Intercity travel multi-objective vehicle path planning method based on diversity enhancement mechanism

Abstract

This invention provides a multi-objective vehicle path planning method for intercity travel based on a diversity enhancement mechanism, belonging to the field of intelligent transportation technology. The method includes: initializing the population using a time-series-based greedy strategy to improve the quality of the initial solution; introducing an objective selection strategy to dynamically allocate computational resources based on performance improvements during population evolution; designing a two-stage enhanced crossover operation to retain high-quality path genes from parent generations; and proposing eight diversity mutation operators with different granularities and methods, combined with an adaptive selection operator strategy to escape local optima. This effectively solves the diversity loss problem in large-scale multi-objective VRP, providing a high-quality path planning solution for the platform.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and more specifically to a multi-objective vehicle path planning method for intercity travel based on a diversity enhancement mechanism. Background Technology

[0002] With the acceleration of urbanization and the development of regional economic integration, the demand for intercity passenger travel has shown a significant growth trend. Compared with traditional intra-city ride-hailing services or fixed-route long-distance buses, emerging intercity carpooling services are gradually becoming an important part of intercity transportation due to their advantages such as door-to-door pick-up and drop-off, high travel flexibility, and good economy. In the core operation of this service model, the quality of solving the Vehicle Routing Problem (VRP) directly determines the platform's service efficiency and operating costs. Specifically, the platform needs to plan reasonable routes and dispatch schemes for the fleet based on the pre-booked order information of passengers (including pick-up point, destination, number of passengers, and expected time window). However, intercity travel introduces unique complexities: First, the long distances and time required for a single trip mean that any detours will result in extremely high additional costs. Second, the problem is subject to a series of complex constraints, including not only the usual vehicle capacity and time window constraints, but also the sequential constraint that "all orders within the same city must be delivered before accepting orders for the next city," as well as safety constraints such as mandatory rest after the driver's longest continuous driving time. Finally, the optimization objective of this problem has typical multi-objective conflict characteristics, involving a game of interests among the platform (e.g., minimizing the number of vehicles and total travel distance), drivers (e.g., maximizing vehicle occupancy to increase revenue), and passengers (e.g., minimizing waiting time), making it an extremely complex non-convex, nonlinear multi-objective optimization problem.

[0003] For multi-objective vehicle routing problems with time windows (MOVRPTW), existing technologies mainly rely on multi-objective evolutionary algorithms (MOEAs). Among these, the most representative algorithms include the non-dominated sorting genetic algorithm (NSGA-II) and the multi-objective evolutionary algorithm based on decomposition (MOEA / D). NSGA-II maintains the convergence and diversity of the population through fast non-dominated sorting and crowding distance calculation; while MOEA / D decomposes the multi-objective problem into a set of single-objective sub-problems for collaborative solution. In addition, some studies have introduced improvement strategies such as multi-directional local search to enhance algorithm performance.

[0004] While these traditional algorithms perform reasonably well in handling small- to medium-scale or standard VRP problems, they reveal several significant shortcomings when dealing with large-scale, highly constrained intercity travel scenarios. First, the algorithms converge slowly and are highly dependent on the quality of the initial solutions. Traditional algorithms often employ random initialization strategies to generate the initial population. However, in intercity travel problems, constraints such as time windows, intercity travel order, and driving time are extremely strict. Randomly generated initial solutions often contain a large number of infeasible or poor-quality solutions, causing the algorithm to consume significant computational resources in the early stages of the search to find feasible regions, severely slowing down the overall convergence process. Second, the diversity of the solution set is easily lost during evolution. As the number of orders increases and the problem size expands, the difficulty of searching the solution space increases exponentially. In the later stages of evolution, traditional algorithms (such as NSGA-II) often see their population rapidly cluster near a local Pareto front, resulting in "premature convergence." This means the algorithm loses its ability to explore unexplored regions, and the generated non-dominated solution set is unevenly distributed and has a narrow coverage, making it difficult to provide decision-makers with diverse and well-balanced compromise solutions. Third, existing algorithms have limited local search capabilities. Common crossover and mutation operators fail to effectively utilize the structured characteristics of intercity travel problems (such as the holistic and grouped nature of journeys). Simple random perturbation operations often disrupt the optimized path structure in the parent generation, making it difficult for the algorithm to escape local optima, especially when optimizing objectives with strong conflicting relationships, such as "number of vehicles" and "total travel distance".

[0005] In summary, when applied to the multi-objective vehicle routing problem in the specific scenario of intercity travel, existing technologies struggle to maintain the diversity of solution sets while ensuring solution accuracy.

[0006] In view of the above, this application is hereby submitted. Summary of the Invention

[0007] This invention provides a multi-objective vehicle route planning method for intercity travel based on a diversity enhancement mechanism, which can at least partially improve the above-mentioned problems.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A multi-objective vehicle routing planning method for intercity travel based on a diversity enhancement mechanism is proposed. This method uses a pre-built multi-objective vehicle routing problem model for intercity travel and includes the following steps: Obtain the set of intercity orders to be processed (Ord), perform trip initialization processing on the intercity order set to generate a trip set (Trips), and perform vehicle allocation initialization processing on the trip set (Trips) to generate an initial population; The initial population is iteratively optimized. Specifically, computing resources are dynamically allocated based on the performance improvement of each generation of the population on each objective. The optimization objective function for the current iteration is selected. Based on the selected optimization objective function, an enhanced crossover operation is performed on the corresponding parent individuals to generate a crossover scheme. A mutation operator is selected from a preset mutation operator pool to perform a mutation operation on the crossover scheme to generate a mutant scheme. The crossover scheme and the mutant scheme are compared, and the scheme with the smaller value of the optimization objective function is selected to be added to the offspring population. Merge the current iteration's offspring population with the historical reserve, update the external archive according to the dominance determination rule, and when the preset maximum running time is reached, output the non-dominated solution set in the external archive as the final intercity travel vehicle route planning scheme.

[0010] In summary, to address the shortcomings of existing intercity travel route planning models in reflecting characteristics and balancing algorithm convergence and diversity, this invention constructs an optimization model (i.e., a multi-objective vehicle routing problem model) that incorporates four objectives: minimizing the number of empty seats, minimizing the number of vehicles, minimizing the total travel distance, and minimizing the total passenger waiting time. This method is based on an improved Non-Dominated Sorting Genetic Algorithm (NSGA-II) framework, integrating dynamic resource allocation strategies and adaptive operator selection techniques.

[0011] Compared with existing technologies, this method has the following advantages: For large-scale intercity travel scenarios, it defines the scenario as a multi-objective problem with four optimization objectives, more comprehensively and realistically reflecting the essence of intercity travel problems. It overcomes the shortcomings of traditional algorithms that easily get trapped in local optima and have uneven solution set distribution when dealing with high-dimensional multi-objectives and complex constraints. Greedy initialization ensures the quality of the search starting point, objective selection strategy improves search efficiency, and enhanced crossover and adaptive mutation mechanisms greatly enrich the diversity of the solution set. By combining these mechanisms, it successfully achieves a high-quality set of planning solutions for intercity travel with different needs using multi-objective optimization methods, thereby achieving a deep synergy between economic benefits and service levels in complex intercity operation environments. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the multi-objective vehicle route planning method for intercity travel based on a diversity enhancement mechanism provided in this embodiment of the invention.

[0013] Figure 2 This is a schematic diagram of the hierarchical structure of the solution provided in the embodiments of the present invention.

[0014] Figure 3 This is a physical schematic diagram of the vehicle driving path provided in an embodiment of the present invention.

[0015] Figure 4 This is a framework diagram of the intercity travel multi-objective vehicle route planning method based on diversity enhancement mechanism provided in the embodiments of the present invention.

[0016] Figure 5 This is a schematic diagram of the enhanced crossover operation provided in an embodiment of the present invention.

[0017] Figure 6 This is a schematic diagram of the enhanced cross-operation process provided in an embodiment of the present invention.

