An AI-based supply chain scheduling method and system

Abstract

The application relates to the technical field of supply chain scheduling, and discloses a supply chain scheduling method and system based on AI, which comprises the following steps: acquiring order, resource, logistics topology data and historical performance data of a supply chain; generating a delivery uncertainty embedding vector by using a graph neural network; determining the center time and slope of an asymmetric trapezoidal function and the mean and variance of a Gaussian function according to the delivery uncertainty embedding vector, and constructing a fuzzy time window for a scheduling task of each destination node; acquiring a reinforcement learning model, and outputting a weight combination of a cost term, a time term and a robustness term in a scheduling target function; determining the scheduling target function according to the weight combination, and solving the scheduling target function; and performing cross operation of a genetic algorithm, wherein the cross operation is key subgraph exchange, and the key path subgraphs of two parent scheduling schemes are identified and exchanged to realize the cross operation. The application can not only accurately depict complex performance risk distribution, but also adapt to real-time changes, and the calculation amount is reduced.

Description

Technical Field

[0001] This invention relates to the field of supply chain scheduling technology, specifically to an AI-based supply chain scheduling method and system. Background Technology

[0002] Supply chain scheduling aims to reduce operating costs and improve order delivery efficiency by optimizing resource allocation across multiple stages, including order processing, production, warehousing, and logistics. This is a typical combinatorial optimization problem, and its inherent NP-hard nature means that the time required to find the optimal solution increases exponentially with the problem size. Therefore, while traditional mathematical programming scheduling methods can theoretically find the optimal solution, they are often difficult to solve in a timely manner in practical applications due to excessive computational complexity.

[0003] Faced with this challenge, researchers have turned to supply chain scheduling methods employing metaheuristic algorithms, such as genetic algorithms and simulated annealing. These metaheuristic algorithms can quickly find near-optimal scheduling schemes, but the quality of the found schemes is highly dependent on how the problem itself is encoded. However, the modern supply chain environment is fraught with uncertainty; for example, random events such as transportation delays and equipment failures can severely impact the stability and reliability of the found scheduling schemes. To address these temporal uncertainties, the fuzzy time window method has been introduced, using a membership function to represent the satisfaction level of task completion time. Nevertheless, the fuzzy time window still has limitations; it often employs symmetric membership functions, making it difficult to accurately characterize complex performance risk distributions.

[0004] Furthermore, existing supply chain scheduling methods generally lack the ability to dynamically adjust based on real-time status, making it difficult for them to find the globally optimal scheduling solution in a dynamic operating environment. Summary of the Invention

[0005] This invention provides an AI-based supply chain scheduling method and system to solve the problems of difficulty in accurately depicting the distribution of complex performance risks, difficulty in adapting to real-time changes, and large computational load in the existing technologies.

[0006] In a first aspect, the AI-based supply chain scheduling method of the present invention includes the following steps:

[0007] Acquire order, resource, logistics topology data and historical fulfillment data of the supply chain; use graph neural networks to process the logistics topology data and historical fulfillment data to generate delivery uncertainty embedding vectors for the destination nodes of each scheduling task;

[0008] Based on the components of each dimension of the delivery uncertainty embedding vector, the center time and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function are determined. A composite fuzzy time window is constructed for the scheduling task of each destination node. The membership function of the fuzzy time window is obtained by the weighted sum of the asymmetric trapezoidal function and the Gaussian function.

[0009] Obtain the reinforcement learning model. The action space of the reinforcement learning model is a multi-dimensional continuous space, which is used to output the weight combination of cost, time and robustness terms in the scheduling objective function. The reward function of the reinforcement learning model includes a performance penalty term calculated based on a fuzzy time window and a stability penalty term representing the structural difference between scheduling schemes in previous and subsequent scheduling cycles.

[0010] The scheduling objective function is determined based on the weight combination, and the scheduling objective function is solved using a genetic algorithm. The crossover operation of the genetic algorithm is a critical subgraph exchange, which is achieved by identifying and exchanging the critical path subgraphs in the directed acyclic graph of the task dependencies of the two parent scheduling schemes.

[0011] Preferably, the mutation operation of the genetic algorithm is as follows: select a task corresponding to the high-norm delivery uncertainty embedding vector, match alternative resource paths for the task, the cost increment of the matched alternative resource paths is constrained by an upper limit, and the upper limit decreases as the membership value of the task in the original scheduling scheme increases.

