Large-scale dynamic logistics scheduling method, system and electronic device

Abstract

The application discloses a large-scale dynamic logistics scheduling method and system and electronic equipment, receives order, node, vehicle, capacity, time window and task priority data in a logistics scheduling system, reduces a vehicle path planning problem into a QUBO quadratic coefficient matrix Q, performs low-rank and sparse joint representation on Q to obtain a low-rank component L and a sparse component S, constructs a compressed QUBO solving model based on L and S, obtains a scheduling sequence on a parallel computing platform or a QUBO heuristic solver, and performs local incremental updating and re-optimization on only the low-rank factor and the sparse triple affected by dynamic logistics events such as order insertion, order cancellation, batch order change, vehicle failure, vehicle type switching or time window change. The application can reduce storage and computing overhead in a large-scale logistics scheduling scenario and improve re-optimization efficiency under dynamic events.

Description

Technical Field

[0001] This invention relates to the fields of intelligent logistics scheduling, vehicle routing planning, warehouse AGV scheduling, port vehicle scheduling, combinatorial optimization, QUBO modeling, low-rank matrix decomposition, sparse matrix computation, parallel optimization computation, and logistics information system integration technology, specifically large-scale dynamic logistics scheduling methods, systems, and electronic equipment. Background Technology

[0002] Vehicle routing is a core problem in modern logistics scheduling. In scenarios with large order volumes, numerous vehicles, dense time windows, complex task priorities, and frequently changing road network constraints, scheduling systems need to quickly generate task sequences for vehicles, AGVs, or other transportation units. Common constraints include vehicle capacity, service time windows, task priorities, node uniqueness, road traffic restrictions, warehouse aisle occupancy, local traffic restrictions, and dynamic anomalies.

[0003] In recent years, reducing the vehicle routing problem to a QUBO model and solving it using digital annealing, simulated bifurcation, quantum annealing, GPU parallel simulated annealing, or other heuristic solvers has become an important approach in combinatorial optimization engineering. A typical QUBO objective function describes the problem's objective and constraints using a quadratic coefficient matrix, and the solver iteratively searches for the binary decision vector that minimizes the objective function.

[0004] Existing publicly available technologies can be broadly categorized into four types. The first type involves candidate path generation and clustering to compress routes, such as candidate path pre-generation and optimization problem clustering for AGV operating systems. The main object of compression is the set of feasible paths or the candidate solution space. This type of method reduces the number of paths that need to be considered during the solution process by pre-generating a large number of possible paths and then clustering or filtering them. However, when new orders are inserted, orders are canceled, vehicles break down, roads become congested, or time windows change, it is usually necessary to regenerate candidate paths or re-execute clustering, making it difficult to directly reuse the original problem structure, resulting in low efficiency for dynamic updates.

[0005] The second category involves direct QUBO modeling followed by computational annealing, simulated bifurcation, quantum annealing, or GPU-inspired solvers, with a core focus on the solver's search efficiency. This type of method directly constructs the VRP problem as a QUBO model and then processes it using a dedicated solver. However, for medium- to large-scale logistics scheduling scenarios, the size of the Q matrix increases quadratically with the number of nodes and vehicles. Explicitly storing the complete dense Q matrix and frequently calculating matrix-vector multiplications during solver iterations leads to a sharp increase in memory usage and computational latency, becoming the main performance bottleneck of the system.

[0006] The third category includes heuristic approaches such as traditional ALNS, LKH, genetic algorithms, and tabu search, which typically search directly in the path or neighborhood space. These methods perform well on small to medium-sized problems, but are prone to getting trapped in local optima on large-scale problems and struggle to effectively handle complex multi-constraint conditions and dynamic events.

