Supply chain scheduling method and system based on AI

Abstract

The invention relates to the technical field of supply chain scheduling, and discloses an AI-based supply chain scheduling method and system, and the method comprises the steps: obtaining the order, resource, logistics topological data and historical performance data of a supply chain; generating a delivery uncertainty embedding vector by using the graph neural network; determining a central moment and a slope of an asymmetric trapezoidal function and a mean value and a variance of a Gaussian function according to the delivery uncertainty embedded vector, and constructing a fuzzy time window for a scheduling task of each destination node; obtaining a reinforcement learning model, and outputting a weight combination of a cost item, a time item and a robustness item in the scheduling objective function; determining a scheduling objective function according to the weight combination, and solving the scheduling objective function; the crossover operation of the genetic algorithm is key sub-graph exchange, and is realized by identifying and exchanging key path sub-graphs of two parent scheduling schemes. According to the method, complex performance risk distribution can be accurately described, real-time changes can be adapted, and meanwhile the calculation amount is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of supply chain scheduling, and particularly relates to an AI-based supply chain scheduling method and system. BACKGROUND

[0002] Supply chain scheduling aims to reduce operating costs and improve order delivery efficiency by optimizing resource allocation in multiple links such as orders, production, warehousing, and logistics. It is a typical combinatorial optimization problem, and its inherent NP-hard nature means that the time required to find the optimal solution increases exponentially as the problem size increases. Therefore, although traditional mathematical programming scheduling methods can theoretically find the optimal solution, they are often difficult to solve within a reasonable time due to high computational complexity.

[0003] In the face of this challenge, researchers have turned to supply chain scheduling methods that apply meta-heuristic algorithms, such as genetic algorithms and simulated annealing. These meta-heuristic algorithms can quickly find scheduling solutions close to the optimal solution, but the quality of the scheduling solution found is highly dependent on the coding method for the problem itself. However, the modern supply chain environment is full of uncertainties, such as random events such as transportation delays and equipment failures, which can seriously affect the stability and reliability of the scheduling solution found. To deal with these temporal uncertainties, fuzzy time window methods are introduced to represent the satisfaction of task completion time through membership functions. However, the fuzzy time window still has limitations, as it often uses symmetric membership functions, which makes it difficult to accurately depict complex fulfillment risk distribution.

[0004] In addition, existing supply chain scheduling methods generally lack the ability to dynamically adjust according to real-time conditions, making it difficult for existing supply chain scheduling methods to find globally optimal scheduling solutions in a changing operating environment. SUMMARY

[0005] The present application provides an AI-based supply chain scheduling method and system to solve the problems of inaccurate depiction of complex fulfillment risk distribution, difficulty in adapting to real-time changes, and large amount of calculation in the prior art.

[0006] In a first aspect, the AI-based supply chain scheduling method of the present application comprises the following steps: Obtain order, resource, logistics topology data and historical fulfillment data of the supply chain; use a graph neural network to process the logistics topology data and historical fulfillment data to generate a delivery uncertainty embedding vector for each destination node of the scheduling task; The center time and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function are determined according to the dimensional components of the delivery uncertainty embedding vector, and a fuzzy time window in a composite form is constructed for the scheduling task of each destination node, and the membership function of the fuzzy time window is obtained by a weighted sum of the asymmetric trapezoidal function and the Gaussian function. An reinforcement learning model is obtained, the action space of the reinforcement learning model is a multi-dimensional continuous space, and the reinforcement learning model is used to output a weight combination of a cost term, a time term and a robustness term in a scheduling objective function; a reward function of the reinforcement learning model includes a fulfillment penalty term calculated based on a fuzzy time window and a stability penalty term representing a structural difference between scheduling schemes of adjacent scheduling periods; A scheduling objective function is determined according to the weight combination, and a genetic algorithm is used to solve the scheduling objective function; the crossover operation of the genetic algorithm is a key subgraph exchange, and the key path subgraph in the task dependency directed acyclic graph of two parent scheduling schemes is identified and exchanged to realize the key subgraph exchange.

[0007] Preferably, the mutation operation of the genetic algorithm is: selecting a task corresponding to a delivery uncertainty embedding vector with a high norm, matching an alternative resource path for the task, the cost increment of the matched alternative resource path being subject to an upper limit, and the upper limit decreasing as the membership value of the planned completion time of the task in the original scheduling scheme increases.

