A stochastic programming method and related equipment for cabinet-based collaborative delivery networks
By using a two-stage stochastic programming model and a deep neural network model, the problems of long computation time and resource waste in large-scale cabinet collaborative delivery networks are solved, achieving fast and robust network planning and improving search efficiency and solution quality.
Patent Information
- Application Number
- CN202611432211.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-15
- Publication Date
- 2026-11-17
AI Technical Summary
Existing cabinet-based collaborative delivery network design algorithms take too long to compute when faced with large-scale, dynamic, and uncertain demands, and are prone to resource waste or insufficient capacity. Traditional precise algorithms and heuristic algorithms are inefficient in large-scale scenarios.
A two-stage stochastic programming model is combined with an adaptive large neighborhood search algorithm and a deep neural network model. Heterogeneous demand scenarios are generated through Monte Carlo sampling. The deep neural network is used to quickly infer truck replenishment costs and update the search operator weights, replacing the traditional time-consuming vehicle path solver.
It enables rapid and high-quality collaborative delivery network planning in large-scale network design, improves search depth and breadth, outputs globally optimal and robust planning solutions, and avoids resource waste and capacity shortage.
Smart Images

Figure CN123117970A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of smart logistics and relates to a stochastic planning method and related equipment for cabinet-based collaborative delivery networks. Background Technology
[0002] With the rapid development of e-commerce, last-mile delivery is facing increasingly severe challenges in terms of cost and efficiency. Drones, with their advantages of flexibility, speed, and immunity to ground traffic congestion, are gradually becoming a new favorite in logistics delivery. However, limited by battery life and payload capacity, drones cannot independently undertake large-scale, long-distance delivery tasks. Therefore, the "cabinet-based collaborative delivery model," which uses parcel lockers as take-off and landing platforms and transfer stations for drones, combined with trucks for replenishment, has emerged. This model expands the service range of drones through parcel lockers and utilizes the large capacity of trucks for centralized replenishment, achieving complementary advantages and significantly improving delivery efficiency.
[0003] In actual logistics operations, customer demand is often highly uncertain. Traditional deterministic network design models struggle to cope with dynamic fluctuations in demand, easily leading to insufficient parcel locker capacity or resource waste. Therefore, the design of locker collaborative delivery networks considering stochastic demand has become a research hotspot. Existing exact algorithms (such as the L-shaped algorithm and the branch-and-price cutting plane method) can guarantee optimality, but their computation time increases exponentially with the increase in customer scale and the number of scenarios, making it impossible to solve medium-to-large-scale cases within a reasonable timeframe. Commonly used metaheuristic algorithms (such as adaptive large neighborhood search) have good optimization capabilities, but during the search process, after each iteration generates a new solution, traditional VRP solvers (such as the C&W saving algorithm and the LKH algorithm) are still called to evaluate the subproblems. This nested solver approach in iterations is extremely time-consuming, severely limiting the search depth and breadth of the algorithm in large-scale scenarios. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a stochastic planning method and related equipment for cabinet-based collaborative delivery networks, which enables rapid and high-quality solutions for logistics network design schemes under large-scale stochastic scenarios.
[0005] To achieve the above objectives, the present invention employs the following technical solution: A stochastic programming method for a cabinet-based collaborative delivery network includes: Obtain delivery parameters for the target delivery area and generate multiple sets of heterogeneous demand feature scenarios; Based on the set of multiple heterogeneous demand characteristics, a two-stage stochastic programming model is constructed, which includes a first-stage strategic decision and a second-stage tactical decision. The first-stage strategic decision includes parcel locker location parameters and drone configuration parameters. An adaptive large neighborhood search algorithm is used to iteratively search the parcel locker location parameters and UAV configuration parameters within the first stage of strategic decision-making. When a new solution is generated in each search iteration, the new solution is input into a pre-trained deep neural network model. The truck replenishment cost corresponding to the new solution in the heterogeneous demand characteristic scenario set is inferred by a deep neural network model. The truck replenishment cost belongs to the second stage tactical decision. The truck replenishment cost is fed back as a fitness value to the adaptive large neighborhood search algorithm, and the operator weights of the adaptive large neighborhood search algorithm are updated. When the adaptive large neighborhood search algorithm reaches the preset termination condition, iterative search stops and the final cabinet collaborative delivery network planning scheme is output.