[0012] Preferably, the step of determining the center time and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function based on the components of each dimension of the delivery uncertainty embedding vector, and constructing a composite fuzzy time window for the scheduling task of each destination node, wherein the membership function of the fuzzy time window is obtained by the weighted sum of the asymmetric trapezoidal function and the Gaussian function, includes: the membership function of the fuzzy time window. The calculation formula is: ;in, It is an asymmetric trapezoidal function. Let w be a Gaussian function and w be the weights; embed the vector using delivery uncertainty. The first five components are used to determine the parameters of the membership function through the following transformation: the central time of the asymmetric trapezoidal function. The left slope of an asymmetric trapezoidal function The right-hand slope of an asymmetric trapezoidal function Mean of Gaussian function Variance of Gaussian function ;in to and to These are the preset transformation coefficients.

[0013] Preferably, the reward function of the reinforcement learning model is calculated as follows: for each task i in the scheduling scheme, the planned completion time of the task is obtained. And calculate the membership value based on the composite morphological fuzzy time window of the destination node of the task. ; Calculate performance penalties : The process involves summing all tasks in the current scheduling scheme; comparing the resources allocated to each task in the current scheduling scheme with those in the previous scheduling cycle, and counting the number of tasks with resource changes. ; Calculate the stability penalty term : Where c is the preset normal weight; reward function .

[0014] Preferably, the crossover operation of the genetic algorithm is a critical subgraph exchange, which is achieved by identifying and exchanging the critical path subgraphs in the task-dependent directed acyclic graphs of the two parent scheduling schemes. This includes: representing the two parent scheduling schemes as task-dependent directed acyclic graphs; using the critical path method, identifying all task nodes in each task-dependent directed acyclic graph whose earliest start time equals its latest start time, and using the subgraph formed by these task nodes and their inter-node dependencies as the critical path subgraph; randomly selecting a path segment D1 from the critical path subgraph of the first parent scheme, consisting of several consecutive task nodes and their dependencies; randomly selecting a path segment D2 from the critical path subgraph of the second parent scheme, which has the same number of tasks as D1; ​​exchanging D1 and D2, and reconnecting the broken dependency arcs in the two subgraphs after the exchange through a repair operation to generate two new child scheduling schemes.

[0015] Preferably, the mutation operation of the genetic algorithm is as follows: selecting a task corresponding to the delivery uncertainty embedding vector with a high norm, matching alternative resource paths for the task, wherein the cost increment of the matched alternative resource paths is constrained by an upper limit, and the upper limit decreases as the membership value of the task's planned completion time in the original scheduling scheme increases, including: calculating the L2 norm of the delivery uncertainty embedding vector corresponding to each task in the scheduling scheme, and classifying tasks with norm values ​​in the preset highest percentile interval as high-risk tasks; matching a new resource path for each high-risk task from the alternative resource pool, wherein the planned transportation time of the new resource path shall not be higher than that of the original resource path; and calculating the cost increment of the new resource path. And obtain the membership value of the task's planned completion time in the original scheduling scheme. The upper limit of cost increment is calculated using the following formula;

[0016] ,in Based on the basic cost constraint value, Here, k represents the upper limit of the cost increment, and k is an adjustment coefficient greater than or equal to 1. The membership value of the task's planned completion time in the original scheduling scheme; when At that time, resources will be reallocated.

[0017] Secondly, the AI-based supply chain scheduling system of the present invention includes the following modules:

[0018] The embedding module is used to acquire order, resource, logistics topology data and historical fulfillment data of the supply chain; it uses graph neural networks to process the logistics topology data and historical fulfillment data to generate delivery uncertainty embedding vectors for the destination nodes of each scheduling task.

[0019] The fuzzy time window construction module is used to determine the center time and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function based on the components of each dimension of the delivery uncertainty embedding vector, and to construct a composite fuzzy time window for the scheduling task of each destination node. The membership function of the fuzzy time window is obtained by the weighted sum of the asymmetric trapezoidal function and the Gaussian function.

[0020] The reinforcement learning module is used to obtain the reinforcement learning model. The action space of the reinforcement learning model is a multi-dimensional continuous space, which is used to output the weight combination of the cost, time and robustness terms in the scheduling objective function. The reward function of the reinforcement learning model includes a performance penalty term calculated based on a fuzzy time window and a stability penalty term representing the structural difference between scheduling schemes in the previous and subsequent scheduling cycles.

[0021] The scheduling scheme generation module is used to determine the scheduling objective function based on the weight combination, and to solve the scheduling objective function using a genetic algorithm. The crossover operation of the genetic algorithm is a key subgraph exchange, which is achieved by identifying and exchanging the key path subgraphs in the directed acyclic graph of the task dependencies of the two parent scheduling schemes.

[0022] Preferably, the mutation operation of the genetic algorithm is as follows: select a task corresponding to the high-norm delivery uncertainty embedding vector, match alternative resource paths for the task, the cost increment of the matched alternative resource paths is constrained by an upper limit, and the upper limit decreases as the membership value of the task in the original scheduling scheme increases.