[0008] Preferably, the center time and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function are determined according to the dimensional components of the delivery uncertainty embedding vector, and a fuzzy time window in a composite form is constructed for the scheduling task of each destination node, and the membership function of the fuzzy time window is obtained by a weighted sum of the asymmetric trapezoidal function and the Gaussian function, comprising: a calculation formula of the membership function of the fuzzy time window . ; wherein is the asymmetric trapezoidal function, is the Gaussian function, and w is the weight; the first five components of the delivery uncertainty embedding vector are used to determine the parameters of the membership function through the following transformations: the center time of the asymmetric trapezoidal function; the left slope of the asymmetric trapezoidal function; the right slope of the asymmetric trapezoidal function; the mean of the Gaussian function; and the variance of the Gaussian function; wherein to and to are preset transformation coefficients.

[0009] Preferably, the reward function of the reinforcement learning model is calculated in the following manner: for each task i in the scheduling scheme, the task plan completion time is obtained, and the membership value is calculated according to the composite form fuzzy time window of the destination node of the task ; a compliance penalty term is calculated : , where the summation is performed on all tasks in the scheme; the resources allocated to each task in the current scheduling scheme are compared with those in the scheduling scheme of the last scheduling period one by one, and the number of tasks with resource changes is counted ; a stability penalty term is calculated : , where c is a preset normal number weight; the reward function .

[0010] Preferably, the crossover operation of the genetic algorithm is a key subgraph exchange, which is implemented by identifying and exchanging the key path subgraphs in the task dependency directed acyclic graphs of two parent scheduling schemes, including: representing the two parent scheduling schemes as task dependency directed acyclic graphs; using the critical path method to identify all task nodes with the earliest start time equal to the latest start time in each task dependency directed acyclic graph, and regarding the subgraph composed of the task nodes and the dependency edges between the task nodes as a key path subgraph; randomly selecting a path segment D1 composed of a plurality of continuous task nodes and dependency edges from the key path subgraph of the first parent scheme; randomly selecting a path segment D2 from the key path subgraph of the second parent scheme, which has the same number of task nodes 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.

[0011] Preferably, the mutation operation of the genetic algorithm is: selecting a task corresponding to a high-norm delivery uncertainty embedding vector, matching a candidate resource path for the task, the cost increment of the matched candidate resource path is subject to an upper limit constraint, and the upper limit decreases as the membership value of the planned completion time of the task 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 regarding the tasks with norm values in a preset highest percentile interval as high-risk tasks; matching a new resource path for each high-risk task from the candidate resource library, and the planned transportation time of the new resource path cannot be higher than that of the original resource path; calculating the cost increment of the new resource path, and obtaining the membership value of the planned completion time of the task in the original scheduling scheme ; the upper limit of the cost increment is calculated by the following formula: , where is a basic cost constraint value, k is an adjustment coefficient greater than or equal to 1, and is an upper limit of the cost increment; is a membership value of the task in the original scheduling scheme; when resource reallocation is performed.

[0012] In a second aspect, the AI-based supply chain scheduling system comprises the following modules: An embedding module is configured to obtain order, resource, logistics topology data and historical performance data of the supply chain, and to generate a delivery uncertainty embedding vector for a destination node of each scheduling task by processing the logistics topology data and the historical performance data using a graph neural network. A fuzzy time window construction module is configured to determine a center time and a slope of an asymmetric trapezoidal function and a mean and a variance of a Gaussian function according to each dimension component of the delivery uncertainty embedding vector, and to construct a composite fuzzy time window for each scheduling task of the destination node, wherein a membership function of the fuzzy time window is obtained by a weighted sum of the asymmetric trapezoidal function and the Gaussian function. A reinforcement learning module is configured to obtain a reinforcement learning model, wherein an action space of the reinforcement learning model is a multi-dimensional continuous space, and the reinforcement learning model is configured to output a weight combination of a cost term, a time term and a robustness term in a scheduling objective function; and a reward function of the reinforcement learning model comprises a performance penalty term calculated based on the fuzzy time window and a stability penalty term representing a structural difference between scheduling schemes of adjacent scheduling periods. A scheduling scheme generation module is configured to determine a scheduling objective function according to the weight combination, and to solve the scheduling objective function by using a genetic algorithm; and a crossover operation of the genetic algorithm is a key subgraph exchange, which is implemented by identifying and exchanging a key path subgraph in a task dependency directed acyclic graph of two parent scheduling schemes.

[0013] Preferably, a mutation operation of the genetic algorithm is as follows: a task corresponding to a delivery uncertainty embedding vector with a high norm is selected, and an alternative resource path is matched for the task, wherein a cost increment of the matched alternative resource path is subject to an upper limit, and the upper limit decreases as a membership value of the task in the original scheduling scheme increases.