[0006] Optionally, the objective function of the two-stage stochastic programming model is obtained by adding the fixed construction cost and the expected operating cost; Fixed construction costs include the construction cost of the parcel lockers and the cost of purchasing drones; Expected operating costs include drone delivery costs, truck transportation costs, and penalty costs arising from demand fluctuations; The constraints of the two-stage stochastic programming model include the capacity limit of the parcel locker service, the divisibility of customer demand, and the service range constraint of the drone.
[0007] Optionally, the process of constructing a two-stage stochastic programming model includes: A sample mean approximation framework is adopted, and independent and identically distributed requirement scenario samples are generated based on Monte Carlo sampling. The stochastic programming problem is transformed into a deterministic equivalent problem, and a statistical estimate of the objective function value is calculated to obtain a statistical upper bound. Calculate the mean of the optimal values of multiple independent approximation problems to obtain a statistical lower bound; The outer iteration of the sample mean approximation framework is terminated when the difference between the statistical upper bound and the statistical lower bound is less than a preset gap value.
[0008] Optionally, the deep neural network model includes an encoder module, a decoder module, and a masking mechanism; The encoder module maps the coordinates and normalized requirement features of the parcel lockers, distribution centers, and customer nodes into high-dimensional embedding vectors. The decoder module combines the context vector and the masking mechanism to output the probability distribution of the truck visiting the next node; The masking mechanism blocks access to nodes and nodes that violate capacity constraints.
[0009] Optionally, the processing steps for deep neural network models include: Within the encoder module, a multi-head attention layer is used to extract the spatial and demand features of the parcel locker nodes and the distribution center, and generate corresponding node embedding vectors. Inside the decoder module, the node embedding vector and the current vehicle state are input into the bidirectional gating recurrent unit network, and the probability distribution of the truck accessing the next node is output based on the bidirectional gating recurrent unit network.
[0010] Optionally, the adaptive large neighborhood search algorithm includes both destruction and repair operators; The destruction operators include random removal operators, worst-cost removal operators, correlation removal operators, and drone number adjustment removal operators; The repaired operators cover the greedy insertion operator, the regret insertion operator, and the drone redistribution operator.
[0011] Optionally, the iterative search process using the adaptive large neighborhood search algorithm includes: Based on the historical call to the operator's performance, some of the selected parcel lockers can be removed from the current solution or the configuration parameters of the drone can be reduced. The repair operator is invoked to re-insert the unselected candidate points into the current solution and generate a new solution to be evaluated, which includes new location parameters and UAV configuration parameters. The total cost of the new solution to be evaluated is calculated by combining the truck replenishment cost, and the acceptance of the new solution to be evaluated is determined by the simulated annealing criterion.
[0012] A stochastic programming system for a cabinet-based collaborative delivery network includes: The scene generation module is used to obtain delivery parameters for the target delivery area and generate multiple sets of heterogeneous demand feature scene sets. The model building module is used to construct a two-stage stochastic programming model based on the set of multiple heterogeneous demand characteristic scenarios, which includes a first-stage strategic decision and a second-stage tactical decision. The first-stage strategic decision includes parcel locker location parameters and drone configuration parameters. The search iteration module is used to iteratively search the parcel locker location parameters and UAV configuration parameters within the first stage strategic decision using an adaptive large neighborhood search algorithm, and inputs the new solution into a pre-trained deep neural network model when a new solution is generated in each search iteration. The cost inference module is used to infer the truck replenishment cost corresponding to the new solution under the heterogeneous demand feature scenario set through a deep neural network model. The truck replenishment cost belongs to the second stage tactical decision. The weight update module is used to feed the truck replenishment cost as a fitness value back to the adaptive large neighborhood search algorithm and update the operator weights of the adaptive large neighborhood search algorithm. The result output module is used to stop the iterative search and output the final cabinet collaborative delivery network planning scheme when the adaptive large neighborhood search algorithm reaches the preset termination condition.
[0013] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the stochastic planning method for the cabinet collaborative delivery network.
[0014] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the stochastic planning method for the cabinet-based collaborative delivery network.
