure on scheduling incentives while covering the total demand to be transferred in, is expressed as follows: In the formula, This represents the minimum output excitation value; It should be noted that the candidate stimulus value set should be sorted in ascending order of value and duplicate values ​​should be removed before entering the traversal step. If there are candidate stimulus items with the same value but different identifiers in the set, only one of them should be retained for calculation to avoid repeatedly performing the same feedback aggregation. Substitute the optimal pull-out stimulus into the feedback pull-out amount. The formula yields the optimal feedback transfer amount under the optimal transfer-out incentive, which is used as the final number of vehicles to be transferred out. The final number of vehicles dispatched will be used as the supply node capacity, and the number of vehicles to be dispatched to the delivery points will be used as the demand node capacity, taking into account the comprehensive path cost. After setting the edge weights, perform minimum total scheduling cost modeling and set constraints; The model for minimizing total scheduling cost is specifically expressed as follows: In the formula, This represents the minimum total scheduling cost. Indicates delivery point Actual dispatch to the delivery point The number of vehicles (to be solved); Set the constraint condition, the expression is: It should be noted that the final number of vehicles dispatched represents the maximum number of vehicles that the supply node can provide, not the number of vehicles that have already been determined to perform transportation.

[0105] The minimum total scheduling cost of the model is solved by using operations research algorithms. After the maximum number of iterations, the specific number of vehicles allocated from each candidate outgoing deployment point to each incoming deployment point and the scheduling path are output. Based on the specific number of vehicles allocated and the dispatch route, a dispatch task work order is generated, which includes the origin and destination points, the number of dispatched vehicles, the estimated arrival time, and the execution incentive amount. The dispatch task work order is then pushed to the handheld smart terminal of the dispatch staff through the Internet of Things communication network (such as 4G / 5G), and the staff executes the dispatch according to the dispatch task work order.

[0106] It should be noted that during the solution process, virtual source nodes and virtual sink nodes are added to the transportation network. The virtual source nodes connect to each candidate outgoing delivery point, with the capacity of the corresponding connection edge set to the feedback outgoing quantity of the corresponding candidate outgoing delivery point, and the cost of the connection edge set to zero. Each delivery point to be delivered connects to the virtual sink node, with the capacity of the corresponding connection edge set to the delivery quantity to be delivered to the corresponding delivery point to be delivered, and the cost of the connection edge set to zero. Subsequently, the target flow to be transported from the virtual source nodes to the virtual sink nodes is set to the sum of the delivery quantities to be delivered from all delivery points to be delivered. The target flow represents the total number of shared electric vehicles that need to be allocated in this round of scheduling. The successive shortest augmenting path algorithm first initializes the vehicle allocation on all transportation edges to zero and performs iterations. During iteration, starting from the virtual source node, it searches for a path with the lowest cumulative cost to reach the virtual sink node. This path sequentially passes through the virtual source node, one candidate outgoing delivery point, one delivery point to be delivered, and the virtual sink node. After finding the lowest-cost path, determine the number of vehicles that can be allocated on that path this time; after iterating to the maximum number of times, output the specific number of vehicles allocated from each candidate dispatch point to each dispatch point to be dispatched and the scheduling path.

[0107] This embodiment also provides a shared electric vehicle dispatching system based on big data analysis, including: The sample generation module is used to obtain the original data records of the deployment points from the shared electric vehicle operation platform, perform validity verification, and form a time-series sample set of multiple deployment points. The return prediction module is composed of a multi-deployment point time series sample set. It calculates the actual road distance between any two different deployment points, the two-way vehicle flow intensity, and the Pearson correlation coefficient of rental demand changes. Then, it constructs a deployment point association map, performs spatial aggregation, feature transformation, and prediction to obtain the predicted number of rented vehicles and the predicted number of returned vehicles. The status determination and parameter calculation module is used to calculate the safe inventory and expected number of vehicles at each deployment point based on the predicted number of rented vehicles and the predicted number of returned vehicles. After determining the scheduling status of the deployment point, such as pending transfer, candidate transfer, and supply-demand balance, it then calculates the unit service value and marginal service loss increment coefficient of the deployment point. The scheduling task generation module is used to construct a set of comprehensive path costs and candidate dispatch incentive values. After combining the unit service value and the marginal service loss increment coefficient, it obtains the feedback dispatch volume of the delivery point, performs the minimum total scheduling cost modeling and solution, and generates a scheduling task work order for execution and distribution.