[0014] Preferably, the determination of the center time and the slope of the asymmetric trapezoidal function and the mean and the variance of the Gaussian function according to each dimension component of the delivery uncertainty embedding vector, and the construction of the composite fuzzy time window for each scheduling task of the destination node, and the obtaining of the membership function of the fuzzy time window by the weighted sum of the asymmetric trapezoidal function and the Gaussian function, comprise: the calculation formula of the membership function of the fuzzy time window is as follows: ; wherein, the asymmetric trapezoidal function is ​is a Gaussian function, w is a weight; the delivery uncertainty embedding vector is determined by embedding the first five components of the delivery uncertainty vector through the following transformation ; the parameters of the membership function are determined by the following transformation: ; the center time of the asymmetric trapezoidal function ; the left slope of the asymmetric trapezoidal function ; the right slope of the asymmetric trapezoidal function ; the mean of the Gaussian function ; the variance of the Gaussian function ; wherein to and to are preset transformation coefficients.

[0015] Preferably, the reward function of the reinforcement learning model is calculated in the following manner: for each task i in the scheduling scheme, the task plan completion time is obtained, and the membership value is calculated according to the composite form fuzzy time window of the destination node of the task ; the compliance penalty term is calculated: , wherein the summation traverses all tasks in the scheme; the resources allocated to each task in the current scheduling scheme are compared with the scheduling scheme of the last scheduling period one by one, and the number of tasks that have undergone resource changes is counted ; the stability penalty term is calculated: , wherein c is a preset normal number weight; the reward function .

[0016] The beneficial effects of the present application are: through the graph neural network, the logistics network topology and historical data are deeply mined to generate an embedding vector that can represent multi-dimensional delivery uncertainty, and a fuzzy time window composed of an asymmetric trapezoidal function and a Gaussian function is constructed, which can better fit the complex compliance risks in the actual supply chain. Through the reinforcement learning model, the weight combination of cost, time and robustness in the multi-objective scheduling function is determined, which helps to find a better balance point among conflicting objectives, and the stability penalty term in the reward function ensures the smooth transition between scheduling schemes of different periods. In the solving process, the crossover operation of the key subgraph exchange effectively retains and reorganizes the excellent structure in the parent scheme, accelerating the convergence of the algorithm; the mutation operation concentrates the computing resources on improving the weak links of the scheduling scheme, and specifically reallocates resources for high-risk tasks. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The flowchart of the AI-based supply chain scheduling method provided by the embodiments of the present application. DETAILED DESCRIPTION

[0018] Embodiments of the present application are described below in detail, examples of which are shown in the accompanying drawings. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0019] As shown in Figure 1 An embodiment of the AI-based supply chain scheduling method provided by the present application includes the following steps: S1, obtaining order, resource, logistics topology data and historical performance data of the supply chain; processing the logistics topology data and the historical performance data by using a graph neural network to generate a delivery uncertainty embedding vector for each destination node of the scheduling task.

[0020] The supply chain includes factories, warehouses, logistics, customers, etc., which are nodes in the graph structure constructed by the supply chain, and the commodity flow direction or process is the edge of the graph structure. Initial features are assigned to each node and edge, and the node features include production capacity, inventory level, and the edge features include historical transportation time, cost, transportation mode, etc. A graph neural network, specifically a graph attention network, is constructed, which takes the graph structure and features of the supply chain as input, uses the historical performance data of historical orders, such as the deviation of the actual delivery time and the planned delivery time of each order, as training labels, and minimizes the mean square error between the predicted delivery time deviation and the actual deviation of the graph attention network to train the graph attention network end-to-end. After training, for any destination node, input its related data, and the output of the graph attention network is the delivery uncertainty embedding vector of the node.

[0021] Specifically, a logistics topology graph G=(V, E) of the supply chain is constructed, where the node set V represents each entity in the supply chain, such as factories, warehouses, ports, distribution centers, and final destination nodes; and the edge set E represents the actual logistics path between entities. Each node in the logistics topology graph is initialized with a feature vector, which includes static attributes of the node and dynamic attributes obtained based on historical performance data. The static attributes include warehouse capacity, geographic location and processing capacity, and the dynamic attributes include average delay time, delay variance and historical throughput of the node. The graph attention network is used to process the logistics topology graph data in multiple iterations. In each layer, each node aggregates the feature information of its neighbor nodes, combines its own feature information, and updates its own representation vector through a nonlinear transformation function. After L iterations, the representation vector of each node incorporates the topology structure and neighbor attribute information within its L-hop neighbor range. The representation vector of each scheduling destination node obtained after multiple layers of processing is output as a delivery uncertainty embedding vector.