[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a two-stage stochastic programming model encompassing strategic and tactical decision-making by combining a heterogeneous demand feature scenario set generated through Monte Carlo sampling. This model can effectively simulate and cope with highly uncertain dynamic demand fluctuations in actual logistics, avoiding insufficient parcel locker capacity or resource waste caused by traditional deterministic models. Simultaneously, a pre-trained deep neural network model is introduced into the iterative process of adaptive large neighborhood search to quickly infer the truck replenishment cost in the second stage, and uses it as a fitness value to update the search operator weights. This replaces the traditional, extremely time-consuming vehicle routing solver, eliminates the time bottleneck caused by nested solutions in the iteration, and achieves a significant improvement in search depth and breadth in large-scale network design, quickly outputting a globally optimal and robust collaborative delivery network planning scheme.
[0016] Furthermore, by scientifically decomposing the model's objective function into fixed construction costs and expected operating costs, and introducing a penalty cost mechanism, the actual cost consumption in the process of drone procurement, parcel locker construction, and real operation in multiple scenarios can be accurately quantified. Combined with the constraints of parcel locker service capacity limitations, divisible customer demand, and drone service range constraints, the constructed model is more in line with the business constraints of real-world cabinet collaborative delivery scenarios, ensuring the feasibility of the generated strategic site selection and tactical scheduling schemes in actual application, effectively avoiding overloading, boundary crossing, and other failed planning, and further guaranteeing the actual operational benefits of the planning scheme.
[0017] Furthermore, Monte Carlo sampling is used to generate independent and identically distributed demand scenario samples, transforming the stochastic programming problem into a computationally feasible deterministic equivalent problem, providing a stable and reliable simulation environment for multi-scenario inference. A statistical lower bound is constructed using the mean of the optimal values of the independent approximation problem, and a statistical upper bound is constructed by evaluating candidate solutions for a large-scale reference sample set. The difference between the statistical upper bound and the lower bound is used as the termination condition for the outer iteration, ensuring the convergence accuracy of the approximate solution in approximating the true optimal solution, and ensuring the reliability and high quality of the optimization results of the algorithm in large-scale complex stochastic demand scenarios.
[0018] Furthermore, by using the encoder module to map various node coordinates and demands into high-dimensional embedding vectors, a deep representation and extraction of the original state information of the delivery network is achieved. By combining the context vector with the decoder module to output the probability distribution of trucks visiting the next node, vehicle route planning is transformed into a sequential decision-making process, reducing the solution complexity. With the introduction of a masking mechanism, visited nodes and infeasible nodes that violate capacity constraints are directly masked in the node generation stage, preventing the generation of invalid paths and ensuring that the replenishment path inferred and output by the neural network model strictly meets the feasibility constraints, thereby improving the accuracy of operating cost assessment.
[0019] Furthermore, a multi-head attention layer is introduced at the encoder end. By adaptively calculating the attention weights between nodes, the spatial topological dependency between node distribution and demand is fully captured and extracted, so that the embedding vector contains rich global spatial features. At the decoder end, a bidirectional gated recurrent unit network is used to process the node embedding vector and vehicle state, effectively capturing the forward and backward temporal dependencies in truck path sequence decision-making. This makes the probability distribution prediction of trucks visiting the next node take into account both global spatial topology and local temporal constraints, thereby establishing an inference mapping from network state to actual operating cost, improving the generalization ability and accuracy of the neural estimator.
[0020] Furthermore, by introducing disruptive operators such as random removal, worst-cost removal, relevance removal, and drone quantity adjustment removal, the current local solution can be effectively disrupted from multiple dimensions such as geographical distance, demand relevance, and cost utilization, thus escaping the local optimum trap, releasing redundant resources, and optimizing inefficient cabinet configuration. At the same time, through repair operators such as greedy insertion, regret insertion, and drone reassignment, nodes are reassigned based on cost savings and the cost of suboptimal selection, realizing intelligent repair and reconstruction of the solution structure. The rich operator library ensures the diversity of the heuristic algorithm's search in the solution space and the breadth of optimization.
