End distribution network load balancing method and system based on swarm intelligence algorithm

Abstract

The invention discloses a terminal distribution network load balancing method and system based on a swarm intelligence algorithm, and relates to the field of intelligent logistics, and the method comprises the steps: dividing a distribution region into a plurality of sub-regions through an improved clustering algorithm; a load balancing model with capacity constraint is adopted, and an improved ant colony algorithm is used for order distribution; designing a pheromone updating rule fusing the load information; establishing a dynamic load feedback mechanism; and a multi-target Pareto optimal solution set screening mechanism is adopted. Compared with a traditional method, the method has the advantages that the distribution efficiency is maintained, the load balance degree of the system is improved, the resource utilization rate is increased, the redistribution response time in a dynamic scene is shortened, the problems of resource idleness and overload caused by non-uniform load in tail end distribution are effectively solved, and the throughput and stability of the distribution system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics, and in particular to a load balancing method and system for last-mile delivery networks based on swarm intelligence algorithms. Background Technology

[0002] With the rapid development of e-commerce and on-demand delivery services, last-mile delivery systems face multiple challenges, including large order volumes, dispersed resources, and complex road conditions. Traditional delivery optimization methods are mostly based on Euclidean distance or static path planning, which are difficult to adapt to dynamically changing delivery environments. Existing ant colony algorithms are mostly used for path optimization, lacking direct modeling of resource capacity constraints and load balancing, and cannot effectively respond to real-time state changes. Therefore, there is an urgent need for a last-mile delivery load balancing method that can balance efficiency, load balancing, and dynamic adaptability. Summary of the Invention

[0003] To address the technical problems of delivery load balancing in existing technologies, this invention provides a method and system for load balancing in last-mile delivery networks based on swarm intelligence algorithms.

[0004] This invention is achieved through the following technical solution: A load balancing method for last-mile delivery networks based on swarm intelligence algorithms includes: S1: Construct a cluster-based spatial partitioning model, including preprocessing the last-mile delivery network and dividing large delivery areas into multiple sub-regions; S2: Construct a load balancing model with capacity constraints, including building a refined load balancing model within each sub-region and solving it using an improved ant colony model; S3: Design pheromone update rules that integrate load information, including introducing load balance index as the basis for pheromone updates and introducing forward-looking pheromone updates based on load trends. S4: Establish a closed-loop dynamic load feedback mechanism, including continuous monitoring of the real-time load and road conditions of all delivery resources; trigger dynamic adjustments when a triggering event is detected. S5: Establish a multi-objective Pareto optimal solution set selection mechanism, including using Pareto to make optimal solution set decisions for candidate solutions generated by the ant colony algorithm.

[0005] Furthermore, the clustering-based spatial partitioning model employs an improved K-Means++ clustering algorithm, which introduces the actual distance of the road network as a similarity metric to obtain a composite distance based on the traditional Euclidean distance.

[0006] Furthermore, the capacity constraint of the load balancing model with capacity constraints includes the maximum load capacity of each delivery resource within the sub-region. The capacity is a multi-dimensional constraint, including weight and / or volume capacity, daily effective working hours limit, order capacity, and the maximum number of orders that can be processed per day.

[0007] Furthermore, the load balancing index is introduced as the basis for pheromone updates, and the pheromone increment expression is as follows: Δτ ij =Q / (L k (1+b F balance (k)), where Δτ ij To place orders i Allocate to resource r j "The incremental pheromone reward given by this specific decision, F" balance (k) is the load balance degree of the allocation scheme constructed by the kth ant, and b is the load balance influencing factor.

[0008] Furthermore, the forward-looking pheromone update based on load trends includes calculating the load trend and modifying the heuristic information according to the trend, wherein the heuristic information η ij Revised to: in, To place orders i Allocate to resource r j The resulting costs, As the load impact factor, For resource r j Real-time load rate, As an indicator function, when the load trend slope s of resource j j The value is 1 when the positive threshold θ is exceeded, and 0 otherwise. λ is the trend penalty factor.