[0022] S2, determine the center time and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function according to the dimensional components of the delivery uncertainty embedding vector, 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.

[0023] Specifically, the membership function of the fuzzy time window The calculation formula is: ; wherein, is the asymmetric trapezoidal function, is the Gaussian function, and w is the weight; the first five components of the delivery uncertainty embedding vector are used to determine the parameters of the membership function through the following transformation: The center time of the asymmetric trapezoidal function The left slope of the asymmetric trapezoidal function The right slope of the asymmetric trapezoidal function The mean of the Gaussian function The variance of the Gaussian function ; wherein to and to are preset transformation coefficients.

[0024] For example, assuming that a task sent to warehouse A has the first five components of the delivery uncertainty embedding vector as 0.8, -0.5, -0.2, 0.9 and -1.0. The 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. Through calculation, the center time of the asymmetric trapezoidal function is 1448 minutes, the left slope is 0.78, the right slope is -0.92, the mean of the Gaussian function is 1449 minutes, and the variance is 1.49. The trapezoidal part represents that the customer has a core expected delivery time of 1448 minutes, and the tolerance for early delivery is slightly higher than that for late delivery. The Gaussian part forms a secondary satisfaction peak around 1449 minutes, representing an alternative customer receiving period. The membership function is obtained by weighted sum of the two parts. Compared with a single trapezoidal or rectangular time window, the membership function of the present application can better represent the customer's complex preferences for delivery time.

[0025] S3, obtain a reinforcement learning model, the action space of the reinforcement learning model is a multi-dimensional continuous space, used to output the weight combination of the cost term, the time term and the robustness term in the scheduling objective function; the reward function of the reinforcement learning model includes a performance penalty term calculated based on the fuzzy time window and a stability penalty term representing the structural difference between the scheduling schemes of the previous and subsequent scheduling periods.

[0026] Specifically, the reinforcement learning model adopts an actor-critic architecture, including two deep neural networks, an actor network and a critic network. The actor network is responsible for the policy function, whose input is the state vector of the current supply chain state, e.g., containing the number of orders to be processed, resource load, current uncertainty level, etc., and the output is a specific action in a multi-dimensional continuous space, i.e., a combination of weights. The critic network is responsible for the value function, whose input is the state vector and the action vector output by the actor network, and the output is the evaluation value of the state-action pair, which is used to evaluate the pros and cons of the action taken. Through training, the critic network can evaluate the pros and cons of the action, and the gradient information of the evaluation value is used to update the parameters of the actor network.

[0027] The action that can be output at each decision-making moment is a vector composed of multiple real components, and each component can take a value within a pre-set continuous interval. Each dimension component of the action vector corresponds to the weight coefficient of an optimization objective in the scheduling objective function, where the optimization objective is the cost term, the time term, and the robustness term. For example, output a three-dimensional action vector , , ], where 、 、 can take continuous values within the interval [0, 1], and 、 and the sum of the three is equal to 1, 、 、 represent the weights of the cost term, the time term, and the robustness term, respectively.

[0028] Specifically, the reward function of the reinforcement learning model is calculated as follows: For each task i in the scheduling scheme, the task plan completion time is obtained, and the membership value is calculated according to the composite form fuzzy time window of the destination node of the task; the compliance penalty term is calculated: , where the summation iterates over all tasks in the scheme; the number of tasks with resource changes is counted by comparing the resources allocated to each task in the current scheduling scheme with the scheduling scheme of the previous scheduling period; the stability penalty term is calculated: , where c is a pre-set normal weight; the reward function .

[0029] Exemplarily, it is assumed that a scheduling scheme contains three tasks, task A, task B and task C. According to the scheduled delivery time, the calculated membership values are 0.95, 0.80 and 0.60 respectively. Then, the fulfillment penalty is 0.65, and the smaller the value is, the higher the customer satisfaction is. At the same time, compared with the resource allocation of the last scheduling, it is assumed that the execution vehicles of task A and task C do not change, but the execution vehicle of task B is changed from vehicle No. 2 to vehicle No. 5, and the number of tasks with resource changes is 1. If the preset normal number weight c is 0.5, the stability penalty term 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 can maximize the total reward value through continuous optimization.

