Shared bicycle multi-target dynamic scheduling method based on double-matrix variable neighborhood search

Abstract

The invention relates to the technical field of shared bicycle scheduling methods, in particular to a shared bicycle multi-target dynamic scheduling method based on double-matrix variable neighborhood search. The system comprises four core components, namely a scheduling problem modeling module, a double-matrix coding module, an improved genetic algorithm module and a dynamic scheduling decision module. Constructing a multi-objective optimization model containing site topological data, operation constraint parameters and demand prediction data through a scheduling problem modeling module; the dual-matrix coding module adopts a dual coding structure of a path matrix and a scheduling quantity matrix, and respectively represents vehicle path planning and station loading and unloading quantity distribution; the improved genetic algorithm module implements a collaborative optimization strategy of hybrid crossover operation and variable neighborhood search; and the dynamic scheduling decision module finally outputs an optimal scheduling scheme meeting multi-target balance. According to the invention, through a double-matrix decoupling optimization architecture and an intelligent search mechanism, the problems of complex constraint and multi-target optimization in dynamic scheduling of shared bicycles are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of shared bicycle dispatching methods, in particular to a shared bicycle multi-objective dynamic dispatching method based on double-matrix variable neighborhood search. BACKGROUND

[0002] With the acceleration of urbanization and the popularity of green travel concept, shared bicycles, as an important supplement to "last mile" travel, have been widely used around the world. However, due to the dynamic changes in user demand and the complexity of urban spatial structure, shared bicycle systems generally face the problem of unbalanced distribution of vehicles. To solve this problem, existing technologies usually adopt a static dispatching method, i.e., dispatching bicycles between stations based on historical data at night to meet user demand during the day.

[0003] Although the existing technology can reduce the dispatching cost to some extent and alleviate part of the supply and demand contradiction, it still has obvious deficiencies in real-time dynamic response capability, which is specifically manifested as follows: the existing dispatching scheme is mainly based on fixed time window and historical data for planning. This dispatching method, which only aims to minimize the dispatching cost, cannot respond to the dynamic fluctuations in demand in real time, resulting in a serious imbalance in the distribution of vehicles at stations during peak demand periods such as early morning and evening, which not only fails to meet the actual needs of users, but also makes it difficult to achieve reasonable allocation of resources. SUMMARY

[0004] The present application provides a shared bicycle multi-objective dynamic dispatching method based on double-matrix variable neighborhood search, which constructs a dispatching model through station topological data, operation constraint parameters and demand prediction data, and uses a double-encoding structure of path matrix and dispatching matrix to represent the scheme. Through mixed crossover operation and variable neighborhood search operator, the double matrix is optimized cooperatively, and the population diversity is maintained by combining congestion distance calculation. Finally, the dispatching scheme that meets the multi-objective balance is output, thereby solving the problems raised in the background technology, i.e.: The shared bicycle system cannot respond to demand fluctuations in real time due to static dispatching, resulting in an imbalance in the distribution of vehicles at stations.

[0005] To achieve the above purpose, the shared bicycle multi-objective dynamic dispatching method includes the following steps: S1, obtaining station topological data, operation constraint parameters and demand prediction data of a shared bicycle system; S2, constructing a double-matrix encoding structure according to the station coordinates and the distance matrix between stations in the station topological data, and combining the vehicle load limit and the station capacity limit in the operation constraint parameters to generate an initial population; S3, performing mixed crossover operation on the path matrix and the dispatching matrix in the initial population, randomly exchanging path nodes by point crossover, and exchanging continuous gene fragments by block crossover to generate offspring individuals; S4, sequentially applying four path optimization operators of site replacement, exchange, movement and inversion to the path matrix in the offspring individuals, and simultaneously applying two optimization operators of numerical fine-tuning and exchange to the scheduling matrix, to complete variable neighborhood search; S5, dividing the solution set after variable neighborhood search into four quadrant regions in the target space, calculating the crowding distance through the normalized scheduling cost and unmet demand rate, and screening high-quality solutions with sparse distribution; S6, repeatedly performing hybrid crossover operation and variable neighborhood search until the non-dominated solution set is output, and generating the vehicle scheduling path and the loading and unloading amount scheme of each station according to the path matrix and the scheduling matrix in the non-dominated solution set.