[0009] Furthermore, the triggering events for the dynamic load feedback include new tasks, resource overload, resource failure, or environmental degradation, and the dynamic adjustment methods include local reallocation and pheromone local reset.

[0010] Furthermore, the local redistribution includes locking the problem area affected by the event, using the current system state as the initial condition, and calling the improved ant colony algorithm described in S2 to perform a round of rapid iteration.

[0011] Furthermore, the pheromone local reset includes, for state deterioration trigger events, enhancing the local evaporation of pheromones related to the affected resources; for all trigger events: when performing local re-optimization, temporarily increasing the global pheromone evaporation coefficient and introducing a minimum pheromone value to prevent pheromone depletion.

[0012] This invention also provides a load balancing system for last-mile delivery networks based on swarm intelligence algorithms, which, based on the load balancing method for last-mile delivery networks based on swarm intelligence algorithms described above, includes: A clustering-based spatial partitioning model construction module is used to preprocess the last-mile delivery network, dividing the vast delivery area into more manageable sub-regions. A load balancing model building module with capacity constraints is used to build a refined load balancing model within each sub-region and solve it using an improved ant colony model. The pheromone update rule design module integrates load information, which is used to introduce load balance index as the basis for pheromone updates and introduce forward-looking pheromone updates based on load trends. The dynamic load feedback mechanism module is used to establish a closed-loop dynamic load feedback system. The system continuously monitors the real-time load and road conditions of all delivery resources; when a trigger event is detected, dynamic adjustments are triggered. The module for establishing a multi-objective Pareto optimal solution set selection mechanism is used to make optimal solution set decisions for candidate solutions generated by the ant colony algorithm using Pareto.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing program instructions for a load balancing method for last-mile delivery networks based on a swarm intelligence algorithm. The program instructions for the load balancing method for last-mile delivery networks based on a swarm intelligence algorithm can be executed by one or more processors to implement the steps of the load balancing method for last-mile delivery networks based on a swarm intelligence algorithm as described above.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: By employing composite distance clustering, the system ensures the operability of sub-regions in terms of both space and access costs, guaranteeing the rationality of zoning. The introduction of capacity constraints and load awareness mechanisms prevents resource overload, improves system stability, and ensures load balance. Event-driven feedback and pheromone reset enable rapid response to environmental changes, achieving dynamic adaptability. The use of Pareto unsets with cluster density control provides diverse optimization schemes, maintaining delivery efficiency while improving system load balance, increasing resource utilization, and shortening redistribution response time in dynamic scenarios. This effectively solves the problem of resource idleness and overload caused by uneven load in last-mile delivery, significantly improving the throughput and stability of the delivery system. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of a load balancing method for last-mile delivery networks based on a swarm intelligence algorithm, according to an embodiment of this application. Detailed Implementation

[0016] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0019] See Figure 1 A load balancing method for last-mile delivery networks based on swarm intelligence algorithms includes the following steps: S1: Construct a spatial partitioning model based on clustering; Preprocessing the last-mile delivery network divides the vast delivery area into more manageable sub-regions.

[0020] The spatial partitioning model inputs include historical order data (including geographical location), delivery station / rider location, and road network structure data.

[0021] The spatial partitioning model employs an improved K-Means++ clustering algorithm, introducing actual road network distance as a similarity metric on top of traditional Euclidean distance. This ensures the clustering results are reasonable both spatially and in terms of actual travel costs. Traditional methods calculate the straight-line distance between two points, which is unrealistic for urban delivery because it ignores real-world obstacles such as buildings, rivers, highways, and one-way streets. Therefore, this invention constructs a composite distance metric, introducing actual road network distance as a similarity measure. Specifically, the composite distance formula in the improved K-Means++ clustering algorithm is as follows: in, This is a distance function used to calculate the distance of an order o. i To a cluster center c j The distance; DrivingDist(o i ,c j () represents the actual road network distance, and () represents the distance from point o. i Point C j The actual shortest driving distance or time; optionally, obtained by calling an online map API or a localized route planning algorithm. This is the core element that makes the clustering results practical.