[0030] S4, determine a scheduling target function according to the weight combination, and solve the scheduling target function by using a genetic algorithm; the crossover operation of the genetic algorithm is key subgraph exchange, which is realized by identifying and exchanging the key path subgraphs in the task dependency directed acyclic graph of two parent scheduling schemes.

[0031] Specifically, a complete scheduling scheme is encoded as a chromosome, and the fitness function of the genetic algorithm is the scheduling target function to be optimized , ; Among them, is the cost term weight, is the time term weight, is the robustness term weight, is the total transportation and production cost, is the sum of penalty values calculated by substituting the planned completion time of all tasks into the respective fuzzy time window function, 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 converted into task dependency directed acyclic graphs, and the longest task sequence, i.e., the key path subgraph, of each is found by using the critical path algorithm. The two key path subgraphs are exchanged, and the start times of the non-key path tasks affected 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, and the probability of being selected is higher for the task with a larger L2 norm. For the selected high-risk task, a new resource path with lower cost or shorter transportation time is searched.

[0032] In a preferred embodiment, the crossover operation of the genetic algorithm is a critical subgraph exchange, which is implemented by identifying and exchanging critical path subgraphs in the task dependency directed acyclic graphs of two parent scheduling schemes, comprising: representing the two parent scheduling schemes as task dependency directed acyclic graphs; identifying all task nodes with an earliest start time equal to a latest start time in each task dependency directed acyclic graph using the critical path method, and regarding the subgraph formed by the task nodes and the dependency edges between the task nodes as a critical path subgraph; randomly selecting a path segment D1 consisting of a plurality of continuous task nodes and dependency edges from the critical path subgraph of the first parent scheme; randomly selecting a path segment D2 equal in number of tasks to D1 from the critical path subgraph of the second parent scheme; exchanging D1 and D2, and reconnecting the broken dependency arcs in the two exchanged subgraphs through a repair operation to generate two new child scheduling schemes.

[0033] For example, assume that the critical path of the first parent scheme is task A to task C to task E, and this path determines the shortest completion time of the entire scheme. The critical path of the second parent scheme is task F to task G to task H. The scheduling of the tasks on the critical path is very compact, with no redundant time. A path segment D1 is randomly selected from the critical path of the first scheme, such as task C. Correspondingly, a path segment D2 equal in length to D1 is selected from the critical path of the second scheme, such as task G. D1 and D2 are exchanged, and after the exchange, the execution sequence of the first scheme becomes task A to task G to task E, while the sequence of the second scheme becomes task F to task C to task H. Not only the tasks themselves but also the resource allocation, time arrangement, and other information associated with the tasks are exchanged. After the exchange, a repair operation is performed, for example, the earliest start time of task E and task H is recalculated, resource conflicts are checked, and it is ensured that the two new child schemes generated are feasible in logic and resources. By exchanging the critical task segments that determine the overall duration, the crossover operation can obtain a high-quality solution space that can shorten the total completion time.

[0034] In a preferred embodiment, the mutation operation of the genetic algorithm is: selecting a task corresponding to a high-norm delivery uncertainty embedding vector, matching an alternative resource path for the task, the cost increment of the matched alternative resource path is subject to an upper limit, and the upper limit decreases as the membership value of the planned completion time of the task in the original scheduling scheme increases.

[0035] Specifically, the L2 norm of the delivery uncertainty embedding vector corresponding to each task in the scheduling scheme is calculated, and the tasks with norm values in a preset highest percentile interval are regarded as high-risk tasks; a new resource path is matched for each high-risk task from the alternative resource library, and the planned transportation time of the new resource path cannot be higher than that of the original resource path; the cost increment of the new resource path is calculated and obtain the membership value of the planned completion time of the task in the original scheduling scheme ; the upper limit of the cost increment is calculated by the following formula: , wherein is a basic cost constraint value, is the upper limit of the cost increment, k is an adjustment coefficient greater than or equal to 1; is the membership value of the planned completion time of the task in the original scheduling scheme; when , resource reallocation is performed.