[0018] Figure 7 This is a diagram of the operator structure for the diversity mutation operation provided in the embodiments of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] refer to Figures 1 to 4As shown, the first embodiment of the present invention discloses a multi-objective vehicle routing planning method for intercity travel based on a diversity enhancement mechanism. This method uses a pre-built multi-objective vehicle routing problem model for intercity travel, which can be executed by a multi-objective vehicle routing planning device for intercity travel based on a diversity enhancement mechanism (hereinafter referred to as the planning device). Specifically, it is executed by one or more processors within the planning device to implement the following method: It should be noted that the optimization objective functions of the multi-objective vehicle routing problem model for intercity travel include: minimizing the number of empty seats, minimizing the number of vehicles, minimizing the total travel distance, and minimizing the total passenger waiting time. The formulas are as follows: , , , ; in, To optimize the objective function by minimizing the number of empty seats in a vehicle, To optimize the objective function by minimizing the number of vehicles, To optimize the objective function by minimizing the total travel distance, To optimize the objective function by minimizing the total passenger waiting time, K is the set of all vehicles, and k is the vehicle index. Let Q be the road segment between node i and node j, Q be the maximum passenger capacity of a vehicle, and A be the set of all road segments. This represents the path variable used to determine whether vehicle k travels directly from node i to node j. For vehicle k on the road section Real-time load on Let k be the activation variable for vehicle k. For road section Let P be the distance between all orders and p be the order index of the current trip. The service start time for order p for vehicle k. Let p be the earliest expected departure time. Assign a variable to the order p of vehicle k.

[0021] In this embodiment, the present invention first establishes a multi-objective optimization model for intercity carpooling scenarios. A complete solution consists of a set of vehicle routes, each route containing multiple trips ordered chronologically. Based on the actual needs of the platform, drivers, and passengers, the following four optimization objectives are determined. The objective function for minimizing the number of empty seats is to increase driver profits by minimizing the number of empty seats in each trip; the objective function for minimizing the number of vehicles is to reduce platform expenses by minimizing the number of vehicles; the objective function for minimizing the total travel distance is to reduce vehicle fuel consumption and thus reduce costs by minimizing the total travel distance; and the objective function for minimizing the total passenger waiting time is to improve passenger satisfaction by minimizing the total passenger waiting time.

[0022] The output of the multi-objective vehicle routing problem model for intercity travel satisfies the constraint rules, and its formula is as follows: (The real-time passenger load and the number of passengers in a single order on the restricted road section shall not exceed the maximum passenger capacity of the vehicle.) (The earliest pickup time for a vehicle to accept a new trip is no earlier than the latest disembarkation time for all passengers in the current trip.) (Each order is limited to being served by a single vehicle) (Limited to pickup and delivery points for the same order being located on the same route, and delivery time being later than pickup time). (The number of times a vehicle visits the same node during a single trip is limited to no more than once.) (The pickup and delivery points of a restricted order must satisfy flow balance, and the flow must be consistent with the allocation variables.) (The pickup service is restricted to start within the passenger's preset time window.) (When the cumulative driving time of a vehicle reaches a preset threshold (e.g., 4 hours), a mandatory rest period (e.g., 20 minutes) will be included in the total task time.) in, Let p be the number of passengers. For vehicle k, the set of orders for the current journey. For vehicle k, the set of orders for the next trip. For order indexing of the next trip, Let k be the time when vehicle k arrives at the drop-off point of order p. For order collection The latest time for all passengers to drop off the bus. Let this be the drop-off point for order p. Let k be the time it takes for vehicle k to reach the next pickup point. For order collection The earliest passenger's pickup and boarding time. The travel time from the pick-up point to the drop-off point for order p. Let this be the pick-up point for order p. This indicates whether vehicle k travels directly from node i to the boarding point. Path variables, This indicates determining whether vehicle k has disembarked from the designated drop-off point. The path variable that leads directly to node j. Let p be the expected latest departure time. Let be the time when vehicle k finally arrives at its destination, and n be the total number of orders in the set P of all orders. Let K be the cumulative driving time of vehicle k. For rest duration, Let k be the resting variable for vehicle k.

[0023] In this embodiment, all generated solutions must satisfy the following hard constraints, capacity constraints: the number of passengers served by all vehicles does not exceed the maximum passenger capacity; specifically, the constraint formula restricts from two dimensions: order demand and real-time path load. First, it requires that the number of passengers in a single order must not exceed the maximum passenger capacity of the vehicle. Second, it requires that the total real-time passenger load of the vehicle on any travel arc in the same travel path must not exceed the maximum passenger capacity of the vehicle.

[0024] Service Constraint 1: To ensure the quality of ride-hailing services, any vehicle must allow all passengers on the current route to disembark before picking up the next group of new passengers. Specifically, the constraint formula is used to limit the vehicles... The constraint governs the temporal order between two adjacent trips; it mandates that a vehicle can only begin the next trip's pickup task after completing all delivery tasks for the current trip and returning to an empty state. Service Constraint Two: Each order is served by one vehicle. Specifically, the constraint formula ensures that each order is not split or repeatedly assigned throughout the entire scheduling cycle by summing the allocation variables of all vehicle sets with respect to orders and setting the result to always equal to 1.

[0025] Service Constraint 3: In intercity travel, the pick-up and drop-off points for any order can only appear on one and the same route, and the pick-up point in the departure city must appear before the drop-off point in the destination city. Specifically, the constraint formula stipulates that the time for a vehicle to reach the order's delivery point must be greater than the sum of its time to reach the pickup point and the necessary travel time between the two. This constraint, by establishing the temporal and spatial coupling relationship between pick-up and delivery nodes, ensures the continuity of order fulfillment, effectively avoids the logical risk of "reverse service," and thus ensures the physical rationality of vehicle scheduling schemes in intercity travel.

[0026] Service Constraint 4: All vehicles serve a location no more than once per trip. Specifically, this constraint formula sums the path variables of all arcs originating from a node, limiting the result to no greater than 1. This ensures that a vehicle visits each geographic node at most once per trip. Service Constraint 5: Ensure service continuity and path integrity for order nodes. Specifically, this constraint formula stipulates that the total flow entering a node equals the total flow leaving that node, and both values ​​are determined by the assignment variable. This constraint mandates that if an order is assigned to a vehicle, the vehicle must enter and exit the node corresponding to that order, thus ensuring that vehicles serving the origin station can reach their corresponding destination, guaranteeing the spatial continuity of the path scheme.

[0027] Time Constraint: To ensure the service quality of intercity ride-hailing services, the start time must meet the service time window. Specifically, the constraint formula stipulates that the vehicle's arrival time at the pick-up point must fall between the earliest and latest departure times. This constraint ensures on-time delivery and guarantees the service quality of intercity carpooling services. Safety Constraint: Considering that intercity travel takes longer than intra-city travel, to ensure the safety of passengers and drivers and to guarantee driver efficiency, all vehicles are required to take the minimum required rest period after reaching the maximum continuous driving time stipulated by the state. Specifically, the constraint formula stipulates that if the accumulated driving time triggers the rest variable, the vehicle's final arrival time at the destination must be included in the stipulated rest period.