[0022] EuclideanDist(o i ,c j ) is the Euclidean distance, which serves as an auxiliary and stabilizing factor, retaining its advantages of simple and fast computation, and is used to maintain the relatively compact shape of the cluster on a macroscopic scale.

[0023] Norm() is a normalization function. Since the dimensions and numerical ranges of road distance and Euclidean distance differ greatly, direct weighted summation is meaningless. Normalization aims to eliminate the influence of dimensions.

[0024] To integrate weighting factors, when When the value is close to 1, clustering is almost entirely determined by actual road distances, yielding results that best reflect reality, but also incurring the highest computational cost. When the value approaches 0, it degenerates back to approximate Euclidean distance clustering, which is computationally fast but has low accuracy. This value is usually set based on a trade-off between accuracy and computational efficiency; it is optional. =0.7~0.9, giving higher weight to actual road distance.

[0025] The spatial partitioning model outputs multiple collaborative delivery sub-regions. Each sub-region is coordinated by a "virtual delivery center" or a main delivery station, with multiple delivery resources (riders or small stations) working collaboratively within the region. This step decomposes the global load balancing problem into multiple parallel sub-region load balancing problems, reducing the problem complexity.

[0026] S2: Construct a load balancing model with capacity constraints; This step, based on the sub-regions output by S1, constructs a refined load balancing model within each sub-region and solves it using an improved ant colony model. The capacity-constrained load balancing model constructed in this invention imposes multi-dimensional constraints on the maximum load capacity of each delivery resource within the sub-region. Its core advantage lies in using the upper limit of resource capacity as a hard prerequisite for optimization. This fundamentally eliminates the awkward situation where the theoretically optimal solution cannot be executed in reality due to the overload of individual resources, ensuring that every scheduling scheme derived by the algorithm is feasible, safe, and can significantly improve the overall throughput and stability of the system.

[0027] Specifically, the model input includes the set of real-time orders to be assigned within a defined sub-region. z ={o1,o2,...,o m Real-time traffic information matrix T z This represents the travel time between points within a sub-region and the real-time load status of each delivery resource within the sub-region, Lcurrent. z ={l1,l2,...,l n}

[0028] Capacity constraints refer to subregion Z k Each delivery resource within r i The maximum load capacity is a multi-dimensional constraint, including but not limited to: weight / volume capacity: the maximum weight and volume of packages that can be transported per day; time capacity: the maximum effective working hours per day; and order capacity: the maximum number of orders that can be processed per day.

[0029] The constraint condition is formally expressed as: for ∀r i ∈Rz, assigned to r i The total load of all orders must not exceed their current remaining capacity C i -l i .

[0030] The model output includes sub-region Z k An optimal or near-optimal order-resource allocation mapping scheme A z :O z ->R z .

[0031] An improved ant colony optimization algorithm is adopted, in which each order to be assigned is regarded as a node. The ant's path construction process simulates the decision-making process of sequentially allocating orders to different delivery resources, and finally forms a complete allocation scheme.

[0032] This invention improves the structure of the state transition rules in the ant colony algorithm, in the current order o i When allocation is required, select resource r. j The probability P(r) j |o i It is determined by the following formula: in: τ ij To make "order o" i Allocate to resource r j "In this decision-making process, the pheromone concentration is α, where α is the pheromone heuristic factor; η..." ij For improved heuristic information, β is the expected heuristic factor. It is no longer based on distance, but is defined as: Among them, Cost(o i ,r j ): refers to placing an order o i Allocate to resource r j The costs incurred can be distance costs, time costs, or a combination of both, ρ j It is a resource r j Real-time load factor ρ j =(l j / C j ), where γ is the load influence factor.

[0033] The state transition rule considers both historical experience (pheromone) and the immediate benefits of the current state (heuristic information). The heuristic information ensures that the higher the current load rate of a resource, the lower the probability that it will be selected to receive new orders, thus proactively guiding the system towards load balancing.