[0036] For example, in a scheduling scheme containing 200 tasks, the L2 norm of the delivery uncertainty embedding vector of each task is calculated. Assuming that the tasks ranked in the top five percent of the norm value are defined as high-risk tasks, task X is selected because its norm value is high. At present, a standard truck is allocated to task X, and the calculated membership value of the planned completion time is 0.4, indicating that the customer satisfaction is low. A new resource path is found in the alternative resources, and a small truck is used to perform this task, and the transportation time does not change, but the cost increases by 50 yuan, so the cost increment is 50. Assuming that the basic cost constraint value is 100 yuan, and the adjustment coefficient k is set to 1.0. Then, the upper limit of the cost increment is 60 yuan. Because the actual cost increment of 50 yuan is less than the calculated upper limit of 60 yuan, the mutation operation is accepted, and the resource of task X is updated to a small truck. The additional resources are invested to improve those high-risk and low-customer-satisfaction tasks, thereby improving the overall service quality under the premise of controllable cost.

[0037] The implementation principle of the AI-based supply chain scheduling method of the embodiment of the application is as follows: the logistics network topology and historical data are deeply mined through a graph neural network to generate an embedding vector capable of representing multi-dimensional delivery uncertainty, so as to construct a fuzzy time window composed of an asymmetric trapezoidal function and a Gaussian function, which can better fit the complex performance risk in the actual supply chain. In addition, the weight combination of cost, time and robustness in the multi-objective scheduling function is determined through a reinforcement learning model, which helps to find a better balance point between conflicting objectives, and the stability penalty term in the reward function ensures the smooth transition between scheduling schemes of different periods. Moreover, in the solving process, the crossover operation of key subgraph exchange effectively retains and reorganizes the excellent structure in the parent scheme, accelerating the convergence of the algorithm; the mutation operation concentrates the computing resources on improving the weak link of the scheduling scheme, and specifically reallocates resources for high-risk tasks.

[0038] The embodiment of the AI-based supply chain scheduling system provided by the application includes the following modules: An embedding module is configured to obtain order, resource, logistics topology data and historical fulfillment data of a supply chain; a graph neural network is used to process the logistics topology data and the historical fulfillment data to generate a delivery uncertainty embedding vector for a destination node of each scheduling task.

[0039] A fuzzy time window construction module is configured to determine a center time and a slope of an asymmetric trapezoidal function and a mean and a variance of a Gaussian function according to each dimension component of the delivery uncertainty embedding vector, and to construct a fuzzy time window in a composite form for a scheduling task of each destination node, wherein a membership function of the fuzzy time window is obtained by a weighted sum of the asymmetric trapezoidal function and the Gaussian function.

[0040] A reinforcement learning module is configured to obtain a reinforcement learning model, wherein an action space of the reinforcement learning model is a multi-dimensional continuous space, and the reinforcement learning model is configured to output a weight combination of a cost term, a time term and a robustness term in a scheduling objective function; a reward function of the reinforcement learning model includes a fulfillment penalty term calculated based on the fuzzy time window and a stability penalty term representing a structural difference between scheduling schemes of adjacent scheduling periods.

[0041] A scheduling scheme generation module is configured to determine a scheduling objective function according to the weight combination, and to perform scheduling solving on the scheduling objective function by using a genetic algorithm; a crossover operation of the genetic algorithm is a key subgraph exchange, which is implemented by identifying and exchanging a key path subgraph in a task dependency directed acyclic graph of two parent scheduling schemes.

[0042] In a preferred embodiment, a mutation operation of the genetic algorithm is as follows: a task corresponding to a delivery uncertainty embedding vector with a high norm is selected, an alternative resource path is matched for the task, a cost increment of the matched alternative resource path is subject to an upper limit constraint, and the upper limit decreases as an increase of a membership value of a planned completion time of the task in an original scheduling scheme.

[0043] In a preferred embodiment, the fuzzy time window construction module is configured to determine a center time and a slope of an asymmetric trapezoidal function and a mean and a variance of a Gaussian function according to each dimension component of the delivery uncertainty embedding vector, and to construct a fuzzy time window in a composite form for a scheduling task of each destination node, wherein a membership function of the fuzzy time window is obtained by a weighted sum of the asymmetric trapezoidal function and the Gaussian function, and the method comprises the following steps: calculating a membership function of a fuzzy time window is as follows: ; wherein, is an asymmetric trapezoidal function, is a Gaussian function, and w is a weight; the first five components of the delivery uncertainty embedding vector are used to determine parameters of the membership function by the following transformation: a center time of the asymmetric trapezoidal function ; a left slope of the asymmetric trapezoidal function ; right slope of asymmetric trapezoidal function ; mean of Gaussian function ; variance of Gaussian function ; where to and to are preset transform coefficients.

