Supply chain management optimization method and system based on large model

Abstract

The invention discloses a supply chain management optimization method and system based on a large model. The method comprises the steps of data acquisition and processing, large model demand prediction, inventory management optimization, supply chain operation optimization and supply chain management optimization. The invention relates to the technical field of supply chain management, in particular to a supply chain management optimization method and system based on a large model, according to the scheme, process design of demand prediction, inventory management optimization and operation optimization is sequentially carried out, and complex integrated supply chain data can be better processed; a long-short-term memory neural network combined with a transformer model is adopted to carry out large model demand prediction, and the processing performance of time sequence trend data is optimized; inventory management is carried out by adopting a deep enhanced hybrid linear programming method combined with graph convolution enhancement, and the extraction quality of transport dynamic edge features is improved by adding an aggregation embedding mode of node features and edge features; and a generation bidirectional encoder combined with a transformer model is adopted, so that the intelligence of the method is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of supply chain management, and in particular to a supply chain management optimization method and system based on a large model. Background Art

[0002] The supply chain management optimization method based on large models uses deep learning technology driven by large language models to integrate and analyze massive data in various links of the supply chain to improve the accuracy and efficiency of decision-making. This method helps companies predict market demand, optimize inventory levels, and improve the coordinated operation of various nodes in the supply chain through three steps: demand forecasting, inventory management optimization, and supply chain operation optimization, thereby reducing costs, improving efficiency, and enhancing market competitiveness. This method can cope with complex and changing market environments and provide intelligent, data-driven decision support for supply chain management.

[0003] However, in the existing supply chain management methods, there is a technical problem that supply chain data generally comes from different special processing systems, which leads to the need for supply chain management to face greater data complexity integration and more difficult decision-making; in the existing supply chain management methods, there is a technical problem that demand forecasting relies on large-scale market data, and accurate demand forecasting provides a basis for subsequent inventory adjustments and operation adjustments to a large extent, and the accuracy of existing methods needs to be improved; in the existing inventory management optimization methods, inventory management modeling through graph structure is a common technical solution, but traditional methods pay more attention to node features and easily ignore edge features when constructing graph convolutional networks, which further leads to excessive consideration of fixed-point features of suppliers, warehouses and retailers as nodes during inventory management, but ignores edge features of transportation dynamics, thereby reducing the accuracy of management; in the existing supply chain operation optimization methods, there is a technical problem that the supply chain operation output of traditional methods is not intuitive enough and cannot effectively provide direct help to actual staff. Summary of the invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a supply chain management optimization method and system based on a large model. In view of the technical problem that in the existing supply chain management method, supply chain data generally comes from different special processing systems, which leads to the supply chain management facing a large data complexity integration and more difficult decision-making, this solution creatively adopts a process design that sequentially performs demand forecasting, inventory management optimization and operation optimization, which can better handle complex integrated supply chain data and perform various specific management tasks, and provide a basis for subsequent inventory management and operation management by predicting market demand, and improve the quality of supply chain management through inventory management and operation management. At the same time, the sub-item functional design is also helpful for the operation and maintenance of the system; in view of the technical problem that in the existing supply chain management method, demand forecasting depends on large-scale market data, and accurate demand forecasting provides a basis for subsequent inventory adjustment and operation adjustment to a large extent, and the accuracy of the existing method needs to be improved, this solution creatively adopts a long short-term memory neural network combined with a transformer model to perform large-model demand forecasting, optimizes the processing performance of time series trend data, and improves the overall flexibility of the model. In the existing inventory management optimization methods, inventory management modeling through graph structure is a common technical solution, but the traditional method pays more attention to node features and easily ignores edge features when constructing graph convolution networks, which further leads to the problem that when managing inventory, too much consideration is given to the fixed-point features of suppliers, warehouses and retailers as nodes, but the edge features of transportation dynamics are ignored, thereby reducing the accuracy of management. This solution creatively adopts a deep reinforced hybrid linear programming method combined with graph convolution enhancement for inventory management. By adding an aggregate embedding method of node features and edge features, the extraction quality of transportation dynamic edge features is improved, thereby optimizing the dynamics, cost and ability to cope with uncertainty problems in the inventory management process; In the existing supply chain operation optimization methods, there is a technical problem that the supply chain operation output of the traditional method is not intuitive enough and cannot effectively provide direct help to actual staff. This solution creatively adopts a generative bidirectional encoder combined with a transformer model to optimize supply chain operations. By training based on text information and then training output text references, the referenceability of the supply chain operation output results is improved, and the intelligence and transparency of the method itself are further improved.