[0007] The fourth category involves mathematical theories such as general low-rank matrices, low-rank tensors, and RobustPCA. While these approaches can explain some structural features, they typically do not provide an engineering closed loop within the logistics scheduling system. This type of research primarily focuses on the theoretical properties of the mathematical model itself, without addressing practical engineering aspects such as order processing, WMS / WCS / TMS interfaces, AGV or vehicle execution loops, dynamic anomaly monitoring, local re-optimization, and scheduling result distribution. Consequently, it is difficult to directly apply these approaches to real-world logistics systems. Summary of the Invention

[0008] The purpose of this invention is to provide a method, system, and electronic device for large-scale dynamic logistics scheduling to solve the problems existing in the prior art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a large-scale dynamic logistics scheduling method, comprising the following steps: Step 1: Receive order data, node data, vehicle data, capacity constraint data, time window constraint data, and task priority data from the logistics scheduling system; Step 2: Based on the received data, the vehicle routing problem is reduced to a QUBO quadratic coefficient matrix Q, and a QUBO objective function is constructed; Step 3: Perform a low-rank and sparse joint representation on the QUBO quadratic coefficient matrix Q, decomposing Q into a low-rank component L, a sparse component S, and a controllable residual E; Step 4: Construct a compressed QUBO solution model based on the low-rank component L and sparse component S, and solve it on a parallel computing platform or QUBO heuristic solver to obtain the vehicle or AGV scheduling sequence. Step 5: Output the scheduling sequence to the logistics scheduling system;

[0010] Step 6: When a dynamic logistics event is received, identify the affected nodes, vehicles, time windows and constraints, perform local incremental updates on the low-rank component L and sparse component S, and perform re-optimization based on the updated L and S.

[0011] Furthermore, the QUBO quadratic coefficient matrix Q consists of at least two of the following: a distance term, a node uniqueness constraint term, a capacity constraint term, a time window constraint term, and a task priority constraint term; the calculation equation for the QUBO objective function is: In the formula, For a binary decision vector, It is a quadratic coefficient matrix. The coefficient of the linear term, This is a constant term.

[0012] Furthermore, in step three, the computational equation for the joint low-rank and sparse representation is: In the formula, As a low-rank component, it is used to express globally relevant structures related to distance, capacity, and task priority; As a sparse component, it is used to express node uniqueness constraints, local time window constraints, local forbidden constraints, or abnormal event constraints. This is a controllable residual term.

[0013] Furthermore, in step three, the low-rank component From the factor matrix Singular value matrix or weight matrix and factor matrix The calculation equation is as follows: In the formula, and A low-dimensional relational structure used to express relationships between nodes, vehicles, time slices, or task priorities. Used to express the weight of the main direction.

[0014] Furthermore, in step three, the sparse component It is represented by a list of sparse element triples, which includes at least a row index, a column index, and a corresponding non-zero coefficient.

[0015] Furthermore, in step four, the complete dense Q matrix is ​​not explicitly reconstructed during the solution process. Instead, the calculation is performed using structured matrix-vector multiplication, and the calculation equation is as follows: In the formula, For a binary decision vector, It is a low-rank factor matrix. It is a singular value matrix or a weight matrix. It is a low-rank factor matrix. It is a sparse component.

[0016] Furthermore, the computational cost of the structured matrix-vector multiplication operation is similar to the rank of the low-rank component. Variable size and the number of non-zero elements in the sparse component The relevant computational cost expression is as follows: .

[0017] Furthermore, in step six, the local perturbations of the low-rank components are updated using the Sherman-Morrison formula, the calculation equation of which is: In the formula, This is the original low-rank matrix. and This is the perturbation vector.

[0018] A large-scale dynamic logistics scheduling system includes a data access module, a QUBO construction module, a low-rank sparse decomposition module, a compression solution module, a dynamic event monitoring module, a local incremental update module, and a scheduling scheme output module. Each module is used to execute the large-scale dynamic logistics scheduling method described above.

[0019] An electronic device includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the large-scale dynamic logistics scheduling method.

[0020] Compared with the prior art, the beneficial effects of the present invention are: This invention performs a low-rank and sparse joint representation on the QUBO coefficient matrix Q after VRPTW reduction, decomposing Q into a low-rank component L and a sparse component S. This fully utilizes the structural characteristics of the logistics scheduling problem and significantly reduces storage overhead in large-scale logistics scheduling scenarios. The low-rank component L is represented by a factor matrix, and the sparse component S is stored in triplet form, which saves a significant amount of storage space compared to the complete dense Q matrix.

[0021] This invention achieves local incremental re-optimization by updating only the affected low-rank factors and sparse triples when dynamic logistics events occur. Compared with traditional global reconstruction methods, local incremental updates can significantly reduce computational load, improve the response speed of dynamic events, and enable the scheduling system to quickly adapt to changes in the logistics environment.