[0021] Furthermore, by dynamically invoking destruction and repair operators based on their historical performance and adaptively updating their weights, the algorithm can automatically tilt towards optimization directions with better improvement effects, thereby improving search efficiency. After generating a new solution to be evaluated, it is directly input into the neural network model to obtain an approximate cost, avoiding the large-scale recalculation overhead caused by calling traditional vehicle path solvers. At the same time, the simulated annealing criterion based on cost difference is used to determine whether a solution is accepted. While giving inferior solutions a certain probability of acceptance to avoid the algorithm getting trapped in local extrema, the efficient cost feedback of the neural network accelerates the convergence speed of global optimization, ensuring that the optimal cabinet location and drone configuration scheme is output within a reasonable time. Attached Figure Description
[0022] Figure 1 This is a flowchart of the stochastic planning method for the cabinet-based collaborative delivery network of the present invention; Figure 2 This is a schematic diagram of the network structure of the neural path inference model of the present invention. Detailed Implementation
[0023] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0024] The following disclosure provides many different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. In addition, examples of various specific processes and materials are provided in this invention, but those skilled in the art will recognize the application of other processes and / or the use of other materials.
[0025] Example 1 This embodiment discloses a stochastic programming method for cabinet-based collaborative delivery networks that integrates spatiotemporal feature reasoning and adaptive search, including the following process: S1 obtains the distribution of customer nodes, potential parcel locker candidate points, and operational parameters of trucks and drones in the target delivery area. Combined with historical demand fluctuation characteristics, it generates a heterogeneous demand feature scenario set that follows a specific probability distribution through Monte Carlo sampling.
[0026] S2 establishes a two-stage stochastic programming model for the cabinet collaborative delivery network with the goal of minimizing the expected total operating cost. By introducing constraints on the service capacity of the parcel lockers, the divisibility of customer demand, and the service range of drones, a nonlinear coupling is achieved between the first-stage strategic decision (parcel locker location and drone configuration) and the second-stage tactical decision (customer allocation and truck replenishment routes).
[0027] The objective function of the constructed two-stage stochastic programming model consists of two parts: the first part is the fixed construction cost, including the rental or construction cost of parcel lockers and the purchase cost of drones; the second part is the expected operating cost, including drone delivery cost, truck transportation cost, and penalty cost or replenishment cost due to demand fluctuations.
[0028] S3 employs a sample mean approximation framework, generating N independent and identically distributed demand scenario samples through Monte Carlo sampling. This transforms the stochastic programming problem in S2 into a deterministic equivalence problem, providing a simulation environment for multi-scenario inference for the subsequent neural evaluator.
[0029] The sample mean approximation framework also includes a boundary evaluation step: calculating the mean of the optimal values of multiple independent approximation problems as a statistical lower bound; evaluating candidate solutions for a large-scale reference sample set and calculating a statistical estimate of their objective function value as a statistical upper bound; stopping the outer iteration when the difference between the statistical upper bound and the statistical lower bound is less than a preset gap value.
[0030] S4. Construct a deep neural network model based on the attention mechanism, defining the truck replenishment problem as a sequential decision-making process. The encoding layer uses a multi-head attention mechanism to extract the spatial topological features of the delivery node coordinates and demand. The decoding layer uses a bidirectional gated recurrent unit (Bi-GRU) to define the truck replenishment cost as the mapping output of the sequential decision-making process. The inference mapping from network state to operating cost is established through offline pre-training.
[0031] The deep neural network model is based on an encoder-decoder architecture, including: an encoder module, which maps the coordinates and normalized requirements of parcel lockers, distribution centers, and customer nodes into high-dimensional embedding vectors, and captures the spatial dependencies between nodes through a multi-head attention layer; a decoder module, which uses a bidirectional GRU network, combined with the current context vector and a masking mechanism, to output the probability distribution of a truck visiting the next node; and a masking mechanism, which masks visited nodes and nodes that violate capacity constraints when generating paths, ensuring that the generated replenishment paths are feasible.
[0032] The construction process of the deep neural network model specifically includes: in the encoder part, a multi-head attention mechanism is used to extract the spatial features and demand features of the currently selected parcel locker node and the distribution center, and generate node embedding vectors; in the decoder part, based on the node embedding vectors and the current vehicle state, a bidirectional gating loop unit is used to output the probability distribution of the truck accessing the next node.
[0033] S5 uses the Adaptive Large Neighborhood Search (ALNS) algorithm to iteratively search the decision variables of the first stage. In each iteration evaluation process, the neural path evaluator trained in S4 is called to quickly infer and calculate the truck replenishment cost of the second stage for the current location scheme. The cost sensitivity feedback is used as the fitness value to dynamically adjust the weight of the search operator and guide the algorithm to converge to the global optimal region.