[0034] By making the above improvements to the ant colony algorithm, the following transformations were achieved: from "path optimization" to "allocation optimization," the core objective of the algorithm has been fundamentally changed; from "sequence decision-making" to "constraint decision-making," the ants' decision-making process is strictly limited by resource capacity constraints, which is more in line with real-world business scenarios; from "distance-guided" to "load and cost-guided": load information is deeply embedded into the core component of the algorithm's state transition rules, giving the algorithm an inherent drive to pursue load balancing.

[0035] This customized structural design for sub-region allocation problems is one of the key innovations of this invention. It enables swarm intelligence algorithms to solve the load balancing problem of last-mile delivery efficiently and directly, rather than indirectly through roundabout path optimization.

[0036] S3: Design pheromone update rules that integrate load information; To address the issue of traditional ant colony algorithms neglecting load balancing, this invention makes personalized improvements to the pheromone update rules.

[0037] Under traditional rules, pheromone concentration is only related to the cost (such as distance) of historical allocation schemes, leading to low-load resources becoming overloaded quickly due to frequent selection. Therefore, this application adopts an improved update rule. After each generation cycle, the pheromone is updated not only based on the total path cost of the allocation scheme, but more importantly, a load balancing index F is introduced. balance As a basis for updating.

[0038] New pheromone increment Δτ ij No longer just Q / L k (L) k (where Q is the total path length and Q is the pheromone intensity), but rather Δτ ij =Q / (L k (1+b F balance (k)). Where, Δτ ij To place orders i Allocate to resource r j "The incremental pheromone reward given by this specific decision, F" balance (k) is the load balance degree of the allocation scheme constructed by the kth ant, and b is the load balance influencing factor.

[0039] Even if the total path is short, the pheromone update rule will significantly reduce the pheromone increase if the load is highly uneven. Conversely, a scheme with a reasonable total path and a very balanced load will receive more pheromone enhancement. This guides the algorithm to search in a direction that balances cost and balance.

[0040] The load balancing degree calculation method is as follows: F balance =Variance=(Σ(ρ j -ρ avg )²) / N. Where, ρ j ρ represents the load factor of the j-th resource. avg This represents the average load rate of all resources, and N represents the total number of resources.

[0041] Preferably, a forward-looking pheromone update based on load trends is introduced; Traditional rules only penalize currently unbalanced solutions but fail to proactively guide the system towards a more balanced state. This application modifies heuristic information by calculating load trends to proactively avoid resources that, while currently under low load, are in a rapid increase phase, thereby preventing potential future congestion and achieving proactive load balancing. Specific implementation methods include: Calculate load trend: Record the load rate changes of each resource in the most recent scheduling cycles, calculate its trend, and optionally obtain the load rate slope s through linear regression. j . s j >0 indicates that the load is increasing, s j <0 indicates a decrease.

[0042] Modify the heuristic information: Change the heuristic information η ij Revised to: in, As an indicator function, when the load trend slope s of resource j j The value is 1 when the positive threshold θ is exceeded, and 0 otherwise. λ is the trend penalty factor.

[0043] S4: Establish a dynamic load feedback mechanism; To enable the model to respond to real-time changes, this invention establishes a closed-loop dynamic load feedback system that continuously monitors the real-time load and road conditions of all delivery resources.

[0044] When a trigger event is detected, dynamic adjustments are initiated.

[0045] Dynamic adjustment methods include local redistribution and local pheromone reset.

[0046] The local redistribution only applies to orders that have not yet started delivery and affected resources, using the current state as the initial condition, and re-executes the improved ant colony algorithm of S2-S3 for fast local optimization.

[0047] The pheromone local reset partially evaporates or resets the pheromones related to the affected area, avoiding old information from misleading new decisions and enhancing the algorithm's ability to explore new environments.

[0048] The core of this dynamic load feedback mechanism is to establish a closed-loop feedback system of "monitoring-judgment-execution," which directly intervenes in and improves the operation of the ant colony algorithm. Specifically, it includes the following steps: S41: Status monitoring and data awareness; Establish a real-time data dashboard to continuously track the following dynamic data: Order flow: Real-time access to new orders, order cancellation, and order completion status.

[0049] Resource status: Each delivery resource r j Real-time location, current load j (t), remaining capacity, and status (such as idle, busy, offline).