[0022] This invention embeds low-rank and sparse structures into the complete engineering closed loop of a logistics scheduling system, including order entry, QUBO construction, compressed solution, dynamic event monitoring, local re-optimization, and scheduling result output. Compared to purely mathematical theoretical research, this invention provides a complete solution that can be directly applied to practical logistics systems, demonstrating strong engineering practicality.

[0023] The compressed QUBO solution model of this invention is compatible with a variety of existing solvers, including GPU parallel simulated annealing, digital annealing, simulated bifurcation, tabu search, adaptive large neighborhood search, and quantum annealing solvers. Users can choose the appropriate solver according to their hardware conditions and solution requirements without modifying the core algorithm, which has strong flexibility and scalability.

[0024] This invention uses the low-rank component L to express globally relevant structures such as distance, capacity, and task priority, and the sparse component S to express local constraints such as node uniqueness, time windows, local forbidden zones, and abnormal events, which have clear physical meaning and interpretability. This interpretability makes the system easier to debug and optimize, and also easier for users to understand and accept. Attached Figure Description

[0025] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the low-rank and sparse decomposition of the QUBO coefficient matrix; Figure 3 This is a diagram illustrating a partial update of a dynamic event. Figure 4 This is a schematic diagram of the structured Qx computation simulation results. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0027] Please see Figure 1-4 The large-scale dynamic logistics scheduling method of the present invention includes the following steps: Step 1: Receive order data, node data, vehicle data, capacity constraint data, time window constraint data, and task priority data from the logistics scheduling system.

[0028] The system interacts with various logistics scheduling systems through standard interfaces to receive real-time logistics scheduling data. Order data includes, but is not limited to, order number, corresponding customer node, order weight, order volume, order priority, and special requirements. Node data includes, but is not limited to, the node's unique identifier, geographic coordinates, node type (e.g., customer address, warehouse location, port storage location, AGV task point, low-altitude logistics take-off and landing point), service time, and special restrictions. Vehicle data includes, but is not limited to, the vehicle's unique identifier, vehicle type, current location, remaining load capacity, remaining volume, current status (e.g., idle, driving, in service, faulty), energy type, and driving range. Capacity constraint data includes, but is not limited to, the vehicle's maximum load capacity, maximum volume, and maximum travel distance. Time window constraint data includes, but is not limited to, the earliest service time, latest service time, service duration, and time window flexibility for each node. Task priority data includes, but is not limited to, the priority weights of different task types and the handling rules for urgent tasks.

[0029] The system preprocesses the received data, including data cleaning, data format conversion, and data validation, to ensure the accuracy and completeness of the data. For missing data, the system fills in the missing data with default values ​​or estimates based on historical data; for abnormal data, the system marks the data and prompts for manual intervention.

[0030] Step 2: Based on the received data, the vehicle routing problem is reduced to a QUBO quadratic coefficient matrix Q, and a QUBO objective function is constructed.

[0031] Based on the received logistics scheduling data, the system transforms the Vehicle Routing Problem with Time Window (VRPTW) into a QUBO quadratic unconstrained binary optimization problem. In one implementation, the system uses a node-vehicle binary variable representation, defining binary variables... This indicates that node i is served by vehicle k, where i = 1, 2, ..., N represents the node number, and k = 1, 2, ..., K represents the vehicle number. In another implementation, the system uses a ternary variable representation of node-time slice-vehicle, defining a binary variable. This indicates that node i is served by vehicle k in time slice t, where t = 1, 2, ..., T represents the time slice number. This notation can describe time window constraints more accurately, but it increases the number of variables.

[0032] The system constructs a QUBO objective function based on distance, capacity, time window, node uniqueness, and task priority. The equation for calculating the QUBO objective function is: In the formula, For a binary decision vector, It is a quadratic coefficient matrix. The coefficient of the linear term, This is a constant term.

[0033] The distance term minimizes the total distance traveled by the vehicle, and its corresponding Q-matrix sub-block contains distance information between nodes. The capacity term ensures that the vehicle's load and volume do not exceed its maximum capacity, and its corresponding Q-matrix sub-block contains a penalty term for vehicle capacity constraints. The time window term ensures that the vehicle arrives at and serves nodes within a specified time window, and its corresponding Q-matrix sub-block contains a penalty term for time window constraints. The node uniqueness term ensures that each node can only be served by one vehicle once, and its corresponding Q-matrix sub-block contains a penalty term for node uniqueness constraints. The task priority term prioritizes high-priority orders, and its corresponding Q-matrix sub-block contains weight information for task priorities.