[0006] In the above technical solution, the double-matrix encoding structure (path matrix + scheduling matrix) is used instead of single encoding, because the traditional single encoding cannot simultaneously consider the discreteness of path planning and the continuity of scheduling quantity distribution. If only path encoding is used, the scheduling quantity distribution will lack accurate control, and if only scheduling quantity encoding is used, it is difficult to ensure the feasibility of the path. The hybrid crossover operation (point crossover + block crossover) is designed instead of single crossover strategy, because point crossover can enhance global search ability but is easy to damage high-quality path fragments, block crossover can preserve good gene combinations but has limited search range, and the complementation of the two can avoid premature convergence of the algorithm. Six variable neighborhood search operators (four paths and two scheduling quantities) are set instead of unified operation, because path optimization requires topology adjustment and scheduling quantity optimization focuses on numerical fine-tuning. If the operator type is simplified, the local search will not be sufficient. The four-quadrant target space is divided for diversity maintenance, which can avoid over-concentration of solution set in a specific target dimension compared to traditional crowding distance calculation. If the partition protection is cancelled, the Pareto front will be unevenly distributed.

[0007] On this basis, the improved genetic algorithm module adopts a dynamic adjustment of crossover probability mechanism, and automatically adjusts the operation proportion of point crossover and block crossover according to the population diversity index during the evolution process.

[0008] In another technical solution, the improved genetic algorithm module transmits the non-dominated solution set to a dynamic scheduling decision module for evaluation and screening of the non-dominated solution set, and selects a scheduling scheme according to a preset decision preference.

[0009] The technical scheme adopts a dynamic adjustment cross probability mechanism instead of a fixed ratio, because a static cross strategy cannot adapt to the requirements of different stages of evolution, and needs to strengthen block crossover to reserve high-quality gene fragments when the population diversity is high, and needs to increase point crossover to jump out of local optimization when the population tends to be the same, and if the fixed ratio is maintained, the search efficiency will be low; the non-dominated solution set evaluation and screening mechanism is designed instead of directly outputting the optimal solution, because actual scheduling needs to flexibly trade off cost and service quality according to different scenes (such as peak / flat periods), and if a single solution is adopted, the decision flexibility will be lost. The combination of the dynamic parameter adjustment and the multi-solution reservation strategy not only ensures the adaptive ability of the algorithm in the search process, but also provides a configurable decision space for actual operation.

[0010] Compared with the prior art, the beneficial effects of the present application are: The present application adopts a hierarchical progressive optimization logic in algorithm design: first, the scheduling problem is decoupled into path planning and load allocation two sub-problems through double-matrix coding, wherein the path matrix adopts a dynamic variable-length structure to record the time-space trajectory of the vehicle, and the scheduling matrix indicates the station operation through signed integer; second, a hybrid evolution strategy is designed, the complementary mechanism of point crossover and block crossover is adopted in the global search stage, and the dedicated neighborhood operator is configured for the path and the scheduling matrix respectively in the local optimization stage; finally, target space partition management is introduced, the normalized double-objective solution set is divided into sector areas according to the angle, and the solution set distribution is maintained through the crowded distance screening. The three-layer architecture of 'decoupling-coordination-control' makes the algorithm not only be able to handle mixed variable types, but also be able to maintain the balance of multi-objective optimization. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 It is a method flowchart of the present application; Figure 2 It is a whole flow structure schematic diagram of embodiment 1 of the present application; Figure 3 It is a gene coding mode schematic diagram of embodiment 1 of the present application; Figure 4 It is a population crossover mode schematic diagram of embodiment 1 of the present application; Figure 5 It is a scheduling route matrix neighborhood search operator schematic diagram of embodiment 1 of the present application; Figure 6 It is a scheduling single vehicle quantity matrix neighborhood search operator schematic diagram of embodiment 1 of the present application; Figure 7 It is a region protection schematic diagram of embodiment 1 of the present application; Figure 8 It is a CR-INS algorithm calculation flowchart of embodiment 1 of the present application; Figure 9 It is a CitiBike scheduling area schematic diagram of embodiment 2 of the present application; Figure 10 Figure for average HV value of embodiment 2 of the present application; Figure 11 Figure for average IGD value of embodiment 2 of the present application; Figure 12 Figure for non-dominated solution set distribution of embodiment 2 of the present application; Figure 13 Figure for non-dominated solution set distribution of embodiment 2 of the present application; Figure 14 Figure for non-dominated solution set distribution of embodiment 2 of the present application; Figure 15 Figure for non-dominated solution set distribution of embodiment 2 of the present application; Figure 16 Figure for non-dominated solution set distribution of embodiment 2 of the present application; Figure 17 Figure for non-dominated solution set distribution of embodiment 2 of the present application.