[0108] In summary, by constructing a multi-source information association graph that integrates actual road distance, two-way vehicle flow intensity, and changes in rental demand, and employing a spatiotemporal graph neural network for multi-step rental demand prediction, the prediction model can effectively capture the spatial dependencies and temporal evolution patterns between deployment points. Furthermore, by extracting the root mean square of historical positive prediction errors as a compensation benchmark and establishing an asymmetric error compensation mechanism in conjunction with time-period attribute guarantee coefficients, the negative impact of prediction deviations on scheduling decisions is effectively reduced. Secondly, by constructing a vehicle retention unit service value model based on predicted demand intensity and demand uncertainty, cumulative retention service loss, a set of candidate withdrawal incentive values, and solutions for optimal withdrawal incentives and minimum total scheduling costs, the synergistic optimization of supply and demand matching, path reliability, and incentive economy is achieved. Therefore, this invention significantly improves the supply and demand matching effectiveness and comprehensive scheduling efficiency of the shared electric vehicle scheduling method.

[0109] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A shared electric vehicle dispatching method based on big data analysis, characterized in that: include, The original data records of the deployment points are obtained from the shared electric vehicle operation platform and their validity is verified to form a time-series sample set of multiple deployment points. Based on a time series sample set of multiple delivery points, after calculating the actual road distance, two-way vehicle flow intensity, and Pearson correlation coefficient of rental demand changes between any two different delivery points, a delivery point association graph is constructed to perform spatial aggregation, feature transformation, and prediction to obtain the predicted number of rented vehicles and the predicted number of returned vehicles. Based on the predicted number of rented vehicles and the predicted number of returned vehicles, the safety stock and expected number of vehicles at each deployment point are calculated. After determining the deployment point's status of pending transfer in, candidate transfer out, and supply-demand balance, the unit service value and marginal service loss increment coefficient of the deployment point are then calculated. Construct a set of comprehensive path costs and candidate outbound incentive values. Combine the unit service value and marginal service loss increment coefficient to obtain the minimum total scheduling cost of the outbound volume of the delivery point. Then, generate a scheduling task work order for execution and distribution.

2. The shared electric vehicle dispatching method based on big data analysis as described in claim 1, characterized in that: The original data records of deployment points obtained from the shared electric vehicle operation platform are validated for validity, forming a time-series sample set of multiple deployment points, as detailed below: The shared electric vehicle operation platform establishes a basic information table of deployment points and collects data according to a unified sampling period to obtain the original data records of each deployment point; Based on the original data records, records with empty placement point identifiers or collection times are removed. Then, the total number of vehicles available for rent in a given period is obtained by summing the number of available vehicles and the number of returned vehicles. After correction, the corrected data is sorted to generate a multi-placement point time-series sample set.

3. The shared electric vehicle dispatching method based on big data analysis as described in claim 2, characterized in that: Based on a multi-deployment-point time-series sample set, the Pearson correlation coefficient of the actual road distance, two-way vehicle flow intensity, and rental demand changes between any two different deployment points is calculated, as follows: Call the road path calculation service to obtain the two-way road distance between any two delivery points; The intensity of two-way vehicle flow is obtained by statistically analyzing the number of two-way rentals and returns between each pair of shared electric vehicle order placement points throughout the history. After performing alignment operations on a time series sample set of multiple delivery points, the Pearson correlation coefficient between the changes in rental demand at each delivery point is calculated.