[0034] The system combines the above items to form a complete QUBO quadratic coefficient matrix Q and linear coefficients q. The constant term c is used to adjust the baseline value of the objective function and can usually be set to 0. When constructing the Q matrix, the system allocates components with strong global relevance (such as distance, capacity, and task priority terms) to the low-rank component L, and components with local sparsity (such as node uniqueness, time window, and local event constraint terms) to the sparse component S. This allocation method can fully utilize the structural characteristics of the logistics scheduling problem and lay the foundation for the subsequent joint low-rank and sparse representation.

[0035] Step 3: Perform a low-rank and sparse joint representation on the QUBO quadratic coefficient matrix Q, decomposing Q into a low-rank component L, a sparse component S, and a controllable residual E.

[0036] The system performs low-rank and sparse joint decomposition on the constructed QUBO quadratic coefficient matrix Q, obtaining the low-rank component L, the sparse component S, and the controllable residual E. The computational equation for the low-rank and sparse joint representation is as follows: In the formula, As a low-rank component, it is used to express globally relevant structures related to distance, capacity, and task priority; As a sparse component, it is used to express node uniqueness constraints, local time window constraints, local forbidden constraints, or abnormal event constraints. This is a controllable residual term.

[0037] This invention does not require Q to be strictly low-rank globally. For scenarios where logistics nodes are distributed in two-dimensional or three-dimensional space, and orders and tasks exhibit regional clustering characteristics, distance-related terms, capacity outer product terms, or task priority terms typically present a compressible, approximately low-rank structure. This is because the distance matrix has low-rank properties in two-dimensional or three-dimensional space, and capacity constraints and task priorities usually have global correlation. Node uniqueness, time windows, local prohibitions, and dynamic event constraints typically exhibit sparse structures because these constraints involve interactions between only a few variables.

[0038] The low-rank component L is represented by the factor matrix U, the singular value matrix or weight matrix, and the factor matrix V. Its calculation equation is as follows: In the formula, and A low-dimensional relational structure used to express relationships between nodes, vehicles, time slices, or task priorities. Used to express the weight of the main direction. and The number of columns is equal to the rank r of the low-rank component, and r is much smaller than the dimension n of the Q matrix. With this representation, the dense matrix L that originally required storing n elements now only needs to store rn + rn + r = r(2n + 1) elements, greatly reducing the storage overhead.

[0039] The system can employ various algorithms to obtain the low-rank component L, including but not limited to stochastic singular value decomposition, principal component tracking, GoDec, online RobustPCA, incremental SVD, Tucker decomposition, or HOSVD. Different algorithms are suitable for different scenarios and requirements. For example, stochastic singular value decomposition is computationally fast and suitable for large-scale matrices; principal component tracking can simultaneously obtain low-rank and sparse components, suitable for situations with high noise; and online RobustPCA can handle streaming data and is suitable for dynamically updated scenarios.

[0040] The sparse component S is stored in sparse triplet form, where each triplet includes a row index, a column index, and a non-zero coefficient. Since the sparse component S contains only a few non-zero elements, storing it in sparse triplet form significantly reduces storage overhead. For example, for an n n matrix, if the sparse component S has only nnz(S) non-zero elements, then only 3nnz(S) elements need to be stored, instead of n n elements. The sparse component S can quickly add, delete, or modify non-zero elements based on dynamic events without reconstructing the complete Q matrix, which facilitates local incremental updates for dynamic events.

[0041] The controllable residual term E represents the error of the joint low-rank and sparse representation. In practical applications, the system can choose to retain or ignore E based on accuracy and storage requirements. If high accuracy is required, E can be retained and its impact considered during the solution process; if high storage and computational speed are required, E can be ignored, and only L and S are used for approximate calculations. In the simulation embodiment, ignoring E did not result in a significant loss of solution quality.

[0042] Step 4: Construct a compressed QUBO solution model based on the low-rank component L and sparse component S, and solve it on a parallel computing platform or a QUBO heuristic solver to obtain the vehicle or AGV scheduling sequence.