[0034] The Adaptive Large Neighborhood Search (ALNS) algorithm includes several targeted operators: Destructive operators include: random removal, worst-cost removal (removing parcel lockers with low utilization), relevance removal (removing parcel lockers that are too close in space), and drone number adjustment removal; Repair operators include: greedy insertion (selecting the position with the greatest cost reduction), regret insertion (considering the cost difference of the second-best choice), and drone redistribution (optimizing the number of drones at existing locker locations).
[0035] The Adaptive Large Neighborhood Search (ALNS) process is as follows: Based on historical performance, an adaptive removal strategy is selected to remove some of the already located parcel lockers or reduce the number of drones in the current solution; the removal strategies include random removal, worst-cost removal, and relevance removal; based on a greedy strategy or a regret value strategy, unselected candidate points are reinserted into the solution, or the number of drones is increased; for new solutions generated after destruction and repair (i.e., new location and drone configuration schemes), instead of directly calling the traditional vehicle path solver, the set of nodes under this scheme is input into the neural path estimator to directly obtain the approximate operating cost of the second stage; combining the fixed cost of the first stage with the cost of the second stage fed back by the neural estimator, the total cost of the new solution is calculated, and the simulated annealing criterion is used to decide whether to accept the new solution.
[0036] S6, when the preset termination condition is met, outputs the coordinates of the parcel locker location that minimizes the expected total cost, the number of drones deployed at each point, and the truck delivery instructions for each scenario.
[0037] This embodiment constructs a stochastic planning framework that integrates spatiotemporal feature reasoning and adaptive search, and uses a deep neural network based on an attention mechanism as a proxy evaluator to replace the traditional time-consuming vehicle path solver. This method not only overcomes the infeasibility of traditional exact algorithms in large-scale network design, but also significantly improves the efficiency and solution quality of heuristic search, making the logistics network solution more robust and with lower expected total operating costs when dealing with demand fluctuations.
[0038] Example 2 This embodiment provides a stochastic planning method for cabinet-based collaborative delivery networks that integrates spatiotemporal feature reasoning and adaptive search, mainly including the following steps: Step 1: Constructing a sample mean approximation framework and multi-scenario dimensionality reduction. To address the uncertainty of requirements, this invention... Figure 1 The outer loop shown uses a sample mean approximation method. The maximum number of independent repeated trials in the outer loop is set. In the first In the repeated experiments ( ), generated through Monte Carlo sampling containing Training sets of independent, identically distributed scenes For any given first-stage decision scheme (parcel locker location and drone deployment), the statistical lower bound estimate of its expected objective function value is... and its variance The calculation formula is as follows:
[0039]
[0040] Using inclusion Candidate solutions are evaluated using a large-scale validation set for each scenario, and a statistical upper bound is calculated. The solution is considered complete when the gap between the statistical upper and lower bounds is less than a preset threshold. When the time is reached, the outer iteration is terminated and the final solution is output.
[0041] Step 2: Initial Solution Generation Based on Greedy Strategy. Before the search begins, an initial solution is generated using a greedy strategy based on efficiency scoring. First, calculate the value of each candidate parcel locker. Efficiency rating:
[0042] In the formula, For parcel lockers The customer base within the service area For customers Expected demand, To fix the construction costs of the parcel lockers, The cost of purchasing a single drone is calculated. Then, considering the divisibility of demand, a competing factor is introduced. Calculation cabinet Estimated service demand:
[0043] In the formula, In order to cover customers A collection of all candidate counter locations. The number of drones to be deployed is determined based on estimated demand. In the formula, The average round-trip service time for drones. This represents the maximum daily operating time.
[0044] Step 3: Perform an adaptive large neighborhood search (ALNS) iteration. For example... Figure 1 As shown, in the inner loop, a neighborhood transformation is performed on the current solution using destruction and repair operators. For the destruction operator, a cabinet point is defined. Reference counter Relevance :
[0045] In the formula, For normalized geographical distance similarity, To share the proportion of potential customers, These are the weighting coefficients. For the repair operator, calculate the unselected counter locations. Regret value :
[0046] In the formula, and These represent the cost savings from inserting the point in the optimal and second-best positions, respectively. The node with the largest regret value is prioritized for insertion.