[0050] Environmental status: Real-time traffic conditions, weather, etc. within the sub-region.

[0051] S42: Feedback triggering and event judgment; Implementation: The system presets key event thresholds. When the monitored data meets one of the following conditions, dynamic adjustments are immediately triggered: Event A (New Task): A batch of new orders arrives, exceeding the preset threshold N. new .

[0052] Event B (Resource Overload): The real-time load rate of any resource exceeds a set threshold, i.e., ρ j (t)>θ overload .

[0053] Event C (Resource Failure): A resource suddenly goes offline due to a malfunction, absence, or other reasons.

[0054] Event D (Environmental Deterioration): Travel time in a certain area exceeds the original planned time by an average of δ due to congestion or severe weather.

[0055] By setting up a lightweight event listener, feedback and judgment are achieved by continuously comparing real-time data with thresholds.

[0056] S43: Dynamically adjust the execution layer; Once an event is triggered, the system does not recalculate from scratch. Instead, it initiates a lightweight, targeted, dynamic adjustment process that directly modifies the state and operating parameters of the ant colony model. This includes the following sub-steps: (1) Triggering local reallocation Lock down the affected areas. For example, for new orders, only reallocate new orders and currently idle or low-load resources. For resource failures, release all unexecuted orders on that resource back to the allocation pool and reallocate them along with current new orders.

[0057] Using the current system state as the initial condition, the improved ant colony algorithm described in S2 is invoked to perform a rapid iteration and find a new local optimal allocation scheme that adapts to the new situation.

[0058] (2) Dynamic evaporation and resetting of pheromones; Since the old pheromone matrix records experience from the environment before the event, it strongly guides ants toward the old, now invalid allocation scheme, hindering the algorithm from quickly adapting to the new environment. Therefore, this invention intervenes in the pheromone matrix before initiating local redistribution: For trigger events B / C / D, i.e., state deterioration categories, the pheromones associated with the affected resources are enhanced through localized volatilization. For example, all pheromones corresponding to the failed resources are affected. Apply an additional volatility coefficient ω (ω>1) to rapidly reduce it to near 0.

[0059] The expression is: ,in, For the updated pheromones, The pheromones before the update. The global volatility coefficient represents the operation as an active forgetting of outdated and bad experiences regarding the resource, forcing the algorithm to explore new paths.

[0060] For all triggered events: during local reallocation, temporarily increase the global pheromone evaporation coefficient and introduce a minimum pheromone value τ. min This prevents pheromone depletion. It enhances the algorithm's ability to explore in changing environments, enabling it to more quickly move beyond old optimal solutions and find new ones.

[0061] The dynamic load feedback mechanism of this invention has the following fundamental improvements: From offline computation to "looping online optimization": Placing the algorithm in a continuous "monitor-optimize-execute" loop enables it to handle dynamic problems.

[0062] An event-driven operating paradigm was introduced: this changed the passive iteration mode of the algorithm, enabling it to respond to external events, achieve on-demand activation and efficient computation.

[0063] The pheromone management strategy has been improved: targeted pheromone reset and evaporation rules have been designed, which is a significant improvement to the algorithm's core memory system. This enables the algorithm to "selectively forget," quickly discarding outdated experiences, thus demonstrating excellent environmental adaptability.

[0064] The allocation of computing resources has been optimized: by “local reallocation” instead of “global recalculation”, the computing overhead of dynamic adjustment has been greatly reduced, meeting the high real-time requirements of last-mile delivery scenarios.

[0065] S5: Establish a multi-objective Pareto optimal solution set selection mechanism; The load balancing problem of this invention is a multi-objective optimization problem, and Pareto is used to make the optimal solution set decision on the candidate solutions generated by the ant colony algorithm.

[0066] Optionally, the optimization objective may include simultaneously optimizing two or more of the following objectives: Objective 1 (Cost): Minimize total delivery distance / time.

[0067] Objective 2 (Balance): Minimize the variance of the load rate of all delivery resources.

[0068] Objective 3 (Efficiency): Minimize the average order delivery time.