[0023] Preferably, the step of determining the center time and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function based on the components of each dimension of the delivery uncertainty embedding vector, and constructing a composite fuzzy time window for the scheduling task of each destination node, wherein the membership function of the fuzzy time window is obtained by the weighted sum of the asymmetric trapezoidal function and the Gaussian function, includes: the membership function of the fuzzy time window. The calculation formula is: ;in, It is an asymmetric trapezoidal function. Let w be a Gaussian function and w be the weights; embed the vector using delivery uncertainty. The parameters of the membership function are determined by the following transformation of the first five components:

[0024] The central moment of the asymmetric trapezoidal function The left slope of an asymmetric trapezoidal function The right-hand slope of an asymmetric trapezoidal function Mean of Gaussian function Variance of Gaussian function ;in to and to These are the preset transformation coefficients.

[0025] Preferably, the reward function of the reinforcement learning model is calculated as follows: for each task i in the scheduling scheme, the planned completion time of the task is obtained. And calculate the membership value based on the composite morphological fuzzy time window of the destination node of the task. ; Calculate performance penalties : The process involves summing all tasks in the current scheduling scheme; comparing the resources allocated to each task in the current scheduling scheme with those in the previous scheduling cycle, and counting the number of tasks with resource changes. ; Calculate the stability penalty term : Where c is the preset normal weight; reward function .

[0026] The beneficial effects of this invention are as follows: By deeply mining the topology and historical data of the logistics network through graph neural networks, embedding vectors that can represent multi-dimensional delivery uncertainties are generated. Based on these vectors, a fuzzy time window composed of an asymmetric trapezoidal function and a Gaussian function is constructed, which can better reflect the complex fulfillment risks in the actual supply chain. The reinforcement learning model is used to determine the weight combination of cost, time, and robustness in the multi-objective scheduling function, which helps to find a better balance between conflicting objectives. The stability penalty term in the reward function ensures a smooth transition between different periodic scheduling schemes. During the solution process, the crossover operation of key subgraph swapping effectively preserves and reorganizes the excellent structure in the parent scheme, accelerating the convergence of the algorithm; the mutation operation concentrates computational resources on improving the weak links of the scheduling scheme, and specifically reallocates resources for high-risk tasks. Attached Figure Description

[0027] Figure 1This is a flowchart illustrating the AI-based supply chain scheduling method provided in an embodiment of the present invention. Detailed Implementation

[0028] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0029] like Figure 1 As shown, an embodiment of the AI-based supply chain scheduling method provided by the present invention includes the following steps:

[0030] S1 acquires order, resource, logistics topology data and historical fulfillment data of the supply chain; uses graph neural networks to process the logistics topology data and historical fulfillment data to generate delivery uncertainty embedding vectors for the destination nodes of each scheduling task.

[0031] The supply chain includes factories, warehouses, logistics, and customers, which act as nodes in the graph structure of the supply chain. The direction or process of goods flow acts as edges in the graph structure. Each node and edge is assigned initial features, including node features such as capacity and inventory level, and edge features such as historical transportation time, cost, and transportation mode. A graph neural network, specifically a graph attention network, is constructed. The graph structure and features of the supply chain are used as input, and historical order fulfillment data, such as the deviation between the actual and planned delivery times for each order, is used as training labels. The graph attention network is trained end-to-end by minimizing the mean squared error between the predicted delivery time deviation and the actual deviation. After training, for any destination node, its relevant data is input, and the output of the graph attention network is the delivery uncertainty embedding vector for that node.

[0032] Specifically, a logistics topology graph G=(V,E) is constructed for the supply chain, where the node set V represents the various entities in the supply chain, such as factories, warehouses, ports, distribution centers, and final destination nodes; and the edge set E represents the actual logistics paths between entities. A feature vector is initialized for each node in the logistics topology graph. The initialized feature vector includes the node's static attributes and dynamic attributes obtained from historical fulfillment data statistics. Static attributes include warehouse capacity, geographical location, and processing capacity, while dynamic attributes include the node's average delay time, delay variance, and historical throughput. A graph attention network is used to perform multi-level iterative processing on the logistics topology graph data. In each layer, each node aggregates the feature information of its neighboring nodes and combines it with its own feature information to update its representation vector through a nonlinear transformation function. After L layers of iteration, the representation vector of each node incorporates the topological structure and neighbor attribute information within its L-hop neighbor range. The representation vectors of each scheduled destination node obtained after multi-level processing are output as the delivery uncertainty embedding vector.

[0033] S2, based on the components of each dimension of the delivery uncertainty embedding vector, determine the center time and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function, and construct a composite fuzzy time window for the scheduling task of each destination node. The membership function of the fuzzy time window is obtained by the weighted sum of the asymmetric trapezoidal function and the Gaussian function.