[0012] The meanings of various labels in the figures are as follows: 100, scheduling problem modeling module; 200, double matrix encoding module; 300, improved genetic algorithm module; 400, dynamic scheduling decision module. DETAILED DESCRIPTION

[0013] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0014] Meanwhile, some technical terms are explained here: The Pareto frontier refers to the mapping of the non-dominated solution set in the objective function space, which is manifested as the trade-off curve of the scheduling cost and the unmet demand rate and other optimization objectives. Each point thereof corresponds to the objective function value of a non-dominated solution, and is used to intuitively show the superior-inferior relationship between schemes; The non-dominated solution set in the present application refers to the scheduling scheme set filtered out by a multi-objective optimization algorithm, wherein any scheme is not dominated by other schemes in all optimization objectives. Specifically, it is manifested as: among the feasible solutions satisfying the vehicle path constraint and the loading and unloading amount distribution, the solution achieving the optimal trade-off in the scheduling cost and the demand satisfaction rate and other objectives.

[0015] Embodiment 1 The present application provides a shared bicycle multi-objective dynamic scheduling method based on double-matrix variable neighborhood search, as shown in Figure 1 and specifically comprises the following steps: S1, acquiring the station topological data, operation constraint parameters and demand prediction data of the shared bicycle system; S2, constructing a double-matrix coding structure according to the station coordinates and the distance matrix between stations in the station topological data, combining the vehicle load limit and the station capacity limit in the operation constraint parameters, and generating an initial population; S3, performing a hybrid crossover operation on the path matrix and the scheduling matrix in the initial population, randomly exchanging path nodes by point crossover and exchanging continuous gene fragments by block crossover to generate offspring individuals; S4, applying four path optimization operators of station replacement, exchange, movement and inversion to the path matrix in the offspring individuals in turn, and simultaneously applying two optimization operators of numerical fine-tuning and exchange to the scheduling matrix to complete variable neighborhood search; S5, dividing the solution set after variable neighborhood search into four fan-shaped regions in the objective space, calculating the crowding distance through the normalized scheduling cost and the unmet demand rate, and screening high-quality solutions with sparse distribution; S6, repeatedly performing the hybrid crossover operation and the variable neighborhood search until the non-dominated solution set is output, and generating the vehicle scheduling path and the station loading and unloading amount scheme according to the path matrix and the scheduling matrix in the non-dominated solution set.

[0016] as shown in Figure 2 and specifically comprises a scheduling problem modeling module 100, a double-matrix coding module 200, an improved genetic algorithm module 300 and a dynamic scheduling decision module 400, which work collaboratively to realize a dynamic optimization scheduling scheme that minimizes the scheduling cost while maximizing the satisfaction of user bike rental and return demands.

[0017] The core task of the scheduling problem modeling module 100 is to convert the shared bicycle dynamic scheduling problem into an optimization model that can be processed by computer algorithms. This model aims to clearly plan the vehicle travel route and station bicycle loading and unloading operation, so as to most effectively solve the problem of station vehicle surplus or shortage caused by user riding, while balancing the company's operation cost and user's use experience.

[0018] The scheduling problem modeling module 100 first constructs a complete scheduling system model. Based on the actual operation of the site distribution and vehicle scheduling demand, the entire system is abstracted into a network graph composed of sites and connecting paths, generating a topological data structure containing site coordinates, site distance matrix, wherein the scheduling center is numbered 0, and other sites are numbered 1 to N in turn, forming a complete network graph model. In this model, we configure two types of scheduling vehicles, trucks and electric tricycles, all vehicles start from the unified scheduling center, and perform dynamic scheduling tasks within a day according to the time division of one hour per period. The specific scheduling tasks are divided into two types: one is to transport surplus bicycles from the vehicle accumulation site (unloading operation), and the other is to supplement bicycles to the vehicle shortage site (loading operation). To ensure the orderly progress of scheduling, the system sets that each site is only allowed to have one scheduling vehicle operating at each site in each period, but different vehicles can schedule different sites in parallel. Based on this model, the invention determines two key decisions: one is to plan the specific driving route for each scheduling vehicle in each period, and to determine which site it starts from and goes to; the second is to determine the specific operation (loading or unloading) of the vehicle at each stop site and the number of bicycles operated.