[0028] S1: Obtain the set of intercity orders (Ord) to be processed, perform trip initialization processing on the intercity order set, and generate a trip set (Trips). Trips), and perform vehicle allocation initialization processing on the trip set Trips to generate an initial population (trips) vehicle); Specifically, step S1 further includes: obtaining the intercity order set Ord to be processed, sorting all orders in the intercity order set Ord in ascending order according to the earliest departure time, and selecting the first order after sorting as the basis to generate the itinerary node vector. Remove it from the intercity order set Ord; Add all remaining orders from the intercity order set Ord to the order queue. Meanwhile, the status of all orders in the order queue is marked as unprocessed; From the order queue Select an unprocessed order from the earliest selected order and attempt to insert it into the travel node vector. In the middle, and for the inserted run-length node vector Perform constraint checks to determine whether the constraints meet the constraint rules (constraint checks should at least include constraints on the maximum passenger capacity of the vehicle, passenger service time window constraints, and the order of pick-up and drop-off points). If the conditions are not met, remove the order from the trip node vector. Removed from the order queue. The system selects an unprocessed order based on the rule that it has not been selected before and is in the earliest order for insertion and constraint validation; If the conditions are met, update the order status to "processed"; Statistical travel node vector The total number of passengers in all orders is checked. If the total number of passengers is determined to be less than the vehicle's maximum capacity, the order queue is inspected. If any orders remain in the process, continue trying to insert a new order. If no orders exist, save the process node vector. And begin building a new itinerary; When it is determined that the total number of passengers has reached the vehicle's maximum capacity, the trip node vector is... Add to the Trips set, remove all orders contained in the trip from the Ord set, and clear the trip node vector. In preparation for the next leg of the journey; Check if the intercity order set Ord is empty. If it is, complete vehicle allocation initialization. If it is not empty, re-acquire the intercity order set to be processed and perform trip initialization, continuing to generate new trips. Specifically, if the total number of passengers in the current trip reaches the vehicle's passenger capacity limit, or if the constraints cannot be met after traversing the order queue, add the current vector to the trip set and clear the vector. Repeat the above process until the order set is empty.

[0029] The trip set Trips is sorted in ascending order by the earliest start time of each trip, and the first trip in the sorted set is selected as the basis for generating vehicle route vectors. And remove the trip from the Trips collection; Add all remaining trips from the Trips set to the trip queue. In the middle, and put the process queue All processes in the process are marked as unprocessed. From the trip queue Randomly select an unprocessed trip and attempt to add it to the vehicle path vector. In the process, the added vehicle path vector Perform constraint verification to determine whether it meets the constraint restriction rules; If the conditions are not met, the trip is removed from the trip set Trips and continues to be removed from the trip queue. In the middle, an unprocessed route is selected according to the rule that it has not been selected before and is the earliest in order for insertion and constraint verification; If the conditions are met, update the status of the trip to "processed"; Determine the process queue Are there any unprocessed processes in the queue? If so, continue processing them from the process queue. In the middle, an unprocessed route is selected according to the rule that it has not been selected before and is the earliest in order for insertion and constraint verification; If the vehicle does not exist, add the currently constructed vehicle to the vehicle set Car, remove all trips contained in that vehicle from the trip set Trips, and clear the vehicle's path vector. ; The process involves checking if the trip set `Trips` is empty. If empty, vehicle allocation initialization is completed, resulting in an initial population containing multiple vehicle sets `Car`. If not empty, the trip is removed from the `Trips` set, and vehicle allocation initialization is performed on the next vehicle, continuing the construction of trip combinations. Specifically, a trip is randomly selected from the set of trips to be processed and attempted to be incorporated into a vector, and its compliance with vehicle continuous driving time safety constraints and cross-city continuous scheduling constraints is verified. If satisfied, the trip is assigned to the vehicle; otherwise, if not satisfied or the traversal ends, the vector is stored in the vehicle set. This two-stage construction process is repeated until all trips are allocated. The final generated vehicle set constitutes an initial feasible path planning scheme, which is then updated in the external archive solution.

[0030] In this embodiment, to overcome the problem of generating a large number of infeasible solutions under strong constraints due to random initialization, this invention employs a time-series-based greedy construction method to generate an initial solution set. Specifically, the optimization objective and constraints are constructed, and an initial feasible solution set is generated using the time-series-based greedy construction method, which forms the initial population. At this point, all vehicles in the two vehicle sets together constitute a complete solution. In simple terms, the greedy construction steps first sort the order set according to the earliest departure time; then, iteratively select the order with the earliest departure time from the order set as the baseline to construct the itinerary; during the itinerary construction process, iterate through the remaining orders, and according to the greedy strategy, sequentially select the order with the earliest departure time in the current order set that still satisfies the constraints after insertion and insert it into the current itinerary until no further insertion is possible; finally, the generated itinerary is assigned to vehicles to form a complete path planning scheme, and this process is repeated until the number of generated schemes reaches the preset population size.

[0031] S2, iteratively optimizes the initial population, where computing resources are dynamically allocated based on the performance improvement of each generation on each objective, the current optimization objective function is selected, and an enhanced crossover operation is performed on the corresponding parent individuals according to the selected optimization objective function to generate a crossover scheme. A mutation operator is selected from the preset mutation operator pool, and a mutation operation is performed on the crossover scheme to generate a mutant scheme. The crossover scheme and the mutant scheme are compared, and the scheme with the smaller value of the optimization objective function is selected to be added to the offspring population. Specifically, step S2 further includes: […]. The potential of each generation of the population is monitored in real time across various target dimensions, and recorded. The set of optimal fitness values ​​of the population across four objective dimensions And compare it with the best value within the historical observation window. A comparison was conducted, among which, This represents the number of population iterations. At that time, the population was the initial population. Set the sliding window size (e.g., 5 or 10); calculate Improvement index of generation population The formula for quantifying the performance improvement of the m-th objective is: , M=4, where M is the number of objective functions to be optimized. for The set of optimal fitness values ​​of the population in the m-th objective dimension. for The set of optimal fitness values ​​of the population in the m-th objective dimension. To prevent division by zero errors, a (very small) positive constant is used. This index can accurately reflect the evolution slope of each objective function in the recent search process. Using an adaptive probability mapping mechanism Improvement index of generation population Convert to The probability of being selected in a generation Its formula is , It is an exponential function. for The first generation of the population The set of optimal fitness values ​​across each objective dimension; to enhance the algorithm's sensitivity to high-reward objectives and suppress noise interference, this embodiment employs the Softmax function to perform nonlinear mapping; The optimization objective function for the current iteration is determined from the target pool using a roulette wheel or random sampling mechanism. This mechanism ensures that if a target (such as "total waiting time") shows significant optimization potential in the near future, the algorithm will automatically increase its optimization frequency in subsequent crossover and mutation operations, thereby dynamically tilting computational resources towards "high-yield" targets. A learning matrix is ​​maintained in real time to store the cumulative improvement of each target within a window, and this is used as a feedback signal to correct the probability distribution in the next round. When a target enters an evolutionary plateau and its improvement exponent approaches zero, its selection probability will automatically decrease to a preset minimum threshold, thereby guiding the search focus to other dimensions that still have optimization potential (such as "vacancy rate" or "fleet size").

[0032] Please see Figure 5 , Figure 6 Two parent individuals are selected from the parent population corresponding to the current iteration's optimization objective function using a binary tournament selection method. A structured recombination process is then executed according to the selected optimization objective function to generate offspring schemes with superior genetic characteristics and strict spatiotemporal constraints. Among these, the "superior vehicle path" in the parent individuals is dynamically defined based on the currently activated optimization objective. If the current activation objective is to minimize the total number of empty seats ( Prioritize identifying and extracting the route vectors of vehicles with the highest passenger load rate among the parent vehicles; if the optimization objective is to minimize the fleet size ( If the objective is to minimize the total travel distance, then the vehicle route with the most transport orders will be prioritized. The system selects the vehicle routes with the shortest path length as the inheritance objects. During inheritance, these optimal route fragments are copied sequentially from the selected parent to the offspring individuals, with order uniqueness checks performed before each copy. If the objective is to minimize the total passenger waiting time... The system selects the vehicle routes with the shortest cumulative waiting time as the inheritance objects. During the inheritance operation, these superior route segments are copied sequentially from the selected parent to the offspring individuals, and an order uniqueness check is performed before each copy.