[0047] set up For the first Operator at the next iteration The selection weight is based on the score of the previous stage. Update:
[0048] In the formula, The reaction coefficient, For operators The number of times it is called during this phase.
[0049] Step 4: Invoke the spatiotemporal feature neural evaluator for inference. This is to quickly evaluate candidate solutions. The second phase of operating costs, this invention constructs as follows Figure 2 The deep neural network model shown. For the encoder, the input is the node coordinates. and normalization requirements The input is mapped to a high-dimensional feature vector through a linear embedding layer. :
[0050] A multi-head attention (MHA) mechanism is employed to extract spatial topological features. The query vector is then calculated. Key vector Sum value vector Attention weights:
[0051] The feature vectors are processed through multiple residual connections and layer normalization (Norm) to output the final node embedding representation. :
[0052] For the decoder, a bidirectional gated recurrent unit (Bi-GRU) is used to capture the temporal dependencies of the path sequence. Forward hidden state. With backward hidden state After concatenation, the access probability is generated by combining the current context.
[0053] For candidate nodes Calculate the logarithmic probability of its selection. :
[0054] In the formula, For graph-level features, This is a dynamic constraint feature.
[0055] A masking mechanism is introduced to block infeasible nodes, and the selection probability is output through Softmax. :
[0056] The path cost corresponding to the node sequence with the highest probability output by the model is used as the cost for the second stage. An approximation of the total cost is used to calculate the total cost. .
[0057] Step 5: Simulated annealing acceptance criteria and final solution output. Calculate the cost difference between the new solution and the current solution. The simulated annealing criterion is used to determine whether to accept a new solution, and the acceptance probability is... for:
[0058] When the maximum number of iterations or the required convergence accuracy is reached, the optimal parcel locker location, drone configuration scheme, and truck replenishment route are output.
[0059] Example 3 The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not omitted in the apparatus embodiments, please refer to the embodiments of the method of the present invention.
[0060] In another embodiment of the present invention, a stochastic programming system for cabinet-based collaborative delivery networks is provided. This stochastic programming system can be used to implement the above-mentioned stochastic programming method for cabinet-based collaborative delivery networks. Specifically, the stochastic programming system for cabinet-based collaborative delivery networks includes a scenario generation module, a model building module, a search iteration module, a cost inference module, a weight update module, and a result output module.
[0061] The scene generation module is used to obtain delivery parameters for the target delivery area and generate multiple sets of heterogeneous demand feature scene sets.
[0062] The model building module is used to construct a two-stage stochastic programming model based on the set of multiple heterogeneous demand characteristics, which includes a first-stage strategic decision and a second-stage tactical decision. The first-stage strategic decision includes parcel locker location parameters and drone configuration parameters.
[0063] The search iteration module is used to iteratively search the parcel locker location parameters and UAV configuration parameters within the first-stage strategic decision using an adaptive large neighborhood search algorithm. When a new solution is generated in each search iteration, the new solution is input into a pre-trained deep neural network model.
[0064] The cost inference module is used to infer the truck replenishment cost corresponding to the new solution under the heterogeneous demand feature scenario set through a deep neural network model. The truck replenishment cost belongs to the second stage tactical decision.
[0065] The weight update module is used to feed the truck replenishment cost as a fitness value back to the adaptive large neighborhood search algorithm and update the operator weights of the adaptive large neighborhood search algorithm.
[0066] The results output module is used to stop the iterative search and output the final cabinet collaborative delivery network planning scheme when the adaptive large neighborhood search algorithm reaches the preset termination condition.