[0044] In a preferred embodiment, the reward function of the reinforcement learning model is calculated in the following manner: For each task i in the scheduling scheme, obtain the task plan completion time , and calculate the membership value according to the composite form fuzzy time window of the destination node of the task ; calculate the compliance penalty term : , where the summation traverses all tasks in the scheme; compare the resources allocated to each task in the current scheduling scheme with the scheduling scheme of the last scheduling period one by one, and count the number of tasks that have changed resources ; calculate the stability penalty term : , where c is a preset normal number weight; the reward function .

[0045] In a preferred embodiment, the crossover operation of the genetic algorithm is a key subgraph exchange, which is implemented by identifying and exchanging the critical path subgraphs in the task dependency directed acyclic graphs of two parent scheduling schemes, including: representing the two parent scheduling schemes as task dependency directed acyclic graphs; using the critical path method to identify all task nodes with earliest start time equal to latest start time in each task dependency directed acyclic graph, and regarding the subgraph composed of the task nodes and the dependency edges between the task nodes as a critical path subgraph; randomly selecting a path segment D1 composed of a plurality of continuous task nodes and dependency edges from the critical path subgraph of the first parent scheme; 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 exchanged subgraphs through a repair operation to generate two new child scheduling schemes.

[0046] In a preferred embodiment, the mutation operation of the genetic algorithm is: selecting a task corresponding to a high-norm delivery uncertainty embedding vector, matching an alternative resource path for the task, the cost increment of the matched alternative resource path being subject to an upper limit, and the upper limit decreasing as the membership value of the planned completion time of the task in the original scheduling scheme increases, comprising: calculating the L2 norm of the delivery uncertainty embedding vector corresponding to each task in the scheduling scheme, and taking the task with the norm value in the preset highest percentile interval as a high-risk task; matching a new resource path for each high-risk task from the alternative resource library, the planned transportation time of the new resource path not being higher than the original resource path; calculating the cost increment of the new resource path , and obtaining the membership value of the planned completion time of the task in the original scheduling scheme ; calculating the upper limit of the cost increment by the following formula , wherein is the basic cost constraint value, is the upper limit of the cost increment, k is an adjustment coefficient greater than or equal to 1; is the membership value of the planned completion time of the task in the original scheduling scheme; when , resource reallocation is performed.

[0047] Although the embodiments of the present application have been shown and described above, it should be understood that the above-described embodiments are exemplary and should not be construed as limiting the present application, and those of ordinary skill in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

Claims

1. An AI-based supply chain scheduling method, characterized in that: The steps include: Obtain supply chain orders, resources, logistics topology data, and historical fulfillment data; use graph neural networks to process logistics topology data and historical fulfillment data to generate delivery uncertainty embedding vectors for the destination nodes of each scheduling task; Determine the central moment and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function based on the dimensional components of the delivery uncertainty embedding vector, 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. Obtain a reinforcement learning model. The action space of the reinforcement learning model is a multidimensional continuous space, which is used to output the weighted combination of the cost term, time term, and robustness term 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 that represents the structural difference between the scheduling plans of the previous and next scheduling cycles. The scheduling objective function is determined according to the weight combination, and the scheduling objective function is solved by 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 task dependency directed acyclic graph 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 to select tasks corresponding to high-norm delivery uncertainty embedding vectors, match alternative resource paths for the tasks, and subject the cost increment of the matched alternative resource paths to an upper limit constraint, wherein the upper limit decreases as the membership value of the task's planned completion time in the original scheduling scheme increases.

3. The AI-based supply chain scheduling method according to claim 1, characterized in that: The central moment and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function are determined according to the dimensional components of the delivery uncertainty embedding vector, and 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, including: Membership function of fuzzy time window The calculation formula is: ;in, is an asymmetric trapezoidal function, is a Gaussian function, w is the weight; the delivery uncertainty embedding vector The first five components of are transformed to determine the parameters of the membership function: Central moment of an asymmetric trapezoidal function ; Left slope of the asymmetric trapezoidal function ; Right slope of the asymmetric trapezoidal function ; Mean of the Gaussian function ; Variance of the Gaussian function ; in to and to is the preset transformation coefficient.

4. The AI-based supply chain scheduling method according to claim 1, characterized in that: The reward function of the reinforcement learning model is calculated as follows: For each task i in the scheduling plan, get the task planned completion time , and calculate the membership value based on the composite morphological fuzzy time window of the destination node of the task ; Calculating performance penalties : , where the sum is over all tasks in the scenario; Compare the resources allocated to each task in the current scheduling plan with the scheduling plan of the previous scheduling cycle one by one, and count the number of tasks with resource changes ; Calculating stability penalty : , where c is the preset positive constant weight; Reward Function .