[0069] Pareto solution set generation: During the iterative process of the improved ant colony algorithm, all non-dominated solutions (i.e. solutions on the Pareto front) are recorded to form a solution set.

[0070] Final strategy confirmation: From the Pareto optimal solution set, a final implementation plan is selected based on the decision-maker's preferences. The decision-maker's preferences can be set by weights.

[0071] The final implementation plan specifically includes: Order allocation scheme: Each order is clearly assigned to a specific delivery resource.

[0072] Resource load prediction: Based on this scheme, predict the final load status of each resource after task execution.

[0073] Region selection suggestion: In cases of extreme imbalance, it may be recommended to dynamically allocate a portion of orders from a certain sub-region to adjacent sub-regions for collaborative processing.

[0074] Under traditional mechanisms, the Pareto solution set may be filled with similar solutions that are too concentrated in the objective space, leading to decision-makers facing a large number of duplicate options, while some potentially marginal solutions may be eliminated. Therefore, the preferred multi-objective Pareto optimal solution set screening mechanism of this application is based on introducing an elite solution set maintenance strategy based on cluster density control, as follows: When a new non-dominated solution needs to be added to a full solution set, instead of simply replacing an existing solution, an adaptive grid method is used or the solution set is clustered in the target space. When the solution set is full, a solution from the region with the highest density within the same cluster is removed first, or a representative solution from each cluster center is retained.

[0075] This ensures that the final Pareto solution set is widely and uniformly distributed in the target space, providing decision-makers with truly diverse choices.

[0076] In this implementation, composite distance clustering ensures the operability of sub-regions in terms of both space and access costs, guaranteeing the rationality of zoning. Capacity constraints and load awareness mechanisms are introduced to prevent resource overload, improve system stability, and ensure load balance. Event-driven feedback and pheromone reset enable rapid response to environmental changes, achieving dynamic adaptability. Pareto unsets with cluster density control provide diverse optimization schemes, maintaining delivery efficiency while improving system load balance, increasing resource utilization, and shortening redistribution response time in dynamic scenarios. This effectively solves the problem of resource idleness and overload caused by uneven load in last-mile delivery, significantly improving the throughput and stability of the delivery system.

[0077] This invention also proposes a load balancing system for last-mile delivery networks based on swarm intelligence algorithms, which, based on the load balancing method for last-mile delivery networks based on swarm intelligence algorithms described above, includes: A clustering-based spatial partitioning model construction module is used to preprocess the last-mile delivery network, dividing the vast delivery area into more manageable sub-regions. A load balancing model building module with capacity constraints is used to build a refined load balancing model within each sub-region and solve it using an improved ant colony model. The pheromone update rule design module integrates load information, which is used to introduce load balance index as the basis for pheromone updates and introduce forward-looking pheromone updates based on load trends. The dynamic load feedback mechanism module is used to establish a closed-loop dynamic load feedback system. The system continuously monitors the real-time load and road conditions of all delivery resources; when a trigger event is detected, dynamic adjustments are triggered. The module for establishing a multi-objective Pareto optimal solution set selection mechanism is used to make optimal solution set decisions for candidate solutions generated by the ant colony algorithm using Pareto.

[0078] Furthermore, this embodiment of the invention also proposes a computer-readable storage medium storing program instructions for a load balancing method for last-mile delivery networks based on a swarm intelligence algorithm. These program instructions can be executed by one or more processors to implement the steps of the load balancing method for last-mile delivery networks based on a swarm intelligence algorithm as described above.

[0079] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A load balancing method for last-mile delivery networks based on swarm intelligence algorithms, characterized in that, include: S1: Construct a cluster-based spatial partitioning model, including preprocessing the last-mile delivery network and dividing large delivery areas into multiple sub-regions; S2: Construct a load balancing model with capacity constraints, including building a refined load balancing model within each sub-region and solving it using an improved ant colony model. S3: Design pheromone update rules that integrate load information, including introducing load balance index as the basis for pheromone updates and introducing forward-looking pheromone updates based on load trends. S4: Establish a closed-loop dynamic load feedback mechanism, including continuous monitoring of the real-time load and road conditions of all delivery resources; trigger dynamic adjustments when a triggering event is detected. S5: Establish a multi-objective Pareto optimal solution set selection mechanism, including using Pareto to make optimal solution set decisions for candidate solutions generated by the ant colony algorithm.