[0019] In order to make the model solvable and consistent with the actual situation, the scheduling problem modeling module 100 sets some prerequisites to generate the operation constraint parameter set, including: vehicle maximum load (truck / electric tricycle capacity), site capacity limit, scheduling time window division, unit loading and unloading cost, unit distance transportation cost, prerequisites include: before scheduling starts, the required basic information is ready, such as which vehicles can be used, how many bicycles each vehicle can carry at most, the distance between sites, how many users want to borrow and return bicycles at each site at different times, etc., and finally generate demand prediction data, including the borrowing demand of each site at each period. All vehicles must start from the scheduling center and finally return to the scheduling center. The road conditions and vehicle status during the entire scheduling period are stable, and each transportation task can be completed within one hour.

[0020] Shared bicycle companies aim to reduce scheduling costs and reduce unmet user borrowing and returning needs, thereby increasing revenue. Therefore, this model minimizes the scheduling cost of shared bicycles as one of the optimization objectives, and introduces unmet user demand as another optimization objective, constructing a double-objective optimization function. Among them, the objective function of minimizing scheduling cost is ; In the formula, represents the minimization of the total scheduling cost; represents the scheduling vehicle number (1 to ); ​ denotes the fixed use cost of the vehicle; denotes the station number (1 to ); denotes the time window number (1 to ); denotes whether the vehicle travels from station to in the time period (0 / 1 variable); denotes the loading / unloading amount of the vehicle at station in the time period (positive for loading, negative for unloading); denotes the unit loading / unloading operation cost of a vehicle; =0 / , =0 / denotes the set of stations including the dispatch center (number 0); 1 denotes the travel distance from station to ; denotes the unit distance transportation cost; The objective is composed of three parts: the first part is the vehicle fixed cost, representing the total fixed expenditure of K dispatch vehicles; the second part is the vehicle dispatch cost, which is proportional to the total amount of loading / unloading operations; the third part is the transportation distance cost, which depends on the total travel distance of the vehicles. Another optimization objective is to minimize the unmet user demand: ; In the formula, denotes the second optimization objective to be minimized (the total amount of unmet user demand); denotes the total amount of unmet borrowing and returning demands of all stations in all time periods; denotes the number of unmet user demands of station in time window , including the borrowing demand (no vehicle available to borrow) and the returning demand (no slot available to return).

[0021] ​​​The scheduling scheme must comply with the rules of the real world: all the cars must go out from the dispatch center and return after completing the task. The same car cannot go to the same site for two consecutive periods. A site can only have one car operating in a period. The number of single cars transported by the car on the road cannot exceed the maximum loading capacity of the car. The number of vehicles parked in the site cannot exceed the maximum capacity of the site. When the car arrives at a site to load, the number of single cars that can be loaded cannot exceed the number of idle cars that the site can actually provide, nor the remaining loading space of the car itself; when unloading, the number of single cars that can be unloaded cannot exceed the number of single cars actually loaded on the car, nor the number of idle slots that the site can accommodate at the time.

[0022] The scheduling problem modeling module 100 generates a complete shared bicycle dynamic scheduling mathematical model framework by integrating system input data. The core modeling results output by this module include: operation constraint parameters (including hard constraint conditions such as vehicle maximum load and site capacity limit), demand prediction data based on historical data analysis (accurate to the hourly single bicycle supply and demand difference), and site topology data. These structured modeling data are completely transmitted to the double-matrix coding module 200 through a standardized interface, providing a strict mathematical basis and input parameters for the subsequent genetic algorithm coding scheme design. Among them, the operation constraint parameters ensure that the coding scheme meets the actual operation requirements, the demand prediction data supports reasonable allocation of scheduling quantities, and the double-objective optimization function points out the direction for algorithm optimization.

[0023] The double-matrix coding module 200 is responsible for converting the complex constraint conditions of the shared bicycle scheduling problem into a computable genetic algorithm coding form. This module receives three types of key parameters from the system input layer: site topology data, including the latitude and longitude coordinates of 23 sites and the Manhattan distance matrix between sites; operation constraint parameters, including the capacity limit of trucks and electric tricycles, and the division of the scheduling time window; demand prediction data, including the single bicycle supply and demand difference of each site at different times.