[0034] Specifically, the membership function of the fuzzy time window The calculation formula is:

[0035] ;in, It is an asymmetric trapezoidal function. Let w be a Gaussian function and w be the weights; embed the vector using delivery uncertainty. The parameters of the membership function are determined by the following transformation of the first five components:

[0036] The central moment of the asymmetric trapezoidal function The left slope of an asymmetric trapezoidal function The right-hand slope of an asymmetric trapezoidal function Mean of Gaussian function Variance of Gaussian function ;in to and to These are the preset transformation coefficients.

[0037] For example, suppose a task is delivered to warehouse A, and the first five components of its delivery uncertainty embedding vector are 0.8, -0.5, -0.2, 0.9, and -1.0. Preset transformation coefficients a1 to a5 are 10, 0.5, 0.4, 10, 0.6, and b1 to b5 are 1440, 0, 0, 1440, 1. Calculations show that the asymmetric trapezoidal function has a center time of 1448 minutes, a left slope of 0.78, a right slope of -0.92, and a Gaussian function with a mean of 1449 minutes and a variance of 1.49. The trapezoidal part represents that the customer has a core expected delivery time of 1448 minutes and a slightly higher tolerance for early delivery than late delivery. The Gaussian part forms a secondary satisfaction peak around 1449 minutes, representing that the customer has an alternative delivery time. The membership function is obtained by weighted summation of these two parts. Compared to a single trapezoidal or rectangular time window, the membership function of this invention can better represent customers' complex preferences for delivery time.

[0038] S3, Obtain the reinforcement learning model. The action space of the reinforcement learning model is a multi-dimensional continuous space, used to output the weight combination of cost, time and robustness terms in the scheduling objective function. The reward function of the reinforcement learning model includes a performance penalty term calculated based on a fuzzy time window and a stability penalty term representing the structural difference between scheduling schemes in previous and subsequent scheduling cycles.

[0039] Specifically, the reinforcement learning model employs an actor-critic architecture, comprising two deep neural networks: an actor network and a critic network. The actor network is responsible for the policy function, taking the current supply chain state vector as input (e.g., containing information such as the number of pending orders, resource load, and current uncertainty level), and outputting a specific action in a multi-dimensional continuous space—a set of weights. The critic network is responsible for the value function, taking the state vector and the action vector output by the actor network as input, and outputting an evaluation value for that state-action pair to assess the merits of the action taken. Through training, the critic network can evaluate the merits of actions, and the gradient information of its evaluation values ​​is used by the actor network to update its parameters.

[0040] At each decision moment, the output action is a vector consisting of multiple real-valued components, each of which can take values ​​within a preset continuous interval. Each dimension of the action vector corresponds to a weight coefficient of an optimization objective in the scheduling objective function, where the optimization objective consists of cost, time, and robustness terms. For example, outputting a three-dimensional action vector... , , ],in , , Both can take continuous values ​​in the interval [0, 1], and , and The sum of the three equals 1. , , These represent the weights of the cost, time, and robustness items, respectively.

[0041] Specifically, the reward function of a reinforcement learning model is calculated in the following way:

[0042] For each task i in the scheduling scheme, obtain the planned completion time of the task. And calculate the membership value based on the composite morphological fuzzy time window of the destination node of the task. ; Calculate performance penalties : The process involves summing all tasks in the current scheduling scheme; comparing the resources allocated to each task in the current scheduling scheme with those in the previous scheduling cycle, and counting the number of tasks with resource changes. ; Calculate the stability penalty term : Where c is the preset normal weight; reward function .

[0043] For example, suppose a scheduling scheme includes three tasks: Task A, Task B, and Task C. Based on the planned delivery times, the calculated membership values ​​are 0.95, 0.80, and 0.60, respectively. Therefore, the performance penalty is 0.65, with a lower value indicating higher customer satisfaction. Simultaneously, comparing the resource allocation between this scheduling and the previous scheduling, assuming the vehicles for Task A and Task C remain unchanged, but the vehicle for Task B changes from vehicle number 2 to vehicle number 5, then the number of tasks with resource changes is 1. If the preset positive constant weight c is 0.5, then the stability penalty is 0.5, and the total reward of the scheduling scheme is -1.15. The goal of the reinforcement learning model is to find a scheduling scheme that maximizes the total reward value through continuous optimization.

[0044] S4. Determine the scheduling objective function based on the weight combination, and use a genetic algorithm to solve the scheduling objective function. The crossover operation of the genetic algorithm is a key subgraph exchange, which is achieved by identifying and exchanging the key path subgraphs in the directed acyclic graph of the task dependencies of the two parent scheduling schemes.