[0033] For the set of unassigned orders that have not entered the sub-scheme, a heuristic insertion algorithm associated with the target is used for processing. For each unassigned order, all vehicle paths in the sub-scheme are traversed to retrieve all possible insertion positions in each path, resulting in multiple candidate insertion points. For each candidate insertion point, the constraint detection module is called to verify in real time whether the constraint rules (i.e., hard constraints such as maximum vehicle passenger capacity, passenger service time window, and mandatory rest time for intercity long-distance driving) are still satisfied after inserting the order. If so, among all candidate insertion points that have passed the feasibility check, the insertion cost is calculated according to the currently selected optimization objective function, and the candidate insertion point that minimizes the objective increment is selected to perform the final insertion and generate the cross body scheme. When it is determined that no feasible position satisfying the constraints can be found after traversing all paths, a new virtual vehicle will be started for the order and an independent starting path will be assigned to generate a cross-body solution to ensure the completeness of the solution.

[0034] Obtain the target state parameters before and after the mutation operation, at the 1st In the iteration, for the first The process of performing a mutation operation on the m-th optimization objective function, recording the objective value before the mutation operation. and the target value after performing the mutation operation ; Calculate the first During the process of a mutation operation applied to the m-th optimization objective function, the percentage improvement of the objective function in a single iteration This value reflects the mutation operation. The immediate search contribution on target m; Calculate the average improvement contribution (i.e., the first statistical number) The total number of times each mutation operation is selected on the m-th objective, and the cumulative improvement ratio is averaged by the improvement ratio generated by each mutation operation divided by the total number of mutation operations. For the first The total number of times a mutation operation is selected on the m-th optimization objective function. For the first During the process of the m-th mutation operation acting on the m-th optimization objective function, the... The percentage of target improvement selected in a single iteration; Iterate through all sets of mutation operations, and calculate the ratio of the average improvement contribution of the current operator to the average improvement contribution of all operators to obtain the result. The standardized reward value of each mutation operation on the m-th optimization objective function This value serves as a feedback signal for the selection probability of subsequent update operators, characterizing the operator's flexible adaptive contribution to each objective. The number of mutation operations in the set (e.g., 8). Retrieve the empirical quality and immediate reward value of the current operator, and retrieve the [number]th [operator] from memory. The current empirical quality score of each mutation operation on the m-th optimization objective function; a moving average update of the operator's empirical quality is performed, wherein the empirical quality is updated based on the immediate reward value using a linear weighted moving average algorithm, the formula of which is: , For the updated experience quality rating, The experience quality score before the update. For fitness parameters, This parameter is used to balance the weight of historical experience quality and current immediate reward in the quality update process. Through this step, the immediate performance feedback generated by the mutation operation is integrated into the long-term quality assessment to smooth the random fluctuations in the search process. The updated operator quality score is output and stored, and the updated experience quality is stored in the operator quality database as the basic input data for the next stage of probability allocation or index calculation.

[0035] Based on the updated experience quality score, a random operator is selected from the preset strategies to determine the operator to be executed, and a probability distribution is generated. OR operator decision index value Among them, the preset strategies include probability matching strategy, adaptive pursuit strategy, and multi-armed slot machine strategy; Among them, the probability matching PM strategy calculates the selection probability of each operator. , The preset minimum selection probability ensures that each operator has a chance to be selected. ; Adaptive Pursuit (AP) strategy: Identify the operator with the highest current quality score. And adjust the probability according to the following formula , For learning rate, As the maximum probability cap, this operation aims to give the best-performing operator more selection opportunities; Multi-Armed Slot Machine (MAB) Strategy: Directly compute the decision index value for each operator. C is a balancing factor used to weigh "choosing the high-scoring operator" against "the operator with fewer attempts".

[0036] When the preset strategy is a probability matching strategy or an adaptive pursuit strategy, a mutation operator is randomly selected based on the probability distribution using the roulette wheel algorithm. When the preset strategy is a multi-armed slot machine strategy, the decision index value is directly selected. The largest mutation operator. Apply the selected operator and generate a new solution, then call the selected first mutation operator. Each mutation operator performs a mutation operation on individuals in the current population to generate the next generation of alternative solutions, and records the number of times the operator is executed, i.e., "number of executions = number of executions + 1".

[0037] In this embodiment, in each iteration, a target selection strategy is first executed to determine the optimization focus of the current computing resources. Specifically, before a preset termination condition is met, the following sub-steps are executed cyclically to perform multi-objective optimization on the population: First, target selection: using the target selection strategy, computing resources are dynamically allocated based on the performance improvement of the population on each target to select the optimization target for the current iteration; Second, strong crossover: based on the selected current optimization target, an enhanced crossover operation is performed on the selected parent individuals to generate a crossover scheme; Third, adaptive operator selection: using the operation selection strategy, a mutation operator is selected from a preset mutation operator pool to perform a mutation operation on the crossover scheme to generate a mutant scheme; Finally, individual update: the crossover scheme and the mutant scheme are compared, and the superior and unique scheme is selected to be added to the offspring population.

[0038] Specifically, target selection includes: calculating the improvement index of each optimized population within each preset evolutionary cycle. The index is determined based on the ratio of the difference between the optimal fitness value of the current generation and the optimal fitness value of previous generations, and is used to characterize the evolutionary potential and search efficiency of the corresponding target dimension. The improvement index of each target is converted into a selection probability using the Softmax function. This mapping mechanism assigns a higher selection probability to targets with higher improvement indices, thereby guiding the algorithm to allocate computational resources towards target dimensions with high evolutionary rewards. Based on the calculated selection probability distribution, the active target for the current iteration is determined from a preset target set (including minimizing the number of empty seats, minimizing the number of vehicles, minimizing the total travel distance, and minimizing the total waiting time) through random sampling or a round-robin mechanism. After each iteration, the lift record in the learning matrix is ​​updated in real time based on the quality feedback of the actual generated solutions, serving as a priori basis for the next round of target selection.

[0039] The enhanced crossover operation consists of two phases. The first phase uses a binary tournament method to select two parent individuals from the parent population. Based on the target selection strategy, it identifies valid travel or vehicle routes from the parents that can be directly inherited by the offspring, prioritizing the current optimization objective. (If the current optimization objective is to minimize the number of empty seats, high-load paths are prioritized; if it's to minimize travel distance, paths with compact routes and short ineffective detours are prioritized; if it's to minimize waiting time, paths with high time window fit are prioritized.) Subsequently, the identified potential paths are completely copied into the offspring scheme, with real-time uniqueness checks during the copying process to ensure that the same order is not repeatedly assigned to different vehicles, thus preserving high-quality path segments already generated in the parent generation that satisfy intercity spatiotemporal constraints. The second phase identifies the remaining set of unassigned orders that were not inherited in the first phase. Then, for each unassigned order, it iterates through all existing travel or vehicle route sequences in the offspring scheme to find the insertion position that satisfies all constraints and has the lowest cost. If no feasible insertion point can be found after traversing all existing path segments, a new vehicle path is created to accept the order. In simple terms, the first stage extracts the best from the parent paths; a binary tournament is used to select two parent individuals, and the most promising vehicle path from the better individuals is selected based on a preset probability to insert into the intersection; the most promising vehicle path refers to the path that performs best in the current optimization objective direction; the second stage inserts unassigned orders; for the unassigned orders remaining after the first stage, attempts are made to insert them into the optimal position of the vehicle in the intersection; if an existing vehicle cannot be inserted, a new vehicle is created for assignment; the optimal position is the position that, after insertion, optimizes the current optimization objective value and satisfies all constraints.

[0040] Diversity Mutation: The eight mutation operators described in this step (i.e., the operators in the mutation operator pool) include the following: Mutation Operator 1: Removes a single order node and reassigns it to another position. Mutation Operator 2: Swaps the positions of two customer nodes. Mutation Operator 3: Removes multiple customer nodes centrally and reconstructs their insertion positions as a whole. Mutation Operator 4: Performs pairwise position reorganization on multiple customer nodes to achieve structural adjustment. Mutation Operator 5: Disassembles a single trip and reinserts it into the feasible solution structure. Mutation Operator 6: Swaps the overall positions or contents of two trips. Mutation Operator 7: Removes multiple trips in batches and reconstructs their combination relationships. Mutation Operator 8: Rearranges and swaps the structures between multiple trips as a whole.