[0024] Based on these inputs, the module constructs an innovative double-matrix coding system: the path matrix adopts a dynamic variable-length coding structure, each gene unit records the vehicle ID, time window, departure station, and target station four-tuple information, and through a verification algorithm to ensure that the constraint condition of "the vehicle must depart from the dispatch center and return" is met; the scheduling quantity matrix adopts signed integer coding, positive values represent loading operations, negative values correspond to unloading operations, and an internal capacity verification mechanism is built in to ensure that the scheduling quantity of each site does not exceed the current inventory and the remaining load of the vehicle.

[0025] The first step of genetic algorithm is to code the individuals. Generally, individuals can be coded by a variety of structures, including binary, floating-point and symbolic coding, etc. In this invention, considering the characteristics of the proposed model, the coding design adopts a double-matrix structure, which represents the scheduling path of vehicles and the scheduling quantity of single vehicles respectively. Each individual contains two parts of information: one is the scheduling route of each vehicle per day, which records the complete path from the scheduling center, visiting stations in turn and returning to the scheduling center; the other is the number of scheduling single vehicles in the time window corresponding to each station (positive value for loading, negative value for unloading). The meaning of coding is to ensure that the vehicle path is not in conflict (adjacent time period station is different, different vehicles at the same time period station is different) and implicitly contains the capacity constraint (scheduling quantity does not exceed the vehicle load and station capacity) through structured representation.

[0026] As shown in Figure 3 , a vehicle scheduling scheme in a day can be represented as: "scheduling center 0→1st station→2nd station→3rd station→4th station→9th station→5th station→1st station→6th station→7th station→8th station 1→9th station→scheduling center 0", and the number of scheduling single vehicles is "0 (scheduling center does not operate), +2 (load 2 vehicles), 8, -5 (unload 5 vehicles), 2, 10, -7, -3, 7, 12, -9, 4, 0 (return to scheduling center)".

[0027] In the population initialization stage, the module adopts a random generation algorithm based on constraint satisfaction: first, generate a random path sequence for each scheduling vehicle that satisfies the time window constraint, then calculate the feasible scheduling quantity range based on the demand difference of stations, and finally output the initial population. Each individual contains a complete scheduling scheme, including: vehicle path planning matrix, station scheduling quantity allocation matrix and objective function estimate value. These structured data are transmitted to the improved genetic algorithm module 300 through a standardized data interface, providing a high-quality initial solution set for subsequent optimization calculation.

[0028] The innovation of this module lies in three aspects: first, the double-matrix coding design decouples the discrete path planning and continuous scheduling quantity decision, which not only maintains the feasibility of the solution but also improves the search efficiency; second, the population initialization algorithm based on constraint propagation effectively improves the proportion of feasible solutions compared with the traditional random generation method; finally, the coding scheme reserves a dynamic adjustment interface, which supports real-time updating of station demand data during the algorithm running process, providing technical scalability for emergency response. These characteristics make the generated initial population not only meet the strict operation constraints but also maintain sufficient diversity, laying a solid foundation for the efficient optimization of the improved genetic algorithm module 300.

[0029] Although the double-matrix encoding module 200 has provided a clear structured representation for the bike-sharing dispatching problem, transforming the complex dispatching scheme into a computable path matrix and loading-unloading matrix, this static encoding method still has obvious limitations. In actual operation, we found that simply relying on the initial scheme generated by the encoding module is difficult to adapt to the dynamically changing demand environment, especially in special scenarios such as sudden weather, holidays, or large-scale activities, the preset dispatching plan often deviates seriously from the actual demand, and the traditional genetic algorithm is prone to fall into local optimal solution when dealing with such multi-objective optimization problems, resulting in that the dispatching scheme cannot achieve a good balance between cost and user demand satisfaction. Therefore, we introduced the improved genetic algorithm module 300, which can continuously optimize the dispatching scheme and dynamically respond to the real-time changing demand mode through innovative evolution mechanism and diversity maintenance strategy.