[0045] Specifically, a complete scheduling scheme is encoded as a chromosome, and the fitness function of the genetic algorithm is the scheduling objective function to be optimized. ,

[0046] ;

[0047] in, As the weight of cost items, As the weight of the time item, For the robustness term weights, Total transportation and production costs, Substitute the planned completion time of all tasks into the sum of the penalty values ​​calculated using their respective fuzzy time window functions. This is the sum of the norms of the delivery uncertainty embedding vectors corresponding to all tasks in the scheme. In the crossover operation, the two parent scheduling schemes are transformed into task-dependent directed acyclic graphs, and the critical path algorithm is used to find the longest-running task sequences, i.e., the critical path subgraphs. The two critical path subgraphs are swapped, and the start times of the affected non-critical path tasks are adjusted to generate two new child schemes. In the mutation operation, the L2 norm of the delivery uncertainty embedding vector corresponding to each task is calculated; tasks with larger L2 norms are more likely to be selected. For selected high-risk tasks, new resource paths with lower costs or shorter transit times are searched.

[0048] In a preferred embodiment, the crossover operation of the genetic algorithm is a critical subgraph exchange, which is achieved by identifying and exchanging the critical path subgraphs in the task-dependent directed acyclic graphs of the two parent scheduling schemes. This includes: representing the two parent scheduling schemes as task-dependent directed acyclic graphs; using the critical path method, identifying all task nodes in each task-dependent directed acyclic graph whose earliest start time equals its latest start time, and using the subgraph formed by these task nodes and their inter-node dependencies as the critical path subgraph; randomly selecting a path segment D1 from the critical path subgraph of the first parent scheme, consisting of several consecutive task nodes and their dependencies; randomly selecting a path segment D2 from the critical path subgraph of the second parent scheme, which has the same number of tasks as D1; ​​exchanging D1 and D2, and reconnecting the broken dependency arcs in the two subgraphs after the exchange through a repair operation to generate two new child scheduling schemes.

[0049] For example, suppose the critical path of parent scheme 1 is from task A to task C and then to task E, and this path determines the shortest completion time of the entire scheme. The critical path of parent scheme 2 is from task F to task G and then to task H. The task scheduling on the critical path is very compact, with no redundant time. A path segment D1, such as task C, is randomly selected from the critical path of scheme 1. Correspondingly, a segment D2 of equal length, such as task G, is selected from the critical path of scheme 2. D1 and D2 are swapped. After the swap, the execution sequence of scheme 1 becomes task A to task G and then to task E, while the sequence of scheme 2 becomes task F to task C and then to task H. Not only the tasks themselves are swapped, but also the resource allocation, scheduling, and other information associated with the tasks are also swapped. After the swap, a repair operation is performed, such as recalculating the earliest start time of tasks E and H, checking for resource conflicts, and ensuring that the two newly generated child schemes are logically and resource-wise feasible. By swapping the critical task segments that determine the overall project duration, the cross-operation can obtain a high-quality solution space that may shorten the total completion time.

[0050] In a preferred embodiment, the mutation operation of the genetic algorithm is as follows: select a task corresponding to the high-norm delivery uncertainty embedding vector, match alternative resource paths for the task, the cost increment of the matched alternative resource paths is constrained by an upper limit, and the upper limit decreases as the membership value of the task in the original scheduling scheme increases.

[0051] Specifically, the L2 norm of the delivery uncertainty embedding vector corresponding to each task in the scheduling scheme is calculated, and tasks with norm values ​​in the preset highest percentile range are identified as high-risk tasks. A new resource path is matched for each high-risk task from the alternative resource pool, and the planned transportation time of the new resource path must not exceed that of the original resource path. The cost increment of the new resource path is calculated. And obtain the membership value of the task's planned completion time in the original scheduling scheme. The upper limit of cost increment is calculated using the following formula;

[0052] ,in Based on the basic cost constraint value, Here, k represents the upper limit of the cost increment, and k is an adjustment coefficient greater than or equal to 1. The membership value of the task's planned completion time in the original scheduling scheme; when At that time, resources will be reallocated.

[0053] For example, in a scheduling scheme containing 200 tasks, the L2 norm of the delivery uncertainty embedding vector for each task is calculated. Assume that the top 5% of tasks by norm value are defined as high-risk tasks, and task X is selected because of its high norm value. Currently, task X is assigned a standard truck, and its planned completion time has a membership value of 0.4, indicating low customer satisfaction. A new resource path is found among the alternative resources, using a smaller truck to perform this task. The transportation time remains the same, but the cost increases by 50 yuan, hence the cost increment is 50. Assume the basic cost constraint value... Let the cost be 100 yuan, and the adjustment coefficient k be set to 1.0. Then, what is the upper limit of the cost increment? The cost increment is 60 yuan. Since the actual cost increase of 50 yuan is less than the calculated upper limit of 60 yuan, this mutation operation is accepted, and the resources for task X are updated to a small truck. This allows for targeted investment of additional resources to improve high-risk tasks with low customer satisfaction, thereby enhancing overall service quality, while keeping costs under control.