[0043] The system constructs a compressed QUBO solution model based on the low-rank component L and the sparse component S. The compressed QUBO solution model differs from the traditional QUBO model in that it does not use the complete dense Q matrix, but instead uses low-rank factors U, V, and sparse triples S to represent the problem structure. This representation significantly reduces the solver's storage and computational overhead.

[0044] The compressed QUBO solution model can be solved by various solvers, including but not limited to GPU parallel simulated annealing, digital annealing, simulated bifurcation, tabu search, adaptive large neighborhood search, quantum annealing solvers, or other parallel heuristic solvers. The system can select an appropriate solver based on the problem size, hardware conditions, and solution requirements. For example, for large-scale problems, GPU parallel simulated annealing or digital annealing solvers can be used to fully utilize parallel computing capabilities; for problems requiring high-precision solutions, adaptive large neighborhood search solvers can be used; and for scenarios with quantum computing hardware, quantum annealing solvers can be used.

[0045] During solver iteration, matrix-vector multiplication Qx is often required. For large-scale problems, explicitly constructing and using the complete dense Q matrix incurs significant storage and computational overhead. This invention performs equivalent or approximate calculations based on an L+S structure, without explicitly reconstructing the complete dense Q matrix. Instead, it performs matrix-vector multiplication operations according to a structured computational order. The computational equation for the structured matrix-vector multiplication operation is as follows: In the formula, For a binary decision vector, It is a low-rank factor matrix. It is a singular value matrix or a weight matrix. It is a low-rank factor matrix. It is a sparse component.

[0046] The advantage of the above calculation order is that it reduces the original O(n) time complexity to O(n^2). 2 The complexity of matrix-vector multiplication is reduced to... Complexity. The computational cost of structured matrix-vector multiplication is related to the rank r of the low-rank component, the size n of the variable, and the number of non-zero elements in the sparse component. The relevant computational cost expression is as follows: Since r is much smaller than n, and nnz(S) is usually much smaller than n, structured computation can significantly reduce computational overhead and improve solution speed.

[0047] It is important to note that the above formula only illustrates that low-rank terms and sparse terms participate in the calculation together, and does not limit the specific calculation order. Calculating Sx first, calculating low-rank terms first, parallel calculation of low-rank and sparse terms, batch processing, GPU parallelization, or multi-core CPU parallelization are all equivalent implementation methods. The system can select the optimal calculation order and parallelization strategy based on the hardware architecture and computing resources to further improve computational efficiency.

[0048] The solver finds the binary decision vector x that minimizes the QUBO objective function through iterative search. When the solver reaches the convergence condition or the maximum number of iterations, it stops iterating and outputs the optimal solution. The system transforms the optimal solution into a vehicle or AGV scheduling sequence, including information such as vehicle number, task node, execution order, estimated arrival time, service time window, scheduling confidence, and local re-optimization flags.

[0049] Step 5: Output the scheduling sequence to the logistics scheduling system.

[0050] The system outputs the generated scheduling sequences to the corresponding logistics scheduling system through a standard interface. Scheduling results can be output to WMS (Warehouse Management System), WCS (Warehouse Control System), TMS (Transportation Management System), warehouse AGV scheduling systems, port vehicle scheduling systems, urban distribution scheduling systems, or low-altitude logistics scheduling systems. Different logistics scheduling systems require scheduling data in different formats; the system can convert the scheduling sequences to the appropriate format according to the requirements of the target system.

[0051] For example, for a warehouse AGV scheduling system, the system outputs a scheduling sequence including AGV number, task point number, task type (such as pickup, delivery, charging, etc.), task execution order, and estimated arrival time. For an urban delivery scheduling system, the system outputs a scheduling sequence including vehicle number, delivery station number, delivery order, estimated arrival time, and cargo list.

[0052] After receiving a scheduling sequence, the logistics dispatching system distributes it to the corresponding vehicle or AGV for execution. During execution, the vehicle or AGV provides real-time feedback on its status, such as task start, task completion, or encountering an exception. The system receives this feedback and adjusts its subsequent dispatching strategy accordingly. For example, if a vehicle encounters traffic congestion while performing a task, the system can adjust the scheduling sequence of other vehicles to ensure overall dispatching efficiency; if the priority of an order changes, the system can readjust the task execution order.

[0053] Execution feedback can also be used to adjust the weights and penalty coefficients in the QUBO model. Based on historical execution data, the system can learn the optimal weights for distance, capacity, time window, task priority, and dynamic event local update ranges to continuously improve scheduling quality and system performance. This self-learning capability enables the system to adapt to different logistics scenarios and business needs, continuously optimizing scheduling effectiveness.