[0030] The improved genetic algorithm module 300 of the present application adopts a double-matrix variable neighborhood search algorithm (NRMOEA), the core of which is to realize dynamic optimization through the combination of double-matrix encoding structure (path matrix and dispatch matrix) and variable neighborhood search operator. The improved genetic algorithm module 300 first receives the initial population generated by the double-matrix encoding module 200, including the dispatch path matrix and the dispatch matrix, as well as the preset algorithm parameters such as the crossover rate, the neighborhood search rate, the population size, etc. As shown in Figure 4 Based on these inputs, the algorithm first performs evolution operations on the population through a hybrid crossover strategy: point crossover randomly exchanges part of the nodes of the parent individuals, enhancing the global search ability; block crossover exchanges continuous gene fragments, preserving the high-quality path structure. During the crossover process, the model constraints are strictly followed to ensure that the generated offspring individuals always satisfy the vehicle capacity, station inventory, and other limitations. As shown in Figure 5 、 Figure 6 In the mutation phase, the algorithm uses six variable neighborhood search operators to fine-tune the solution, including four path optimization operators (station replacement, exchange, movement, and inversion) and two dispatch quantity optimization operators (quantity change and quantity exchange), which act on the population in random order, significantly improving the local search efficiency.

[0031] As shown in Figure 7 To maintain the diversity of the solution set, the algorithm introduces a region protection mechanism, which divides the target space into four fan-shaped regions with an interval of 22.5°, forcing each region to retain a certain number of solutions, avoiding excessive concentration of the search process in a certain local optimal region. At the same time, as shown in Figure 8As shown, the improved CR-INS non-dominated sorting algorithm prioritizes high-quality solutions with sparse distribution by normalizing objective function values and calculating crowded distances, ensuring the uniformity and comprehensiveness of the Pareto front. During the iterative optimization process, the algorithm dynamically adjusts the search strategy, automatically increasing the disturbance intensity of the neighborhood search operator when detecting a decline in population diversity, effectively avoiding premature convergence. After 1500 generations of evolution, the algorithm outputs a set of non-dominated solutions, including scheduling schemes that achieve the best trade-off between scheduling costs and unmet demand. Based on the path matrix and scheduling matrix in the non-dominated solution set, vehicle scheduling paths and station loading and unloading schemes are generated.

[0032] In terms of algorithm parameter settings, the selection of crossover rate and neighborhood search rate is based on extensive pre-experiments to ensure a balance between exploring new solutions and utilizing known high-quality solutions. The population size ensures sufficient diversity while avoiding excessive computational overhead. For constraint handling, the algorithm uses a repair strategy to ensure that all solutions meet operational constraints, including the requirement that vehicles must depart from and return to the dispatch center and the restriction that only one vehicle is allowed to operate at a single station during the same time period. In terms of computational complexity, NRMOEA optimizes the neighborhood search and non-dominated sorting processes through parallel computing, significantly improving the operational efficiency of large-scale problems.

[0033] Finally, the improved genetic algorithm module 300 passes the non-dominated solution set to the dynamic scheduling decision module 400, which includes detailed vehicle path planning and station-by-station vehicle allocation quantities, directly guiding actual scheduling operations. Each solution in the scheme is verified for feasibility to ensure compliance with all operational constraints. In addition, the algorithm outputs the objective function values and solution set distribution characteristics of each solution, providing decision-makers with multi-dimensional scheme evaluation criteria. These data are transmitted in real-time to the scheduling system through standardized interfaces, supporting dynamic responses to demand fluctuations and unexpected situations, and achieving seamless integration from algorithm optimization to implementation. The entire process fully demonstrates the adaptability and practicality of the NRMOEA algorithm in the actual shared bicycle scheduling scenario, providing a reliable solution for multi-objective optimization problems in complex dynamic environments.

[0034] The dynamic scheduling decision module 400 receives the non-dominated solution set optimized by the improved genetic algorithm module 300, which contains multiple feasible scheduling schemes that achieve the optimal balance between scheduling cost and user demand satisfaction. The module first performs multi-dimensional evaluation on the input solution set, filters the scheduling scheme most suitable for the current scenario from the solution set based on the preset decision preference weight (such as cost priority or demand priority) and real-time operation state (such as the number of available vehicles, traffic conditions, etc.), and performs detailed feasibility verification for each candidate scheme, including checking path continuity, vehicle load limit, station capacity constraint and other key operation conditions to ensure that the scheme can be directly executed. In the scheme selection process, the module uses a multi-criteria decision method to consider the implementation cost, expected demand satisfaction rate, scheduling length and other key indicators of the scheme, while supporting a manual decision intervention interface to allow operation personnel to adjust the automatically selected scheme according to special circumstances.