[0054] The implementation principle of the AI-based supply chain scheduling method in this invention is as follows: By deeply mining the logistics network topology and historical data through graph neural networks, embedding vectors that characterize multi-dimensional delivery uncertainties are generated. This allows the construction of a fuzzy time window composed of an asymmetric trapezoidal function and a Gaussian function, which better reflects the complex fulfillment risks in the actual supply chain. Furthermore, using a reinforcement learning model to determine the weight combination of cost, time, and robustness in the multi-objective scheduling function helps find a better balance between conflicting objectives. The stability penalty term in the reward function ensures a smooth transition between different periodic scheduling schemes. Moreover, during the solution process, the crossover operation of key subgraph swapping effectively preserves and reorganizes the excellent structure in the parent scheme, accelerating the algorithm's convergence; the mutation operation concentrates computational resources on improving the weak links of the scheduling scheme, specifically reallocating resources for high-risk tasks.

[0055] The AI-based supply chain scheduling system provided by this invention includes the following modules:

[0056] The embedding module is used to acquire order, resource, logistics topology data and historical fulfillment data of the supply chain; it uses graph neural networks to process the logistics topology data and historical fulfillment data to generate delivery uncertainty embedding vectors for the destination nodes of each scheduling task.

[0057] The fuzzy time window construction module is used to determine the center time and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function based on the components of each dimension of the delivery uncertainty embedding vector, and to construct a composite fuzzy time window for the scheduling task of each destination node. The membership function of the fuzzy time window is obtained by the weighted sum of the asymmetric trapezoidal function and the Gaussian function.

[0058] The reinforcement learning module is used to obtain the reinforcement learning model. The action space of the reinforcement learning model is a multi-dimensional continuous space, which is used to output the weight combination of the cost, time and robustness terms in the scheduling objective function. The reward function of the reinforcement learning model includes a performance penalty term calculated based on a fuzzy time window and a stability penalty term representing the structural difference between scheduling schemes in previous and subsequent scheduling cycles.

[0059] The scheduling scheme generation module is used to determine the scheduling objective function based on the weight combination, and to solve the scheduling objective function using a genetic algorithm. The crossover operation of the genetic algorithm is a key subgraph exchange, which is achieved by identifying and exchanging the key path subgraphs in the directed acyclic graph of the task dependencies of the two parent scheduling schemes.

[0060] In a preferred embodiment, the mutation operation of the genetic algorithm is as follows: select a task corresponding to the high-norm delivery uncertainty embedding vector, match alternative resource paths for the task, the cost increment of the matched alternative resource paths is constrained by an upper limit, and the upper limit decreases as the membership value of the task in the original scheduling scheme increases.

[0061] In a preferred embodiment, the step of determining the center time and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function based on the components of the delivery uncertainty embedding vector, and constructing a composite fuzzy time window for the scheduling task of each destination node, wherein the membership function of the fuzzy time window is obtained by a weighted sum of the asymmetric trapezoidal function and the Gaussian function, includes:

[0062] Membership function of fuzzy time window The calculation formula is: ;in, It is an asymmetric trapezoidal function. Let w be a Gaussian function and w be the weights; embed the vector using delivery uncertainty. The parameters of the membership function are determined by the following transformation of the first five components:

[0063] The central moment of the asymmetric trapezoidal function The left slope of an asymmetric trapezoidal function The right-hand slope of an asymmetric trapezoidal function Mean of Gaussian function Variance of Gaussian function ;in to and to These are the preset transformation coefficients.

[0064] In a preferred embodiment, the reward function of the reinforcement learning model is calculated as follows:

[0065] For each task i in the scheduling scheme, obtain the planned completion time of the task. And calculate the membership value based on the composite morphological fuzzy time window of the destination node of the task. ; Calculate performance penalties : The process involves summing all tasks in the current scheduling scheme; comparing the resources allocated to each task in the current scheduling scheme with those in the previous scheduling cycle, and counting the number of tasks with resource changes. ; Calculate the stability penalty term : Where c is the preset normal weight; reward function .