[0041] The adaptive operator selection process comprises three stages: The first stage is the performance evaluation of mutation operators. This stage measures the operator's contribution by monitoring changes in the target state before and after the mutation operation. Specifically, it involves recording the changes in the target value before and after the mutation operation in real time, and calculating the immediate improvement ratio of the operation to the current target. The total number of times each mutation operation is selected is counted, and the accumulated improvement ratios are averaged to obtain an index reflecting the average contribution level of the operators. Finally, a standardized reward value is obtained by comparing the average improvement ratio of a specific operator with the sum of the average improvement ratios of all candidate operators. This standardized reward value characterizes the operator's flexible adaptive contribution to different optimization objectives.

[0042] The second stage involves dynamic updating of the operator's empirical quality. This stage aims to balance the operator's historical performance with immediate feedback. Specifically, it includes retrieving the operator's current empirical quality score recorded in memory and the standardized reward value output from the first stage. A preset fitness parameter is introduced, and a linear weighted moving average logic is used to inject the immediate performance reward into the historical quality score. Through this moving average update process, the algorithm can smooth out the random fluctuations during the search process, transforming the operator's immediate performance into reliable long-term quality assessment data, and updating and storing it in the operator quality database.

[0043] The third stage involves mutation operator decision-making and execution. This stage achieves intelligent optimization based on operator quality scores, with the following specific steps: First, based on the updated empirical quality, a probability distribution or decision index for operator selection is established using one of the following strategies: probabilistic matching, adaptive pursuit, or multi-armed slot machine. The probabilistic matching strategy ensures that each operator has a basic selection opportunity; the adaptive pursuit strategy aims to strengthen the selection of the optimal operator; and the multi-armed slot machine strategy uses a decision index to balance the utilization of high-scoring operators with the exploration of low-frequency operators. Roulette wheel sampling is performed based on the probability distribution generated by the selected strategy, or the target operator is directly selected based on the maximum value of the decision index. The selected operator is then applied to perform mutation operations on individuals in the population to generate new generation candidate solutions, and the cumulative execution count of the operator is updated simultaneously.

[0044] Please see Figure 7 It should be noted that this invention designs eight mutation operators specifically tailored to the characteristics of intercity travel problems for local search optimization of solutions. Given that the vehicle routing problem for intercity travel is essentially a multi-objective optimization problem, and multi-objective optimization is one of the core problems in optimization, different objectives (such as minimizing the number of empty seats, minimizing the number of vehicles, minimizing the travel distance, and minimizing the waiting time) have different characteristics in terms of physical meaning, scale, and optimization difficulty. Therefore, for some objectives that are difficult to optimize using conventional neighborhood search operations, it is necessary to design neighborhood search operations specifically tailored to the characteristics of intercity travel. The eight mutation operators and their execution logic are as follows: Mutation Operator 1 (Single Order Relocation): Randomly selects a boarding order in the current path and deletes it. Then, it globally searches for the boarding order location that satisfies the constraints and minimizes the increase in the objective function. The necessity of this operator lies in fine-tuning the service order through precise displacement at the micro level without changing the overall path framework. Subsequently, the TerminalOrderOptimizer function is called to optimize the new disembarkation sequence using polar angle calculation, solving the problem of high flexibility but lack of time-series guidance in intercity last-mile delivery, thereby effectively reducing travel distance.

[0045] Mutation Operator 2 (Single Order Point Swap): The system randomly selects an order point on the path and attempts to swap it with other path points, while simultaneously migrating its associated drop-off point to the target path. The specific location of the drop-off point in the path after the swap is dynamically rearranged using the drop-off order optimization function described in Mutation Operator 1, and this is recorded... The min_delta value is selected to maximize the target improvement. This operator aims to directly resolve load imbalances between paths and time window conflicts of individual orders through point-to-point swapping, and is a fundamental tool for improving convergence accuracy.

[0046] Mutation Operator 3 (Multi-Order Point Relocation): Identifies and extracts a sequence containing multiple consecutive order points from the selected path, and attempts to embed it as a whole into other feasible locations. During this process, the system synchronously verifies the logical integrity of the orders within the sequence, ensuring that the temporal logic of the boarding and alighting points of the extracted orders remains consistent with the path constraints when they are re-inserted at the target location. Its core value lies in adjusting the distribution of multiple related orders at once through block-level migration, thereby effectively resolving the local search bottleneck that cannot be overcome by single-point fine-tuning.

[0047] Mutation Operator 4 (Two-opt): This mutation operator selects a cut-off point on each of two different travel paths and swaps all order segments following that cut-off point on both paths. Specifically, the system first randomly selects a location on each of the first and second paths as the starting point for the swap. Then, it connects all order sequences following that point on the first path to the cut-off point on the second path, while simultaneously connecting the remaining segments from the original second path to the first path. This operation achieves rapid adjustment of the vehicle's travel route framework by recombining the latter half of the tasks on the two paths. When processing a large number of intercity orders, this operation can select the lowest-cost solution by trying different combinations of path segments, thereby effectively reducing the empty driving distance of vehicles traveling between different areas without affecting passengers' normal travel.

[0048] Mutation Operator 5 (Single-Stroke Relocation): This operator is mainly used to process inefficient segments of vehicle travel paths, optimizing the overall solution by breaking down the journey. The specific implementation steps are as follows: Step 1, Journey Selection: The system first randomly selects a journey to be processed from the current scheduling plan (denoted as...). The itinerary includes several customer order points requiring service. The second step is task distribution and transfer: the system attempts to distribute the selected itinerary... For each customer point in the process, the system attempts to insert it into the driving paths of other vehicles. During insertion, the system searches all available vehicle paths and verifies whether the target path still meets the loading and time requirements after adding the point. The third step is integrity judgment: if the trip... If all customer points can be successfully assigned to other paths, the trip becomes an empty path and proceeds to the next verification step; if any customer point cannot find an acceptable other path, the splitting is deemed a failure, and the system restores all points to the original trip. Fourth step, trip association verification: When the trip... After successfully becoming an empty path, the system checks if the trip is the last trip for that vehicle. If it is the last trip, the empty trip is directly deleted from the vehicle's task list. If it is not the last trip, the system further evaluates its immediately following trip (denoted as...). The system attempted to use the same method on... The process involves dismantling and transferring the trips. Only when all associated trips are successfully cleared will the system officially confirm the deletion of these trips. Fifth step, plan update: After confirming successful trip dismantling, the system updates the driving routes and task sequences for all affected vehicles.

[0049] Mutation Operator Six (Single-Trip Exchange): This operator treats the trip as the smallest logical trigger unit and optimizes the global cost by reallocating task blocks among different transportation resources. The specific implementation steps are as follows: First, object selection: Randomly select two different target vehicles from the current vehicle set, denoted as vehicle V1 and vehicle V2 respectively. Then, randomly select a complete trip T from the travel path of vehicle V1. a At the same time, another complete journey T is selected from the driving path of vehicle V2. b The second step is atomization interchange: stroke T a With itinerary T b Tasks considered indivisible are swapped in execution order. That is, the task sequence T originally belonging to V1 is... a The entire task is assigned to V2, while the task sequence T that originally belonged to V2 is transferred to V2. b The entire process is assigned to V1. The third step involves timing constraints: during the swapping process, the operator maintains its travel distance T. a and T bThe internally scheduled order order sequence and time points remain completely unchanged. This "unaltered" migration ensures that the already optimized tight schedule remains effective after the cross-vehicle transfer, avoiding secondary calculations at the micro level. The fourth step is feasibility verification and cost assessment: After the exchange, the system will perform constraint checks on the new routes of V1 and V2, including but not limited to verifying the maximum passenger capacity of vehicles and the connection time between trips. If the constraints are met, the system will calculate the change in total transportation costs before and after the exchange. Step 5: Optimal Result Selection: If this itinerary swap results in a lower total cost solution (i.e....) If so, the mutation operation will be officially accepted and the vehicle's mission list will be updated.