[0035] The final scheduling scheme will be converted into specific execution instructions, including the detailed driving route of each scheduled vehicle, the accurate loading and unloading quantity of each station and the corresponding time window arrangement, which will be transmitted in real time to the vehicle terminal and station management system through the scheduling system interface. In addition, the module also has dynamic adjustment capability, which can trigger the scheme re-optimization process when the system detects that the actual execution deviates from the expected (such as vehicle delay, demand mutation, etc.), and request the algorithm module to generate a new scheduling scheme to ensure that the system always maintains the optimal operation state. The entire decision-making process ensures the scientificity of the scheme while taking into account the flexibility and operability of actual operation, effectively improving the dynamic response capability and overall operation efficiency of the shared bicycle system.

[0036] By innovatively constructing a double-matrix coding system, improving the genetic algorithm optimization mechanism, and establishing an intelligent dynamic decision-making process, the present application realizes the optimal balance between operation cost and user demand for shared bicycle scheduling schemes, effectively controls the operation cost while ensuring service quality, and provides a complete technical solution for intelligent scheduling of the shared bicycle industry.

[0037] Example 2 The present application also verifies the actual optimization effect of the proposed NRMOEA algorithm through the following experimental examples, and demonstrates its performance advantage in the dynamic scheduling problem of shared bicycles through a comparative experiment. For example, Figure 9As shown, the experiment selects the operation data of the New York CitiBike system from October 28 to November 7, 2024 as a test case, the dispatching area covers 23 high-frequency use stations in Lower Manhattan, and the experimental environment is configured as a Python 3.9.7 platform (CPU: Intel Xeon Gold 5218R@2.10GHz, memory 32GB). The positive center of the dispatching area is the dispatching center of the area, and all transport vehicles start from the dispatching center, load a certain number of bicycles, and return to the dispatching center after completing the dispatching task.

[0038] As shown in Table 1, comparative experiments were carried out on the three algorithms, and the HV and IGD values of the solution sets of the three algorithms on 10 instances are shown, and the optimal value of each instance is shown in bold. From the experimental results (see Table 1), it can be analyzed that: Table 1 IGD and HV values of three algorithms on 10 instances

[0039] Based on the double-matrix variable neighborhood search algorithm (NRMOEA) proposed in the present application, two variants NRMOEA-PC (point crossover dominant) and NRMOEA-BC (block crossover dominant) are designed, and compared with the traditional NSGA-II algorithm to verify the influence of different crossover strategies on the optimization effect. As shown in Figure 10 , 11 The HV index comparison curve of the three algorithms clearly shows that NRMOEA-BC obtains the highest hyper-volume value after 1500 iterations, indicating that its solution set has the optimal coverage and diversity in the target space; the IGD index change trend proves that NRMOEA-BC can approach the real Pareto front more quickly. In terms of instance analysis, Figure 12 , 13 , 14, 15 show the non-dominated solution set distribution comparison results of the three algorithms in four typical scenarios. In Figure 12 instance 1, the solution set of NRMOEA-BC (red dots) shows a clear L-shaped distribution characteristic, and achieves an optimization effect of 85%-95% demand satisfaction in the dispatching cost interval of 3000-5000 yuan, while the NSGA-II solution set (blue dots) is only concentrated in the high-cost interval.

[0040] Figure 13 The NRMOEA-BC solution set forms a continuous optimal front, especially in the high service quality interval of demand satisfaction >90%, its dispatching cost is reduced by an average of 18% compared with NSGA-II. Figure 14 In instance 2, NRMOEA-PC (green dots) and NRMOEA-BC solution sets perform similarly in the low demand interval (<80%), but NRMOEA-BC still has a clear advantage in the high service quality interval.Figure 15 As the most difficult optimization case, the solution sets of the three algorithms appear partial overlap, but the NRMOEA-BC still maintains an independent optimal solution cluster in the lower left area, which proves the robustness of the algorithm. The coordinate axes in each subgraph are normalized, the horizontal axis represents the scheduling cost (ten thousand yuan), the vertical axis represents the unmet demand rate (%), and the gray dotted line is the theoretical optimal reference line.