[0067] Example 4 In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the computer device, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function. The processor described in this embodiment can be used in the operation of a stochastic programming method for a cabinet-based collaborative delivery network, including: acquiring delivery parameters of the target delivery area and generating multiple sets of heterogeneous demand characteristic scenario sets; constructing a two-stage stochastic programming model based on the multiple sets of heterogeneous demand characteristic scenario sets, including a first-stage strategic decision and a second-stage tactical decision, wherein the first-stage strategic decision includes... Parcel locker location parameters and drone configuration parameters; an adaptive large neighborhood search algorithm is used to iteratively search for the parcel locker location parameters and drone configuration parameters within the first stage of strategic decision-making. Each time a new solution is generated in the search iteration, it is input into a pre-trained deep neural network model. The deep neural network model infers the truck replenishment cost corresponding to the new solution in a heterogeneous demand characteristic scenario set. The truck replenishment cost belongs to the second stage of tactical decision-making. The truck replenishment cost is fed back as a fitness value to the adaptive large neighborhood search algorithm, and the operator weights of the adaptive large neighborhood search algorithm are updated. When the adaptive large neighborhood search algorithm reaches the preset termination condition, the iterative search stops and the final parcel locker collaborative delivery network planning scheme is output.
[0068] Example 5 In another embodiment, the present invention also provides a computer-readable storage medium (Memory), which is a memory device in a computer device for storing programs and data. It is understood that the computer-readable storage medium here may include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the operating system of the computer device. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here may be high-speed RAM or non-volatile memory, such as at least one disk storage device.
[0069] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the stochastic programming method for the cabinet collaborative delivery network in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by the processor to perform the following steps: including: obtaining delivery parameters of the target delivery area and generating multiple sets of heterogeneous demand feature scenarios; constructing a two-stage stochastic programming model containing a first-stage strategic decision and a second-stage tactical decision based on the multiple sets of heterogeneous demand feature scenarios, wherein the first-stage strategic decision includes cabinet location parameters and drone configuration parameters; and utilizing an adaptive large neighborhood search algorithm. The algorithm iteratively searches for the parcel locker location parameters and drone configuration parameters within the first-stage strategic decision-making process. Each time a new solution is generated during the search iteration, it is input into a pre-trained deep neural network model. The deep neural network model infers the truck replenishment cost corresponding to the new solution in a heterogeneous demand characteristic scenario set. The truck replenishment cost belongs to the second-stage tactical decision-making process. The truck replenishment cost is fed back as a fitness value to the adaptive large neighborhood search algorithm, and the operator weights of the adaptive large neighborhood search algorithm are updated. When the adaptive large neighborhood search algorithm reaches the preset termination condition, the iterative search stops and the final locker collaborative delivery network planning scheme is output.
[0070] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0071] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0072] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0073] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0074] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0075] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.
[0076] It should be understood that the above description is for illustrative purposes and not for limitation. Many embodiments and applications beyond the provided examples will be apparent to those skilled in the art upon reading the above description. Therefore, the scope of this application should not be determined by reference to the above description, but rather by reference to the foregoing claims and the full scope of their equivalents. For purposes of completeness, all articles and references, including patent applications and publications, are incorporated herein by reference. The omission of any aspect of the subject matter disclosed herein in the foregoing claims is not intended as a waiver of that subject matter, nor should it be construed as an indication that the applicant has not considered that subject matter as part of the disclosed inventive subject matter.
Claims
1. A stochastic programming method for a cabinet-based collaborative delivery network, characterized in that, include: Obtain delivery parameters for the target delivery area and generate multiple sets of heterogeneous demand feature scenarios; Based on the set of multiple heterogeneous demand characteristics, a two-stage stochastic programming model is constructed, which includes a first-stage strategic decision and a second-stage tactical decision. The first-stage strategic decision includes parcel locker location parameters and drone configuration parameters. An adaptive large neighborhood search algorithm is used to iteratively search the parcel locker location parameters and UAV configuration parameters within the first stage of strategic decision-making. When a new solution is generated in each search iteration, the new solution is input into a pre-trained deep neural network model. The truck replenishment cost corresponding to the new solution in the heterogeneous demand characteristic scenario set is inferred by a deep neural network model. The truck replenishment cost belongs to the second stage tactical decision. The truck replenishment cost is fed back as a fitness value to the adaptive large neighborhood search algorithm, and the operator weights of the adaptive large neighborhood search algorithm are updated. When the adaptive large neighborhood search algorithm reaches the preset termination condition, iterative search stops and the final cabinet collaborative delivery network planning scheme is output.
2. The stochastic planning method for cabinet-based collaborative delivery networks according to claim 1, characterized in that, The objective function of the two-stage stochastic programming model is obtained by adding the fixed construction cost and the expected operating cost; Fixed construction costs include the construction cost of the parcel lockers and the cost of purchasing drones; Expected operating costs include drone delivery costs, truck transportation costs, and penalty costs arising from demand fluctuations; The constraints of the two-stage stochastic programming model include the capacity limit of the parcel locker service, the divisibility of customer demand, and the service range constraint of the drone.