[0050] Mutation Operator 7 (Vehicle Merging and Relocation): This mutation operator directly reduces the total number of vehicles performing tasks by performing secondary integration of transportation resources, thereby reducing operating costs. The specific implementation steps are as follows: Step 1, Low-load vehicle identification: The operator traverses all currently running vehicles, sorts them according to the number of trips each vehicle undertakes and its task saturation, and prioritizes selecting the vehicles with the fewest tasks, marking them as vehicles to be merged. The second step is batch migration of trip routes: operators extract vehicles. All the travel modules undertaken. For each travel T within it... i The operator attempts to find a location among the remaining active vehicles that can receive the trip. The third step is feasibility and constraint verification: This involves attempting to route T... i When inserting other vehicle routes, the operator strictly calculates the remaining passenger capacity of the target receiving vehicle, the maximum operating time limit, and the transfer connection time between routes. Only when the vehicle... The operator determines the merge operation is feasible only when all trips under the vehicle's name can be successfully received by other vehicles. The fourth step is the formal deregistration of the vehicle: After confirming the successful transfer of all trips, the operator will deregister the vehicle. The vehicle and its corresponding empty path are completely removed from the current task list, and the vehicle's running status is deregistered. Fifth step, target optimization confirmation: After vehicle deregistration, the operator recalculates the overall vehicle fixed cost and total operating cost. This operator is designed to directly address the core objective of "minimizing fleet size." In complex intercity operation scenarios, due to uneven order distribution, some vehicles often have extremely low loads. This forced vehicle merging method maximizes the resource utilization of individual vehicles, reduces vehicle idling, and is the most direct and effective means of reducing overall system operating costs.

[0051] Mutation Operator 8 (Return Trip Switching): This operator is specifically designed for the characteristic of intercity travel where vehicles need to travel back and forth between cities. It optimizes resource allocation by re-matching the return trip tasks of vehicles. The specific implementation steps are as follows: Step 1, Return Trip Segment Identification: The system randomly selects two vehicles performing cross-city tasks from the current vehicle set, denoted as vehicle V1 and vehicle V2 respectively. Based on the driving trajectory of each vehicle, the system identifies the task sequence in its path that returns from the destination city to the origin city, and marks them as return trip segments. and The second step is task sequence extraction: the system extracts the sequence from the V1 path. All order points and V2 paths belonging to All order points are extracted separately. During this process, the system maintains the order sequence within each return segment without changing it. The third step is cross-vehicle swap execution: the system performs a swap operation on the return tasks, that is, swaps the return task sequences originally belonging to V1. The entire process is assigned to vehicle V2, and the return sequence originally belonging to V2 is also removed. Assigned to vehicles as a whole The fourth step, connection constraints and load verification: After the exchange, the system will logically verify the complete paths of the two vehicles (including the first half of the outbound journey and the newly connected return journey). The key verifications include whether the transfer time between the outbound endpoint and the new return start point is sufficient, and whether the passenger capacity during the return journey exceeds the vehicle's rated load limit. The fifth step, scheme optimization and update: The system calculates the total global mileage after the exchange operation. If this re-matching effectively reduces the empty-run rate or increases the vehicle's return load rate, the vehicle's task allocation scheme will be officially updated.

[0052] The necessity of this operator lies in the fact that intercity travel typically involves long-distance round trips between cities, and there is often an imbalance in vehicle demand between the two cities (for example, there are many orders from city A to city B in the morning, but few return orders). This operation can solve the problem of empty return trips caused by the mismatch in demand between the two cities by flexibly reorganizing the tasks of the vehicles in the "second half" of the journey, without affecting the service quality of the first half of the journey. This balances the vehicle load throughout the entire cycle and improves overall operational efficiency.

[0053] S3 merges the current iteration's offspring population with the historical reserve, updates the external archive according to the dominance determination rule, and outputs the non-dominated solution set in the external archive as the final intercity travel vehicle route planning scheme when the preset maximum running time is reached.

[0054] Specifically, step S3 further includes: determining the target space scaling standard, and for each optimization objective function, setting a positive step size based on the actual solution accuracy requirements. This parameter determines the minimum resolution of the final solution set in the target space; Perform a discretization mapping of the objective function to obtain the individual solutions to be evaluated. The original target vector According to the formula Calculate its corresponding gridded coordinate vector ,in, To solve individual In the m-th optimization objective function, To solve individual The gridded coordinates on the dimension of the m-th optimization objective function This is the floor operator, which converts a continuous floating-point target into a discrete integer index value. definition Dominance determination rules, specifically including: for any two solution individuals and If the condition is met: for all targets, their grid coordinates satisfy... And there exists at least one target whose grid coordinates satisfy Then determine the individual solution. of Occupying the superiority of individual solutions (recorded as) M represents the number of objective functions to be optimized; (Perform grid-based deduplication and filtering. When updating the external archive, based on the best-case decision rule) when it is determined that the new solution and an existing solution in the archive are in different grids, The dominant rule determines its survival. When it is determined that a new solution is in the same grid as an existing solution in the archive (i.e., the grid coordinates are completely consistent), only the individual with the closest Euclidean distance to the origin is retained, and the rest are discarded.

[0055] In this embodiment, traditional Pareto dominance is overly sensitive when comparing two solutions; if one objective value is slightly better, the two solutions cease to dominate each other. This leads to an overcrowded solution set in certain regions, and its size becomes difficult to control. In contrast, Dominance is achieved by introducing a scaling factor The target space is divided into countless "small squares." Solutions within the same square are considered "equal," thus achieving forced spacing control. Therefore, this method merges the current iteration's individual solutions with the historical reserve, and based on... Dominance relationships and redundancy elimination rules within the same grid enable dynamic screening, scale control, and coordinated optimization of the spatial distribution uniformity of the globally optimal non-dominated solution.

[0056] Specifically, the process begins with the construction of a discretized grid in the target space. For the objective functions of each dimension of the multi-objective optimization problem, a positive scale factor is set. This scale factor is then used to scale and round down the original objective values ​​of individuals, mapping the continuous target space to discrete grid coordinate vectors and establishing a resolution benchmark for the target space. Next, a dominance relationship is determined based on the grid coordinates. For any individual in the candidate solution set, its coordinate vector in the discrete grid space is compared. If the first individual's grid coordinates are not inferior to the second individual's in all target dimensions, and it is superior to the second individual in at least one target dimension, then the first individual is determined to constitute a dominance constraint on the second individual. Subsequently, redundant individuals within the same grid are eliminated. During the update of the external archive, all sets of individuals falling into the same grid coordinates are traversed, and the Euclidean distance between each individual and the ideal point in the target space is calculated. Within the same grid, only the representative individual with the smallest Euclidean distance is retained, and the remaining redundant individuals are eliminated, thus achieving automatic adjustment of the archive size. Finally, the external archive is dynamically updated and output; based on the dominance determination result and deduplication rules, dominated individuals and redundant individuals are removed, and the external archive set is updated; the grid constraint mechanism is used to force the uniformity of the distribution of the solution set on the Pareto front, and the final non-dominated solution set is output.

[0057] When the preset maximum running time is reached, the non-dominated solution set in the external archive is output as the final intercity travel vehicle route planning scheme.

[0058] Specifically, in this embodiment, to comprehensively and objectively verify the effectiveness and robustness of the method in solving different types of problems, two test case libraries with different dimensions were constructed: one is an improved test case library based on the classic Solomon benchmark, and the other is an intercity travel test case library based on real operational data. The improved test case library based on Solomon (36 cases) was reconstructed in this embodiment to verify the algorithm's performance in standardized scenarios, using the classic Solomon VRP benchmark as a blueprint and incorporating the long-distance characteristics of intercity travel. The data sources were selected from Solomon cases: C-class (customer point clustering distribution), R-class (random distribution), and RC-class (mixed distribution) data. The modification method was to maintain the original number of customer points (100), and to map the original short-distance coordinates to long-distance intercity coordinates by randomly increasing the latitude and longitude range; at the same time, the originally tight time windows were widened proportionally to simulate the relatively loose "soft time window" characteristics of intercity travel. Case composition: A total of 36 standardized test cases were generated by combining different travel demand ratios.