[0054] Step 6: When a dynamic logistics event is received, identify the affected nodes, vehicles, time windows and constraints, perform local incremental updates on the low-rank component L and sparse component S, and perform re-optimization based on the updated L and S.

[0055] Logistics scheduling often encounters various dynamic events, such as order insertion, order cancellation, batch order modification, batch order import or cancellation, vehicle malfunction, vehicle type switching, energy type switching, road congestion, warehouse access blockage, changes in vehicle access areas, time window changes, and upstream anomaly prediction events. These dynamic events can render the original scheduling sequence infeasible and require re-optimization.

[0056] Traditional re-optimization methods typically require reconstructing the entire QUBO matrix and solving it again, which leads to significant computational overhead and long response times. This invention employs a local incremental update method, updating only the portion affected by dynamic events instead of reconstructing the entire Q matrix, thereby significantly improving re-optimization efficiency.

[0057] When the system receives a dynamic logistics event, it first analyzes the event type and scope of impact, identifying the affected nodes, vehicles, time windows, and constraints. Then, based on the event type, it performs corresponding local update operations on the low-rank component L and the sparse component S.

[0058] For order insertion, order cancellation, or batch order modification events, the system updates the rows, columns, sparse constraint terms, and low-rank factors related to the corresponding nodes in the QUBO quadratic coefficient matrix. Specifically, for newly inserted orders, the system adds the corresponding rows and columns to the Q matrix and adds the relevant constraint terms to the sparse component S; simultaneously, it updates the factor matrix of the low-rank component L to reflect the distance and correlation between the new node and other nodes. For cancelled orders, the system removes the corresponding constraint terms from the sparse component S and updates the factor matrix of the low-rank component L to remove the influence of the cancelled nodes.

[0059] For events such as vehicle malfunction, vehicle type switching, energy type switching, or capacity constraint changes, the system locally updates the low-rank component L based on the low-rank perturbation update method and performs addition, deletion, and modification operations on relevant triples in the sparse component S. The low-rank perturbation update method can quickly update the low-rank factor matrix without recalculating the entire low-rank decomposition. For local perturbations of the low-rank component, the system can use the Sherman-Morrison formula for updating, the calculation equation of which is: In the formula, This is the original low-rank matrix. and Let L be the perturbation vector. The Sherman-Morrison formula is applicable to rank 1 perturbations. For higher-rank perturbations, the Woodbury formula or other low-rank update methods can be used. Furthermore, the system can also use incremental SVD, Brand-class incremental SVD, and other methods to update the low-rank component L.

[0060] The system updates the sparse component S by performing add, delete, and modify operations on the triplet list. For example, when a storage channel is blocked, the system adds the corresponding prohibition constraint to the sparse component S; when a time window changes, the system modifies the coefficient of the corresponding time window constraint in the sparse component S. Because the sparse component S is stored in triplet form, these add, delete, and modify operations are highly efficient and do not require reconstructing the entire matrix.

[0061] After the update, the system feeds the updated L and S into the compressed QUBO solution model for local re-optimization. Local re-optimization only requires searching within the affected subspace, rather than re-searching the entire solution space, thus significantly improving the re-optimization speed. The system outputs a new scheduling sequence and sends it to the logistics scheduling system for execution.

[0062] This invention is preferably applicable to medium-to-large-scale logistics scheduling scenarios, such as when the number of nodes N is not less than 50, the number of vehicles K is not less than 10, and the reduced variable size n is not less than 500, the storage and computational advantages brought by structured representation are more easily demonstrated. The above scale conditions are only preferred embodiments and do not limit the scope of protection of this invention. In small-scale scenarios, this invention can also work normally, but the advantages may not be as obvious as in large-scale scenarios.