[0066] In a preferred embodiment, the crossover operation of the genetic algorithm is a critical subgraph exchange, which is achieved by identifying and exchanging the critical path subgraphs in the task-dependent directed acyclic graphs of the two parent scheduling schemes. This includes: representing the two parent scheduling schemes as task-dependent directed acyclic graphs; using the critical path method, identifying all task nodes in each task-dependent directed acyclic graph whose earliest start time equals its latest start time, and using the subgraph formed by these task nodes and their inter-node dependencies as the critical path subgraph; randomly selecting a path segment D1 from the critical path subgraph of the first parent scheme, consisting of several consecutive task nodes and their dependencies; randomly selecting a path segment D2 from the critical path subgraph of the second parent scheme, which has the same number of tasks as D1; ​​exchanging D1 and D2, and reconnecting the broken dependency arcs in the two subgraphs after the exchange through a repair operation to generate two new child scheduling schemes.

[0067] In a preferred embodiment, the mutation operation of the genetic algorithm includes: selecting a task corresponding to the delivery uncertainty embedding vector with a high norm; matching alternative resource paths for the task; the cost increment of the matched alternative resource paths is constrained by an upper limit, and the upper limit decreases as the membership value of the task's planned completion time in the original scheduling scheme increases; calculating the L2 norm of the delivery uncertainty embedding vector corresponding to each task in the scheduling scheme; classifying tasks with norm values ​​in the preset highest percentile interval as high-risk tasks; matching a new resource path for each high-risk task from the alternative resource pool; the planned transportation time of the new resource path must not exceed that of the original resource path; and calculating the cost increment of the new resource path. And obtain the membership value of the task's planned completion time in the original scheduling scheme. The upper limit of cost increment is calculated using the following formula;

[0068] ,in Based on the basic cost constraint value, Here, k represents the upper limit of the cost increment, and k is an adjustment coefficient greater than or equal to 1. The membership value of the task's planned completion time in the original scheduling scheme; when At that time, resources will be reallocated.

[0069] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An AI-based supply chain scheduling method, characterized in that, Includes the following steps: Acquire order, resource, logistics topology data, and historical fulfillment data from the supply chain; utilize graph neural networks to process the logistics topology data and historical fulfillment data, generating delivery uncertainty embedding vectors for the destination nodes of each scheduling task; the input of the graph neural network is the graph structure and features of the supply chain, the edges of the graph structure represent the direction or process of commodity flow, the node features of the graph structure include capacity and inventory level, and the edge features of the graph structure include historical transportation time, cost, and transportation mode; use historical fulfillment data of historical orders as training labels for the graph neural network, and for any destination node, the output of the graph neural network is the delivery uncertainty embedding vector for that node; Based on the components of the delivery uncertainty embedding vector, the center time and slope of the asymmetric trapezoidal function, as well as the mean and variance of the Gaussian function, are determined. A composite fuzzy time window is then constructed for the scheduling tasks of each destination node. The membership function of the fuzzy time window is obtained by a weighted sum of the asymmetric trapezoidal function and the Gaussian function, specifically including: Membership function of fuzzy time window The calculation formula is: ; It is an asymmetric trapezoidal function. Let w be a Gaussian function and w be the weights; embed the vector using delivery uncertainty. The parameters of the membership function are determined by the following transformation of the first five components: The central moment of the asymmetric trapezoidal function ; The left slope of an asymmetric trapezoidal function ; The right slope of an asymmetric trapezoidal function ; Mean of Gaussian function ; variance of Gaussian function ; in to and to The transformation coefficients are preset; where the center time of the asymmetric trapezoidal function represents the customer's expected delivery time, the right slope represents the tolerance for early delivery, and the left slope represents the tolerance for delayed delivery; the Gaussian part forms a minor satisfaction peak near the Gaussian mean, indicating that the customer has an alternative delivery time. The reinforcement learning model is obtained. The input of the reinforcement learning model is the state vector of the current supply chain state, which includes the number of orders to be processed, resource load, and current uncertainty level. The action space of the reinforcement learning model is a multi-dimensional continuous space, which is used to output the weight combination of cost, time, and robustness terms in the scheduling objective function. The reward function of the reinforcement learning model includes a performance penalty term calculated based on fuzzy time windows and a stability penalty term representing the structural difference between scheduling schemes in the previous and subsequent scheduling cycles. Among them, the cost term is the total transportation and production cost, the time term is the sum of the penalty values ​​calculated by substituting the planned completion time of all tasks into their respective fuzzy time window functions, and the robustness term is the sum of the norms of the delivery uncertainty embedding vectors of all tasks. The reward function of a reinforcement learning model is calculated as follows: For each task i in the scheduling scheme, obtain the planned completion time of the task. And calculate the membership value based on the composite morphological fuzzy time window of the destination node of the task. ; Calculate performance penalties : The summation iterates through all tasks in the solution; Compare the resources allocated to each task in the current scheduling plan with those in the previous scheduling cycle, and count the number of tasks that have undergone resource changes. ; Calculate the stability penalty term : , where c is the preset weight of positive constants; reward function ; The scheduling objective function is determined based on the weight combination, and the scheduling objective function is solved using a genetic algorithm. The crossover operation of the genetic algorithm is a critical subgraph exchange, which is achieved by identifying and exchanging the critical path subgraphs in the directed acyclic graph of the task dependencies of the two parent scheduling schemes.