[0059] A library of 45 intercity travel case studies based on real operational data: To verify the algorithm's performance in complex real-world operating environments, this embodiment uses actual operational data (round trips between city A and city B) from a ride-hailing platform in city A. Data characteristics: All coordinates are based on real geographical locations, and travel distances and times are strictly calculated based on the actual road network. Case study composition: By cross-combining three key dimensions—"order size" (20 to 600 orders), "tidal ratio" (1:1 to 1:5), and "time window tightness" (30 to 60 minutes)—45 highly realistic scenario case studies were generated.

[0060] To verify the algorithm's performance, this example selects three classic algorithms in the field of multi-objective optimization as comparison objects: non-dominated sorting genetic algorithm II, decomposition-based multi-objective evolutionary algorithm, and multi-directional local search algorithm. These represent the mainstream population evolution algorithm, the decomposition-based algorithm, and the improved neighborhood search algorithm, respectively. To ensure the fairness of the comparison, the maximum running time is used as the termination condition for all algorithms. The runtime for the example was set to 100 seconds; for real-world examples, the runtime varied from 30 seconds to 2007 seconds depending on the scale. Each algorithm was run independently 30 times on each example.

[0061] This example uses two metrics, HV and IGD, for evaluation. A higher HV value and a lower IGD value indicate better convergence and diversity of the solution set. Figure 5 As shown. The statistical results of the method of this invention, obtained through the Wilcoxon signed-rank test, are as follows: (1) In 36 Performance improvements on simulations: Compared to MDLS: This invention significantly outperforms the IGD metric on 34 simulations and the HV metric on 32 simulations. Compared to MOEA / D: This invention significantly outperforms the IGD metric on 34 simulations and the HV metric on 31 simulations. Compared to NSGA-II: This invention significantly outperforms the IGD metric on 34 simulations and the HV metric on all 36 simulations. This demonstrates that, within a standardized theoretical model, the diversity enhancement mechanism of this invention can effectively overcome the bottlenecks of traditional algorithms.

[0062] (2) Performance on 45 real-world scenario examples: Compared with MDLS: This invention achieves significantly better results in both IGD and HV metrics across all 45 examples (100% win rate). Compared with MOEA / D: This invention outperforms IGD metrics on 43 examples and HV metrics on 44 examples. Compared with NSGA-II: This invention outperforms IGD metrics on 33 examples and HV metrics on all 45 examples.

[0063] Combining the results of the two sets of experiments, it is evident that thanks to the "greedy initialization based on time series" providing a high-quality search starting point, and the "enhanced crossover" and "adaptive mutation" mechanisms effectively breaking local convergence, the multi-objective vehicle path planning method based on diversity enhancement mechanisms proposed in this invention not only performs excellently in idealized standard examples but also demonstrates overwhelming advantages in complex real-world operational scenarios. This method can provide intercity travel platforms with a scheduling solution that offers broad coverage, low cost, and high service quality, possessing extremely high practical application value.

[0064] In summary, this invention addresses the multi-objective vehicle routing problem for intercity travel with four objectives, innovatively combining a diversity enhancement mechanism with a multi-objective evolutionary algorithm. This method, considering the long distances, multiple constraints, and multi-objective conflicts inherent in intercity travel scenarios, achieves a balance between convergence and diversity by introducing mechanisms such as time-series-based greedy initialization, dynamic objective selection, enhanced crossover, and adaptive diversity mutation.

[0065] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A multi-objective vehicle route planning method for intercity travel based on a diversity enhancement mechanism, characterized in that, The problem is addressed using a pre-built multi-objective vehicle routing problem model for intercity travel, and the steps include: Obtain the set of intercity orders to be processed (Ord), perform trip initialization processing on the intercity order set to generate a trip set (Trips), and perform vehicle allocation initialization processing on the trip set (Trips) to generate an initial population; The initial population is iteratively optimized. Specifically, computing resources are dynamically allocated based on the performance improvement of each generation of the population on each objective. The optimization objective function for the current iteration is selected. Based on the selected optimization objective function, an enhanced crossover operation is performed on the corresponding parent individuals to generate a crossover scheme. A mutation operator is selected from a preset mutation operator pool to perform a mutation operation on the crossover scheme to generate a mutant scheme. The crossover scheme and the mutant scheme are compared, and the scheme with the smaller value of the optimization objective function is selected to be added to the offspring population. Merge the current iteration's offspring population with the historical reserve, update the external archive according to the dominance determination rule, and when the preset maximum running time is reached, output the non-dominated solution set in the external archive as the final intercity travel vehicle route planning scheme.

2. The intercity travel multi-objective vehicle route planning method based on diversity enhancement mechanism according to claim 1, characterized in that, The optimization objective functions of the multi-objective vehicle routing problem model for intercity travel include: minimizing the number of empty seats, minimizing the number of vehicles, minimizing the total travel distance, and minimizing the total passenger waiting time. Their formulas are as follows: , , , ; in, To optimize the objective function by minimizing the number of empty seats in a vehicle, To optimize the objective function by minimizing the number of vehicles, To optimize the objective function by minimizing the total travel distance, To optimize the objective function by minimizing the total passenger waiting time, K is the set of all vehicles, and k is the vehicle index. Let Q be the road segment between node i and node j, Q be the maximum passenger capacity of a vehicle, and A be the set of all road segments. This represents the path variable used to determine whether vehicle k travels directly from node i to node j. For vehicle k on the road section Real-time load on Let k be the activation variable for vehicle k. For road section Let P be the distance between all orders and p be the order index of the current trip. The service start time for order p for vehicle k. Let p be the earliest expected departure time. Assign a variable to the order p of vehicle k.

3. The intercity travel multi-objective vehicle route planning method based on diversity enhancement mechanism according to claim 2, characterized in that, The output of the multi-objective vehicle routing problem model for intercity travel satisfies the constraint rules, and its formula is as follows: , , , , , , , ; in, Let p be the number of passengers. For vehicle k, the set of orders for the current journey. For vehicle k, the set of orders for the next trip. For order indexing of the next trip, Let k be the time when vehicle k arrives at the drop-off point of order p. For order collection The latest time for all passengers to drop off the bus. Let this be the drop-off point for order p. Let k be the time it takes for vehicle k to arrive at the next pickup point. For order collection The earliest passenger's pickup and boarding time. The travel time from the pick-up point to the drop-off point for order p. Let this be the pick-up point for order p. This indicates whether vehicle k travels directly from node i to the boarding point. Path variables, This indicates determining whether vehicle k has disembarked from the designated drop-off point. The path variable that leads directly to node j. Let p be the expected latest departure time. Let be the time when vehicle k finally arrives at its destination, and n be the total number of orders in the set P of all orders. Let K be the cumulative driving time of vehicle k. For rest duration, Let k be the resting variable for vehicle k.