5. 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 task dependency directed acyclic graphs of the two parent scheduling schemes, including: Represent the two parent scheduling schemes as a task dependency directed acyclic graph; Using the critical path method, in each task dependency directed acyclic graph, all task nodes whose earliest start time is equal to the latest start time are identified, and the subgraph consisting of the task nodes and the dependency edges between the task nodes 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 several consecutive task nodes and dependency edges; From the critical path subgraph of the second parent solution, randomly select a path segment D2 that is equal to D1 in the number of tasks; D1 is swapped with D2, and the disconnected dependency arcs in the two subgraphs after the swap are reconnected through the repair operation to generate two new child scheduling schemes.

6. The AI-based supply chain scheduling method according to claim 2, characterized in that: The mutation operation of the genetic algorithm is to select tasks corresponding to high-norm delivery uncertainty embedding vectors, match alternative resource paths for the tasks, and constrain the incremental cost of the matched alternative resource paths to an upper limit, which decreases as the membership value of the task's planned completion time in the original scheduling scheme increases. The method includes: Calculate the L2 norm of the delivery uncertainty embedding vector corresponding to each task in the scheduling plan, and classify tasks with norm values ​​in the highest percentile range as high-risk tasks; A new resource path is matched for each high-risk task from the alternative resource pool. The planned transportation time of the new resource path must not be longer than that of the original resource path. Calculate the incremental cost of the new resource path , and obtain the membership value of the task's planned completion time in the original scheduling plan ; The upper limit of the cost increment is calculated by the following formula; ,in is the basic cost constraint value, is the upper limit of cost increment, k is the adjustment coefficient greater than or equal to 1; The membership value of the task's planned completion time in the original scheduling plan; when When the resource allocation is completed, the resource is reallocated.

7. An AI-based supply chain scheduling system, characterized by: Includes the following modules: The embedding module is used to obtain supply chain orders, resources, logistics topology data, and historical fulfillment data. It uses a graph neural network to process the logistics topology data and historical fulfillment data to generate a delivery uncertainty embedding vector for the destination node of each scheduling task. A fuzzy time window construction module is used to determine the central moment and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function based on the dimensional components 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 a reinforcement learning model. The action space of the reinforcement learning model is a multidimensional continuous space, which is used to output the weighted combination of the cost term, time term, and robustness term 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 that represents the structural difference between the scheduling plans of the previous and next scheduling cycles. The scheduling scheme generation module is used to determine the scheduling objective function based on the weight combination and use the genetic algorithm to schedule and solve the scheduling objective function; the crossover operation of the genetic algorithm is the key subgraph exchange, which is achieved by identifying and exchanging the key path subgraphs in the task dependency directed acyclic graph of the two parent scheduling schemes.

8. The AI-based supply chain scheduling system according to claim 7, characterized in that: The mutation operation of the genetic algorithm is to select tasks corresponding to high-norm delivery uncertainty embedding vectors, match alternative resource paths for the tasks, and subject the cost increment of the matched alternative resource paths to an upper limit constraint, wherein the upper limit decreases as the membership value of the task's planned completion time in the original scheduling scheme increases.

9. The AI-based supply chain scheduling system according to claim 7, characterized in that: The central moment and slope of the asymmetric trapezoidal function and the mean and variance of the Gaussian function are determined according to the dimensional components of the delivery uncertainty embedding vector, and 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, including: Membership function of fuzzy time window The calculation formula is: ;in, is an asymmetric trapezoidal function, is a Gaussian function, w is the weight; the delivery uncertainty embedding vector The first five components of are transformed to determine the parameters of the membership function: Central moment of an asymmetric trapezoidal function ; Left slope of the asymmetric trapezoidal function ; Right slope of the asymmetric trapezoidal function ; Mean of the Gaussian function ; Variance of the Gaussian function ; in to and to is the preset transformation coefficient.

10. The AI-based supply chain scheduling system according to claim 7, characterized in that: The reward function of the reinforcement learning model is calculated as follows: For each task i in the scheduling plan, get the task planned completion time , and calculate the membership value based on the composite morphological fuzzy time window of the destination node of the task ; Calculating performance penalties : , where the sum is over all tasks in the scenario; Compare the resources allocated to each task in the current scheduling plan with the scheduling plan of the previous scheduling cycle one by one, and count the number of tasks with resource changes ; Calculating stability penalty : , where c is the preset positive constant weight; Reward Function .

Images