[0063] In the simulation embodiments, the present invention uses Python numerical simulation to verify the technical solution. The simulation results are only used to illustrate the technical effect and do not represent the field test results. The simulation includes structured Qx computation acceleration, storage compression, dynamic event local update, and numerical consistency testing. In three logistics scenarios (urban delivery, warehouse AGV, and port short-haul), L+S indicates that it exhibits a higher storage compression rate and computation speedup compared to the complete Q matrix. In the warehouse AGV scenario, as n increases, the proportion of non-zero elements in the sparse component S increases, leading to a relative decrease in the Qx speedup. This phenomenon indicates that the acceleration effect of the present invention is affected by the rank value of the low-rank component, the density of the sparse component, and the scenario constraint structure. In the dynamic event local update test, the local update method of the present invention shows a significant acceleration advantage compared to the complete reconstruction method. In the numerical consistency test, the numerical error of the structured Qx computation is very small, and no significant solution quality loss was observed in the simulation embodiments.

[0064] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A large-scale dynamic logistics scheduling method, characterized by: Includes the following steps: Step 1: Receive order data, node data, vehicle data, capacity constraint data, time window constraint data, and task priority data from the logistics scheduling system; Step 2: Based on the received data, the vehicle routing problem is reduced to a QUBO quadratic coefficient matrix Q, and a QUBO objective function is constructed; Step 3: Perform a low-rank and sparse joint representation on the QUBO quadratic coefficient matrix Q, decomposing Q into a low-rank component L, a sparse component S, and a controllable residual E; Step 4: Construct a compressed QUBO solution model based on the low-rank component L and sparse component S, and solve it on a parallel computing platform or QUBO heuristic solver to obtain the vehicle or AGV scheduling sequence. Step 5: Output the scheduling sequence to the logistics scheduling system; Step 6: When a dynamic logistics event is received, identify the affected nodes, vehicles, time windows and constraints, perform local incremental updates on the low-rank component L and sparse component S, and perform re-optimization based on the updated L and S.

2. The large-scale dynamic logistics scheduling method according to claim 1, characterized in that: In step two, the QUBO quadratic coefficient matrix Q consists of at least two of the following: a distance term, a node uniqueness constraint term, a capacity constraint term, a time window constraint term, and a task priority constraint term; the calculation equation for the QUBO objective function is: In the formula, For a binary decision vector, It is a quadratic coefficient matrix. The coefficient of the linear term, This is a constant term.

3. The large-scale dynamic logistics scheduling method according to claim 1, characterized in that: In step three, the computational equation for the joint low-rank and sparse representation is: In the formula, As a low-rank component, it is used to express globally relevant structures related to distance, capacity, and task priority; As a sparse component, it is used to express node uniqueness constraints, local time window constraints, local forbidden constraints, or abnormal event constraints. This is a controllable residual term.

4. The large-scale dynamic logistics scheduling method according to claim 1, characterized in that: In step three, the low-rank component From the factor matrix Singular value matrix or weight matrix and factor matrix The calculation equation is as follows: In the formula, and A low-dimensional relational structure used to express relationships between nodes, vehicles, time slices, or task priorities. Used to express the weight of the main direction.

5. The large-scale dynamic logistics scheduling method according to claim 1, characterized in that: In step three, the sparse component It is represented by a list of sparse element triples, which includes at least a row index, a column index, and a corresponding non-zero coefficient.

6. The large-scale dynamic logistics scheduling method according to claim 1, characterized in that: In step four, the complete dense Q matrix is ​​not explicitly reconstructed during the solution process. Instead, the calculation is performed using structured matrix-vector multiplication, and the calculation equation is as follows: In the formula, For a binary decision vector, It is a low-rank factor matrix. It is a singular value matrix or a weight matrix. It is a low-rank factor matrix. It is a sparse component.

7. The large-scale dynamic logistics scheduling method according to claim 6, characterized in that: The computational cost of the structured matrix-vector multiplication operation and the rank of the low-rank component are related. Variable size and the number of non-zero elements in the sparse component The relevant computational cost expression is as follows: 。 8. The large-scale dynamic logistics scheduling method according to claim 1, characterized in that: In step six, the local perturbations of the low-rank components are updated using the Sherman-Morrison formula, the calculation equation of which is: In the formula, This is the original low-rank matrix. and This is the perturbation vector.

9. A large-scale dynamic logistics scheduling system, characterized in that, It includes a data access module, a QUBO construction module, a low-rank sparse decomposition module, a compression solution module, a dynamic event monitoring module, a local incremental update module, and a scheduling scheme output module. Each module is used to execute the large-scale dynamic logistics scheduling method according to any one of claims 1 to 8.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the large-scale dynamic logistics scheduling method according to any one of claims 1 to 8.

Images