4. The intercity travel multi-objective vehicle route planning method based on diversity enhancement mechanism according to claim 3, characterized in that, Obtain the set of intercity orders (Ord) to be processed, perform trip initialization on the intercity order set, and generate a trip set (Trips), specifically: Obtain the set of intercity orders (Ord) to be processed. Sort all orders in the set of intercity orders (Ord) in ascending order according to their earliest departure time, and select the first order in the sorted set as the reference to generate the itinerary node vector. Remove it from the intercity order set Ord; Add all remaining orders from the intercity order set Ord to the order queue. Meanwhile, the status of all orders in the order queue is marked as unprocessed; From the order queue Select an unprocessed order from the earliest selected order and attempt to insert it into the travel node vector. In the middle, and for the inserted run-length node vector Perform constraint verification to determine whether it meets the constraint restriction rules; If the conditions are not met, remove the order from the trip node vector. Removed from the order queue. The system selects an unprocessed order based on the rule that it has not been selected before and is in the earliest order for insertion and constraint validation; If the conditions are met, update the order status to "processed"; Statistical travel node vector The total number of passengers in all orders is checked. If the total number of passengers is determined to be less than the vehicle's maximum capacity, the order queue is inspected. If any orders remain in the process, continue trying to insert a new order. If no orders exist, save the process node vector. And begin building a new itinerary; When it is determined that the total number of passengers has reached the vehicle's maximum capacity, the trip node vector is... Add to the Trips set, remove all orders contained in the trip from the Ord set, and clear the trip node vector. In preparation for the next leg of the journey; Check if the intercity order set Ord is empty. If it is empty, complete the vehicle allocation initialization. If it is not empty, re-acquire the intercity order set to be processed, perform trip initialization processing, and continue to generate new trips.

5. The intercity travel multi-objective vehicle route planning method based on diversity enhancement mechanism according to claim 1, characterized in that, The trip set Trips is initialized with vehicle allocation to generate an initial population, specifically as follows: The trip set Trips is sorted in ascending order by the earliest start time of each trip, and the first trip in the sorted set is selected as the basis for generating vehicle route vectors. And remove the trip from the Trips collection; Add all remaining trips from the Trips set to the trip queue. In the middle, and put the trip queue All processes in the process are marked as unprocessed. From the trip queue Randomly select an unprocessed trip and attempt to add it to the vehicle path vector. In the process, the added vehicle path vector Perform constraint verification to determine whether it meets the constraint restriction rules; If the conditions are not met, the trip is removed from the trip set Trips and continues to be removed from the trip queue. In the middle, an unprocessed route is selected according to the rule that it has not been selected before and is the earliest in order for insertion and constraint verification; If the conditions are met, update the status of the trip to "processed"; Determine the process queue Are there any unprocessed processes in the queue? If so, continue processing them from the process queue. In the middle, an unprocessed route is selected according to the rule that it has not been selected before and is the earliest in order for insertion and constraint verification; If the vehicle does not exist, add the currently constructed vehicle to the vehicle set Car, remove all trips contained in that vehicle from the trip set Trips, and clear the vehicle's path vector. ; Check if the trip set Trips is empty. If it is empty, complete the vehicle allocation initialization and obtain the initial population, which contains multiple vehicle sets Car. If it is not empty, remove the trip from the trip set Trips, perform vehicle allocation initialization for the next car, and continue to build trip combinations.

6. The intercity travel multi-objective vehicle route planning method based on diversity enhancement mechanism according to claim 1, characterized in that, Computational resources are dynamically allocated based on the performance improvement of a certain generation of the population on each objective, and the optimization objective function for the current iteration is selected, specifically as follows: right The potential of each generation of the population is monitored in real time across various target dimensions, and recorded. The set of optimal fitness values ​​of the population across four objective dimensions and compare it with the best value within the historical observation window. A comparison was conducted, in which, This represents the number of population iterations. At that time, the population was the initial population. To adjust the sliding window size; calculate Improvement index of generation population The formula for quantifying the performance improvement of the m-th objective is: , M=4, where M is the number of objective functions to be optimized. for The set of optimal fitness values ​​of the population in the m-th objective dimension. for The set of optimal fitness values ​​of the population in the m-th objective dimension. Positive constants to prevent division by zero errors; Using an adaptive probability mapping mechanism Improvement index of generation population Convert to The probability of being selected in a generation Its formula is , It is an exponential function. for The first generation of the population The set of optimal fitness values ​​across each objective dimension; The objective function for the current iteration is determined from the objective pool using a roulette wheel or random sampling mechanism.

7. The intercity travel multi-objective vehicle route planning method based on diversity enhancement mechanism according to claim 1, characterized in that, Based on the selected optimization objective function, an enhanced crossover operation is performed on the corresponding parent individuals to generate a crossover scheme, specifically: Two parent individuals are selected from the parent population corresponding to the current iteration's optimization objective function using a binary tournament selection method. A structured recombination process is then executed based on the selected optimization objective function to generate offspring schemes with excellent genetic characteristics and strict spatiotemporal constraints. For the set of unassigned orders that have not entered the sub-scheme, a heuristic insertion algorithm associated with the target is used for processing. For each unassigned order, all vehicle paths in the sub-scheme are traversed to retrieve all possible insertion positions in each path, resulting in multiple candidate insertion points. For each candidate insertion point, the constraint detection module is called to verify in real time whether the constraint rules are still met after inserting the order. If so, among all candidate insertion points that have passed the feasibility check, the insertion cost is calculated according to the currently selected optimization objective function, and the candidate insertion point that minimizes the objective increment is selected to perform the final insertion and generate the cross body scheme. When it is determined that no feasible position satisfying the constraints can be found after traversing all paths, a new virtual vehicle will be started for the order and an independent starting path will be assigned to generate a cross-body solution.

8. The intercity travel multi-objective vehicle route planning method based on diversity enhancement mechanism according to claim 6, characterized in that, Select a mutation operator from the preset mutation operator pool, specifically: Obtain the target state parameters before and after the mutation operation, at the... In the next iteration, for the first The process of performing a mutation operation on the m-th optimization objective function, recording the objective value before the mutation operation. and the target value after performing the mutation operation ; Calculate the first During the process of a mutation operation applied to the m-th optimization objective function, the percentage improvement of the objective function in a single iteration ; Calculate the average improvement contribution , For the first The total number of times a mutation operation is selected on the m-th optimization objective function. For the first During the process of the m-th mutation operation acting on the m-th optimization objective function, the... The percentage of target improvement selected in a single iteration; Iterate through all sets of mutation operations, and calculate the ratio of the average improvement contribution of the current operator to the average improvement contribution of all operators to obtain the result. The standardized reward value of each mutation operation on the m-th optimization objective function , The number of mutation operations in the set; The moving average update of the operator's experience quality is performed, where a linear weighted moving average algorithm is used to update the experience quality based on the immediate reward value. The formula is as follows: , For the updated experience quality rating, The experience quality score before the update. For fitness rate parameters; Based on the updated experience quality score, a random operator is selected from the preset strategies to determine the operator to be executed, and a probability distribution is generated. OR operator decision index value Among them, the preset strategies include probability matching strategy, adaptive pursuit strategy, and multi-armed slot machine strategy; When the preset strategy is a probability matching strategy or an adaptive pursuit strategy, a mutation operator is randomly selected based on the probability distribution using the roulette wheel algorithm. When the preset strategy is a multi-armed slot machine strategy, the decision index value is directly selected. The largest mutation operator.

9. The intercity travel multi-objective vehicle route planning method based on diversity enhancement mechanism according to claim 1, characterized in that, Merge the current iteration's offspring population with the historical archive, and update the external archive according to the dominance determination rule, specifically: Determine the scaling criterion for the target space, and for each optimization objective function, set a positive step size based on the actual solution accuracy requirements. ; Perform a discretization mapping of the objective function to obtain the individual solutions to be evaluated. The original target vector, according to the formula Calculate its corresponding gridded coordinate vector, where, To solve individual In the m-th optimization objective function, To solve individual The gridded coordinates on the dimension of the m-th optimization objective function This is the floor operator; definition Dominance determination rules, specifically including: for any two solution individuals and If the condition is met: for all targets, their grid coordinates satisfy... And there exists at least one target whose grid coordinates satisfy Then determine the individual solution. of Occupying the superiority of individual solutions M is the number of objective functions to be optimized; When it is determined that the new solution and an existing solution in the archive are in different grids. The dominant rule determines its survival. When a new solution is found to be in the same grid as an existing solution in the archive, only the individual with the closest Euclidean distance to the origin is retained, and the rest are discarded.

Images