4. The shared electric vehicle dispatching method based on big data analysis as described in claim 3, characterized in that: The process of constructing the deployment point association graph involves spatial aggregation, feature transformation, and prediction to obtain the predicted number of rented vehicles and the predicted number of returned vehicles, as detailed below: The actual road distance, two-way vehicle flow intensity, and Pearson correlation coefficient are compared with the corresponding preset thresholds. If any two different delivery points meet at least one of the road distance condition, vehicle flow condition, or coefficient correlation condition, then an association edge is established between the current delivery points; otherwise, no association edge is established. The actual road distance, two-way vehicle flow intensity, and Pearson correlation coefficient are weighted and integrated to form the weights of the associated edges, and then an adjacency matrix is ​​constructed. The input feature vector of the delivery point is constructed and fed into the feature transformation layer to obtain the spatial correlation features of the delivery point; Based on spatial association features, a gated recurrent unit network is used to obtain the current hidden state, and the predicted number of rented vehicles and the predicted number of returned vehicles at the drop-off point are obtained through the Softplus function.

5. The shared electric vehicle dispatching method based on big data analysis as described in claim 4, characterized in that: The calculation of the safe inventory and expected number of vehicles at each deployment point is used to determine the scheduling status of the deployment points, including those awaiting transfer in, those being considered for transfer out, and those in a supply-demand balance, as detailed below: The root mean square error between the actual rental volume and the predicted rental volume is used as the error compensation benchmark. Set the guarantee coefficient corresponding to the attribute matching of the current time period of the deployment point, combine the current error compensation benchmark and the cumulative predicted rental volume to calculate the number of vehicles required for guarantee, and combine the maximum vehicle capacity of the deployment point to calculate the current required safe inventory. Starting with the current number of available vehicles at the drop-off point, the predicted return volume is added and the predicted rental volume is subtracted. After non-negative constraint correction, the predicted number of vehicles after non-negative constraint correction is generated. The projected number of vehicles is compared with the safe inventory. Under the conditions that the projected number of vehicles is less than and greater than the safe inventory, the quantity to be transferred in and the maximum quantity to be transferred out are generated respectively. The scheduling status of the delivery point is generated based on the amount to be transferred in and the maximum amount that can be transferred out.

6. The shared electric vehicle dispatching method based on big data analysis as described in claim 5, characterized in that: The calculation of the unit service value and marginal service loss increment coefficient of the delivery point is as follows: Based on the scheduling status, establish a set of candidate delivery points to be dispatched out and a set of delivery points to be dispatched in; Extract the cumulative predicted rental volume and maximum vehicle capacity of candidate drop-off points, calculate the predicted demand intensity and impose an upper limit constraint, and then calculate the basic service value of a single vehicle remaining at the drop-off point by combining the preset minimum basic value and value conversion coefficient. Extract the error compensation benchmark and cumulative predicted rental volume, calculate the demand uncertainty and impose an upper limit constraint, and then multiply it by the prediction error value conversion coefficient to obtain the added service value. Sum the added service value with the basic service value to obtain the unit service value of a single vehicle remaining at the deployment point. The marginal service loss increment coefficient is determined based on the maximum available capacity and the safe holding capacity.

7. The shared electric vehicle dispatching method based on big data analysis as described in claim 6, characterized in that: Before obtaining the feedback call-out volume of the delivery point, the comprehensive path cost should be obtained first, which specifically includes: For any delivery point and any delivery point The road route calculation service obtains the shortest time path that dispatched vehicles can take and the estimated total dispatch time. Query delivery points The predicted sequence of steps where the number of vehicles first falls below the safe stock level within the scheduling time window is determined, and the latest arrival time is calculated. Based on the estimated total scheduling time, a combination of delivery points is selected to obtain the retained combination of delivery points; For the retained deployment point combination, the difference between the maximum scheduling duration and the expected total scheduling time will be used as the scheduling time margin, starting from the most recent Extract the actual scheduling time from historical scheduling tasks that pass through the same road area and are in the same time period, calculate the standard deviation of the scheduling time, and combine it with the scheduling time margin to calculate the delay risk. The shortest path time, the estimated total scheduling time, and the delay risk are weighted and integrated to generate a comprehensive path cost.