3. The stochastic planning method for cabinet-based collaborative delivery networks according to claim 2, characterized in that, The process of constructing a two-stage stochastic programming model includes: A sample mean approximation framework is adopted, and independent and identically distributed requirement scenario samples are generated based on Monte Carlo sampling. The stochastic programming problem is transformed into a deterministic equivalent problem, and a statistical estimate of the objective function value is calculated to obtain a statistical upper bound. Calculate the mean of the optimal values of multiple independent approximation problems to obtain a statistical lower bound; The outer iteration of the sample mean approximation framework is terminated when the difference between the statistical upper bound and the statistical lower bound is less than a preset gap value.
4. The stochastic planning method for cabinet-based collaborative delivery networks according to claim 1, characterized in that, Deep neural network models include encoder modules, decoder modules, and masking mechanisms; The encoder module maps the coordinates and normalized requirement features of the parcel lockers, distribution centers, and customer nodes into high-dimensional embedding vectors. The decoder module combines the context vector and the masking mechanism to output the probability distribution of the truck visiting the next node; The masking mechanism blocks access to nodes and nodes that violate capacity constraints.
5. The stochastic planning method for cabinet-based collaborative delivery networks according to claim 4, characterized in that, The processing steps of a deep neural network model include: Within the encoder module, a multi-head attention layer is used to extract the spatial and demand features of the parcel locker nodes and the distribution center, and generate corresponding node embedding vectors. Inside the decoder module, the node embedding vector and the current vehicle state are input into the bidirectional gating recurrent unit network, and the probability distribution of the truck accessing the next node is output based on the bidirectional gating recurrent unit network.
6. The stochastic planning method for cabinet-based collaborative delivery networks according to claim 1, characterized in that, The adaptive large neighborhood search algorithm includes both destruction and repair operators; The destruction operators include random removal operators, worst-cost removal operators, correlation removal operators, and drone number adjustment removal operators; The repaired operators cover the greedy insertion operator, the regret insertion operator, and the drone redistribution operator.
7. The stochastic planning method for cabinet-based collaborative delivery networks according to claim 6, characterized in that, The iterative search process using the adaptive large neighborhood search algorithm includes: Based on the historical call to the operator's performance, some of the selected parcel lockers can be removed from the current solution or the configuration parameters of the drone can be reduced. The repair operator is invoked to re-insert the unselected candidate points into the current solution and generate a new solution to be evaluated, which includes new location parameters and UAV configuration parameters. The total cost of the new solution to be evaluated is calculated by combining the truck replenishment cost, and the acceptance of the new solution to be evaluated is determined by the simulated annealing criterion.
8. A stochastic programming system for a cabinet-based collaborative delivery network, characterized in that, include: The scene generation module is used to obtain delivery parameters for the target delivery area and generate multiple sets of heterogeneous demand feature scene sets. The model building module is used to construct a two-stage stochastic programming model based on the set of multiple heterogeneous demand characteristic scenarios, which includes a first-stage strategic decision and a second-stage tactical decision. The first-stage strategic decision includes parcel locker location parameters and drone configuration parameters. The search iteration module is used to iteratively search the parcel locker location parameters and UAV configuration parameters within the first stage strategic decision using an adaptive large neighborhood search algorithm, and inputs the new solution into a pre-trained deep neural network model when a new solution is generated in each search iteration. The cost inference module is used to infer the truck replenishment cost corresponding to the new solution under the heterogeneous demand feature scenario set through a deep neural network model. The truck replenishment cost belongs to the second stage tactical decision. The weight update module is used to feed the truck replenishment cost as a fitness value back to the adaptive large neighborhood search algorithm and update the operator weights of the adaptive large neighborhood search algorithm. The result output module is used to stop the iterative search and output the final cabinet collaborative delivery network planning scheme when the adaptive large neighborhood search algorithm reaches the preset termination condition.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the stochastic planning method for cabinet collaborative delivery network as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the stochastic planning method for the cabinet collaborative delivery network as described in any one of claims 1 to 7.