2. The AI-based supply chain scheduling method according to claim 1, characterized in that, The mutation operation of the genetic algorithm is as follows: select the task corresponding to the delivery uncertainty embedding vector of the high norm, match the alternative resource path for the task, the cost increment of the matched alternative resource path is constrained by an upper limit, and the upper limit decreases as the membership value of the task in the original scheduling scheme increases.

3. The AI-based supply chain scheduling method according to claim 1, characterized in that, The crossover operation of the genetic algorithm is a critical subgraph exchange, which is achieved by identifying and exchanging the critical path subgraphs in the directed acyclic graph of the task dependencies of the two parent scheduling schemes, including: The two parent scheduling schemes are represented as a task-dependent directed acyclic graph; The critical path method is used to identify all task nodes whose earliest start time is equal to their latest start time in each task-dependent directed acyclic graph. The subgraph formed by the task nodes and the dependency edges between them is used as the critical path subgraph. From the critical path subgraph of the first parent solution, randomly select a path segment D1 consisting of a series of consecutive task nodes and dependent edges; From the critical path subgraph of the second parent solution, randomly select a path segment D2 that has the same number of tasks as D1; Swap D1 and D2, and reconnect the broken dependency arcs in the two subgraphs after the swap through a repair operation to generate two new child scheduling schemes.

4. The AI-based supply chain scheduling method according to claim 2, characterized in that, The mutation operation of the genetic algorithm is as follows: Selecting a task corresponding to the high-norm delivery uncertainty embedding vector, matching alternative resource paths for the task, wherein the cost increment of the matched alternative resource paths is constrained by an upper limit, and the upper limit decreases as the membership value of the task's planned completion time in the original scheduling scheme increases, including: Calculate the L2 norm of the delivery uncertainty embedding vector corresponding to each task in the scheduling scheme, and classify the tasks whose norm value is in the preset highest percentile interval as high-risk tasks. For each high-risk task, a new resource route is matched from the alternative resource pool. The planned transportation time of the new resource route must not be longer than that of the original resource route. Calculate the cost increment of the new resource path And obtain the membership value of the task's planned completion time in the original scheduling scheme. ; The upper limit of cost increment is calculated using the following formula; ,in Based on the basic cost constraint value, Here, k represents the upper limit of the cost increment, and k is an adjustment coefficient greater than or equal to 1. This represents the membership value of the task in the original scheduling scheme, indicating its planned completion time. when At that time, resources will be reallocated.

5. An AI-based supply chain scheduling system for implementing the method as described in claim 1, characterized in that, Includes the following modules: The embedding module is used to acquire order, resource, logistics topology data and historical fulfillment data of the supply chain; it uses graph neural networks to process the logistics topology data and historical fulfillment data to generate delivery uncertainty embedding vectors for the destination nodes of each scheduling task. The fuzzy time window construction module is used to determine the center time and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function based on the components of each dimension of the delivery uncertainty embedding vector, and to construct a composite fuzzy time window for the scheduling task of each destination node. The membership function of the fuzzy time window is obtained by the weighted sum of the asymmetric trapezoidal function and the Gaussian function. The reinforcement learning module is used to obtain the reinforcement learning model. The action space of the reinforcement learning model is a multi-dimensional continuous space, which is used to output the weight combination of the cost, time and robustness terms in the scheduling objective function. The reward function of the reinforcement learning model includes a performance penalty term calculated based on a fuzzy time window and a stability penalty term representing the structural difference between scheduling schemes in the previous and subsequent scheduling cycles. The scheduling scheme generation module is used to determine the scheduling objective function based on the weight combination, and to solve the scheduling objective function using a genetic algorithm. The crossover operation of the genetic algorithm is a key subgraph exchange, which is achieved by identifying and exchanging the key path subgraphs in the directed acyclic graph of the task dependencies of the two parent scheduling schemes.

6. The system of the AI-based supply chain scheduling method according to claim 5, characterized in that, The mutation operation of the genetic algorithm is as follows: select the task corresponding to the delivery uncertainty embedding vector of the high norm, match the alternative resource path for the task, the cost increment of the matched alternative resource path is constrained by an upper limit, and the upper limit decreases as the membership value of the task in the original scheduling scheme increases.

Images