2. The load balancing method for last-mile delivery networks based on swarm intelligence algorithms according to claim 1, characterized in that, The clustering-based spatial partitioning model employs an improved K-Means++ clustering algorithm, which introduces the actual distance of the road network as a similarity metric to obtain a composite distance based on the traditional Euclidean distance.

3. The load balancing method for last-mile delivery networks based on swarm intelligence algorithms according to claim 1, characterized in that, The capacity constraint of the load balancing model with capacity constraints includes the maximum load capacity of each delivery resource in the sub-region. The capacity is a multi-dimensional constraint, including weight and / or volume capacity, daily effective working hours limit, order capacity, and the maximum number of orders that can be processed per day.

4. The load balancing method for last-mile delivery networks based on swarm intelligence algorithms according to claim 3, characterized in that, The load balancing index is introduced as the basis for pheromone updates, and the pheromone increment expression is as follows: Δτ ij =Q / (L k (1+b F balance (k)), where Δτ ij To place orders i Allocate to resource r j "The incremental pheromone reward given by this specific decision, F" balance (k) is the load balance degree of the allocation scheme constructed by the kth ant, and b is the load balance influencing factor.

5. The load balancing method for last-mile delivery networks based on swarm intelligence algorithms according to claim 4, characterized in that, The forward-looking pheromone update based on load trends includes calculating the load trend and modifying the heuristic information η according to the trend. ij Revised to: ； in, To place orders i Allocate to resource r j The resulting costs, As the load impact factor, For resource r j Real-time load rate, As an indicator function, when the load trend slope s of resource j j The value is 1 when the positive threshold θ is exceeded, and 0 otherwise. λ is the trend penalty factor.

6. The load balancing method for last-mile delivery networks based on swarm intelligence algorithms according to claim 1, characterized in that, The triggering events for the dynamic load feedback include new tasks, resource overload, resource failure, or environmental degradation. The dynamic adjustment methods include local reallocation and pheromone local reset.

7. The load balancing method for last-mile delivery networks based on swarm intelligence algorithms according to claim 6, characterized in that, The local redistribution involves locking the problem area affected by the event, using the current system state as the initial condition, and calling the improved ant colony algorithm described in S2 to perform a round of rapid iteration.

8. The load balancing method for last-mile delivery networks based on swarm intelligence algorithms according to claim 6, characterized in that, The pheromone local reset includes enhancing the local evaporation of pheromones related to the affected resources for state deterioration-type trigger events; for all trigger events: when performing local re-optimization, temporarily increasing the global pheromone evaporation coefficient and introducing a minimum pheromone value to prevent pheromone depletion.

9. A load balancing system for last-mile delivery networks based on swarm intelligence algorithms, comprising the load balancing method for last-mile delivery networks based on swarm intelligence algorithms as described in any one of claims 1 to 8, including: A clustering-based spatial partitioning model construction module is used to preprocess the last-mile delivery network, dividing the vast delivery area into more manageable sub-regions. A load balancing model building module with capacity constraints is used to build a refined load balancing model within each sub-region and solve it using an improved ant colony model. The pheromone update rule design module integrates load information, which is used to introduce load balance index as the basis for pheromone updates and introduce forward-looking pheromone updates based on load trends. The dynamic load feedback mechanism module is used to establish a closed-loop dynamic load feedback system. The system continuously monitors the real-time load and road conditions of all delivery resources; when a trigger event is detected, dynamic adjustments are triggered. The module for establishing a multi-objective Pareto optimal solution set selection mechanism is used to make optimal solution set decisions for candidate solutions generated by the ant colony algorithm using Pareto.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions for a load balancing method for last-mile delivery networks based on a swarm intelligence algorithm. These program instructions can be executed by one or more processors to implement the steps of the load balancing method for last-mile delivery networks based on a swarm intelligence algorithm as described in any one of claims 1 to 8.

Images