8. The shared electric vehicle dispatching method based on big data analysis as described in claim 7, characterized in that: The specific details of obtaining the feedback call-out volume of the delivery point are as follows: Pre-set the lower limit of incentives, the upper limit of incentives, and the fixed increment, and generate a set of candidate incentive values ​​to be removed according to the fixed increment; For any candidate outgoing delivery point, query the delivery points to be transferred in that have a time-feasible candidate scheduling path, form a set of delivery points to be transferred in, and use the number of delivery points to be transferred in as a weight, and combine it with the set of delivery points to be transferred in to perform a weighted average of the scheduling cost, so as to obtain the average generalized execution cost corresponding to the delivery point. The net incentive per unit that a released vehicle can obtain is obtained by subtracting the average generalized execution cost from the candidate release incentive value in the candidate release incentive value set, and then subtracting the unit service value. If the net incentive per unit is greater than zero, the feedback release amount is calculated based on the marginal service loss increment coefficient; otherwise, the feedback release amount is zero.

9. The shared electric vehicle dispatching method based on big data analysis as described in claim 8, characterized in that: After modeling and solving for the minimum total scheduling cost, a scheduling task work order is generated and issued, specifically including: With the goal of minimizing the total expenditure of scheduling incentives and covering the total demand for transfers, the smallest candidate transfer incentive value that makes the total feedback transfer amount not less than the total number of transfers is selected as the optimal transfer incentive. Substituting this value into the formula for the feedback transfer amount, the optimal feedback transfer amount under the optimal transfer incentive is obtained as the final number of vehicles to be transferred. The final number of vehicles dispatched is taken as the supply node capacity, and the number of vehicles to be dispatched to the destination points is taken as the demand node capacity. After using the comprehensive path cost as the edge weight, the minimum total scheduling cost model is performed and constraints are set. The minimum total scheduling cost model is solved by the operations research algorithm, and the specific vehicle allocation quantity and scheduling path from each candidate dispatch point to each destination point are output. Based on the specific number of vehicles allocated and the dispatch route, a dispatch task work order is generated, which includes the origin and destination points, the number of dispatched vehicles, the estimated arrival time, and the execution incentive amount. The dispatch task work order is then pushed to the handheld smart terminal of the dispatch staff through the Internet of Things communication network, and the staff executes the dispatch according to the dispatch task work order.

10. A shared electric vehicle dispatching system based on big data analysis, based on the shared electric vehicle dispatching method based on big data analysis as described in any one of claims 1 to 9, characterized in that: include, The sample generation module is used to obtain the original data records of the deployment points from the shared electric vehicle operation platform, perform validity verification, and form a time-series sample set of multiple deployment points. The return prediction module is composed of a multi-deployment point time series sample set. It calculates the actual road distance between any two different deployment points, the two-way vehicle flow intensity, and the Pearson correlation coefficient of rental demand changes. Then, it constructs a deployment point association map, performs spatial aggregation, feature transformation, and prediction to obtain the predicted number of rented vehicles and the predicted number of returned vehicles. The status determination and parameter calculation module is used to calculate the safe inventory and expected number of vehicles at each deployment point based on the predicted number of rented vehicles and the predicted number of returned vehicles. After determining the scheduling status of the deployment point, such as pending transfer, candidate transfer, and supply-demand balance, it then calculates the unit service value and marginal service loss increment coefficient of the deployment point. The scheduling task generation module is used to construct a set of comprehensive path costs and candidate dispatch incentive values. After combining the unit service value and the marginal service loss increment coefficient, it obtains the feedback dispatch volume of the delivery point, performs the minimum total scheduling cost modeling and solution, and generates a scheduling task work order for execution and distribution.

Images