[0041] In terms of algorithm component verification, as shown in Figure 16 , Figure 17 , it shows the independent test results of six neighborhood search operators on instance 1 and instance 2. The position replacement operator ( Figure 16 ) and the number movement operator ( Figure 17 ) show the strongest optimization ability, and their solution sets can cover the key interval of low scheduling cost and high demand satisfaction at the same time. These experimental results fully confirm the effectiveness of the block crossover mechanism, variable neighborhood search operator and region protection strategy proposed in the present application, and provide a superior performance optimization solution for shared bicycle dynamic scheduling.

[0042] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A shared bicycle multi-objective dynamic scheduling method based on double-matrix variable neighborhood search, characterized in that, The method comprises the following steps: S1, acquiring station topological data, operation constraint parameters and demand prediction data of a shared bicycle system; S2, constructing a double-matrix coding structure and generating an initial population according to station coordinates and a distance matrix between stations in the station topological data, in combination with vehicle load limits and station capacity limits in the operation constraint parameters; S3, performing a hybrid crossover operation on a path matrix and a scheduling matrix in the initial population, randomly exchanging path nodes by point crossover and exchanging continuous gene segments by block crossover to generate offspring individuals; S4, applying four path optimization operators of station replacement, exchange, movement and inversion to the path matrix in the offspring individuals in sequence, and applying two optimization operators of numerical fine-tuning and exchange to the scheduling matrix to complete variable neighborhood search; S5, dividing the solution set after variable neighborhood search into four fan-shaped regions in the objective space, calculating congestion distance through scheduling cost and unmet demand rate after normalization processing, and screening high-quality solutions with sparse distribution; S6, repeatedly performing the hybrid crossover operation and variable neighborhood search until a non-dominated solution set is output, and generating a vehicle scheduling path and a loading and unloading amount scheme for each station according to the path matrix and the scheduling matrix in the non-dominated solution set. 2.The shared bicycle multi-objective dynamic scheduling method based on double-matrix variable neighborhood search according to claim 1, wherein: The scheduling problem modeling module (100) collects station location information and calculates the actual distance between stations to generate the station topological data. 3.The shared bicycle multi-objective dynamic dispatching method based on double-matrix variable neighborhood search of claim 2, wherein: The scheduling problem modeling module (100) generates the operation constraint parameters by setting vehicle load limits and station capacity requirements. 4.The shared bicycle multi-objective dynamic scheduling method based on double-matrix variable neighborhood search of claim 2, wherein: The scheduling problem modeling module (100) generates the demand prediction data by analyzing historical vehicle data and predicting demand at each station.

5. The double-matrix variable neighborhood search based shared bike multi-objective dynamic scheduling method according to claim 2, characterized in that: The scheduling problem modeling module (100) constructs a scheduling cost objective function, which includes vehicle fixed cost, loading and unloading operation cost, and transportation distance cost, the transportation distance cost is calculated according to the distance between stations, and the loading and unloading operation cost is calculated according to the loading and unloading amount and unit cost. 6.The shared bicycle multi-objective dynamic dispatching method based on double-matrix variable neighborhood search of claim 1, wherein: The double-matrix coding module (200) encodes using a double-matrix structure of a path matrix and a scheduling matrix, the path matrix records the vehicle travel path, and the scheduling matrix records the loading and unloading amount of each station.

7. The double-matrix variable neighborhood search based shared bike multi-objective dynamic scheduling method according to claim 6, characterized in that: The double-matrix coding module (200) calculates the feasible scheduling matrix range according to the station demand difference and vehicle load limit when generating the initial population. 8.The shared bicycle multi-objective dynamic dispatching method based on double-matrix variable neighborhood search of claim 1, wherein: The improved genetic algorithm module (300) adopts a dynamic adjustment crossover probability mechanism, and automatically adjusts the operation proportion of point crossover and block crossover according to the population diversity index in the evolution process. 9.The double-matrix variable neighborhood search based shared bike multi-objective dynamic scheduling method according to claim 8, characterized in that: The improved genetic algorithm module (300) transmits the non-dominated solution set to the dynamic scheduling decision module (400) for evaluation and screening of the non-dominated solution set, and selects a scheduling scheme according to a preset decision preference.

10. The double-matrix variable neighborhood search based shared bike multi-objective dynamic scheduling method according to claim 9, characterized in that: The dynamic scheduling decision module (400) converts the selected scheduling scheme into specific execution instructions, and the specific execution instructions include vehicle travel routes and loading and unloading amounts of each station.

Images