[0005] The technical solution adopted by the present invention is as follows: The supply chain management optimization method based on a large model provided by the present invention comprises the following steps:

[0006] Step S1: data collection and processing;

[0007] Step S2: large model demand forecast;

[0008] Step S3: Inventory management optimization;

[0009] Step S4: Supply chain operation optimization;

[0010] Step S5: Supply chain management optimization.

[0011] Further, in step S1, the data collection process is used to collect the original data set required for large-model supply chain management optimization, and preliminarily preprocess the original data set, specifically, to obtain the supply chain management original data set from the enterprise supply chain association system through data collection, and to obtain the supply chain management optimization data set by sequentially performing multi-source data fusion, natural language data processing, data cleaning and standardization on the supply chain management original data set;

[0012] The supply chain management optimization data set specifically includes a market demand forecasting data set, an inventory management data set, and an operation optimization data set.

[0013] Further, in step S2, the large model demand forecast is used to combine historical data and external data to predict the basic demand indicators of supply chain management, specifically, based on the market demand forecast data set in the supply chain management optimization data set, a long short-term memory neural network combined with a transformer model is used to perform large model demand forecasting to obtain market demand forecast reference data;

[0014] The long short-term memory neural network combined with the transformer model specifically includes a long short-term memory subnet, a self-attention layer, a transformer layer, a decoding layer and a fully connected layer;

[0015] The step of using a long short-term memory neural network combined with a transformer model to perform large model demand forecasting and obtain market demand forecast reference data specifically includes:

[0016] Step S21: constructing input data, specifically extracting historical data and external information data from the market demand forecast data set, and performing time series data processing to obtain a market demand forecast input data set;

[0017] Step S22: constructing a long short-term memory subnet, specifically constructing a standard long short-term memory neural network including an input gate, an output gate and a forget gate, to obtain the long short-term memory subnet;

[0018] Step S23: constructing a self-attention layer, specifically, using the output hidden state of the long short-term memory subnetwork as the input of the self-attention layer, and constructing a query matrix, a key matrix and a value matrix, and calculating multi-head attention in parallel to construct the self-attention layer;

[0019] Step S24: constructing a transformer layer, specifically by constructing the self-attention layer, constructing a standard transformer module as the transformer layer, and performing feature encoding to obtain demand forecast transformer encoding feature data;

[0020] Step S25: constructing a decoding layer and a fully connected layer, specifically, using a multi-layer perceptron as a decoder to perform feature decoding on the demand forecast transformer coded feature data to obtain demand forecast decoded feature data, and constructing a fully connected layer to convert the demand forecast decoded feature data into a forecast output through a linear transformation to obtain demand forecast output data;

[0021] Step S26: Demand forecasting model training, specifically, through the input data construction, the long short-term memory subnet, the self-attention layer, the transformer layer, the decoding layer and the fully connected layer, using the mean square error loss function as the model loss function, to perform model optimization training to obtain the demand forecasting model Model DP ;

[0022] Step S27: Market demand forecasting, specifically using the demand forecasting model Model DP , performing market demand forecasting based on the market demand forecasting data set to obtain market demand forecasting reference data;

[0023] The market demand forecast reference data specifically includes product category demand forecast data, product demand quantity forecast data, product regional demand forecast data, product sales trend forecast data and external market impact reference data.

[0024] Further, in step S3, the inventory management optimization is used to generate replenishment and inventory optimization solutions according to the inventory information, specifically, based on the inventory management data set in the supply chain management optimization data set, a deep reinforced hybrid linear programming method combined with graph convolution enhancement is used to perform inventory management optimization to obtain inventory management decision reference data;

[0025] The deep reinforcement hybrid linear programming method combined with graph convolution enhancement specifically includes a hybrid linear programming operator, a graph convolution enhancement feature extraction subnet and a deep reinforcement learning subnet;

[0026] The hybrid linear programming operator is used to generate refined inventory management decisions;

[0027] The graph convolution enhanced feature extraction subnet is used to extract relational features in supply chain inventory data;

[0028] The deep reinforcement learning subnet is used to optimize the modeling of inventory management problems and optimize specific inventory management decision-making solutions;

[0029] The step of optimizing inventory management by using the deep reinforced hybrid linear programming method combined with graph convolution enhancement to obtain reference data for inventory management decisions includes:

[0030] Step S31: Inventory information modeling, specifically, supply chain relationship graph structure modeling, using nodes of the supply chain relationship graph structure to represent key elements in inventory management, and using edges between nodes to represent logistics relationships between key elements;

[0031] The key elements specifically include supplier nodes, warehouse nodes and retailer nodes;

[0032] The logistics relationship specifically includes the replenishment path from the supplier to the warehouse and the delivery path from the warehouse to the retailer;

[0033] Step S32: constructing a graph convolution enhanced feature extraction subnet, specifically, enhancing the graph convolution feature extraction process by aggregating node features and edge features, and constructing a node adjacency matrix and an edge adjacency matrix to aggregate and embed node features and edge features, thereby realizing the construction of an enhanced feature extraction subnet and obtaining the graph convolution enhanced feature extraction subnet;

[0034] Step S33: construct a deep reinforcement learning subnet, specifically defining the state space, dynamic space, reward function and update strategy in sequence to construct the deep reinforcement learning subnet;

[0035] Step S34: constructing a hybrid linear programming operator, specifically constructing a cost coefficient, a decision variable and a constraint matrix, and performing hybrid linear programming modeling for refining specific decisions of inventory management, and optimizing the decision variables according to the constraint matrix by minimizing the cost coefficient to obtain the optimal inventory management output;

[0036] Step S35: Inventory management optimization model training, specifically, by constructing the hybrid linear programming operator, the graph convolution enhanced feature extraction subnet and the deep reinforcement learning subnet, to perform inventory management optimization model training to obtain the inventory management optimization model Model SM ;

[0037] Step S36: Inventory management optimization, specifically using the inventory management data set to optimize inventory management and obtain inventory management decision reference data;

[0038] The inventory management decision reference data specifically includes replenishment quantity data, replenishment time data, inventory allocation plan data and replenishment cost data.

[0039] Further, in step S4, the supply chain operation optimization is used to generate supplier selection and rating information based on the text information, specifically, based on the operation optimization data set in the supply chain management optimization data set, using a generative bidirectional encoder combined with a transformer model to perform supply chain operation optimization and obtain supply chain operation decision reference data;

[0040] The generative bidirectional encoder combined with the transformer model specifically includes a bidirectional encoder backbone layer, a generative pre-trained encoder layer and a dense processing layer;

[0041] The bidirectional encoder backbone layer is used to serve as a backbone network of the model and extract context information from input data;

[0042] The generative pre-trained encoder layer is used to assist in recognizing grammar, semantics and contextual understanding in text;

[0043] The dense processing layer is used to integrate the outputs of the bidirectional encoder and the generative pre-trained encoder and obtain the final supply chain operation optimization result;

[0044] The step of using a bidirectional encoder combined with a transformer model to optimize supply chain operations and obtain reference data for supply chain operation decisions specifically includes:

[0045] Step S41: constructing a bidirectional encoder backbone layer, specifically constructing a bidirectional encoder including a pre-trained model, a hidden layer, and a pooling layer as the bidirectional encoder backbone layer;

[0046] The pre-training model specifically adopts the bert-base-uncased model;

[0047] Step S42: constructing a generative pre-trained encoder layer, specifically using a small generative pre-trained encoder model as the generative pre-trained encoder layer;

[0048] The small generative pre-trained encoder model specifically refers to a small GPT-2 model;

[0049] Step S43: constructing a dense processing layer, specifically constructing a first random dropout layer, a first base layer, a second random dropout layer, and a second dense layer in sequence, and using an S-type activation function in the second dense layer for model enhancement, and using a binary cross entropy loss function as a model loss function;

[0050] Step S44: Supply chain operation optimization model training, specifically, by constructing the bidirectional encoder backbone layer, the generative pre-trained encoder layer and the intensive processing layer, the supply chain operation optimization model training is performed to obtain the supply chain operation optimization model Model RM ;

[0051] Step S45: Supply chain operation optimization, specifically using the supply chain operation optimization model Model RM , based on the operation optimization data set, optimizing the supply chain operation and obtaining reference data for supply chain operation decision-making;

[0052] The supply chain operation decision reference data specifically includes supplier rating information generation data, supplier selection recommendation generation data and supply chain disruption risk analysis data.

[0053] Furthermore, in step S5, the supply chain management optimization is used to perform comprehensive supply chain management based on the results of demand forecasting, inventory optimization and operation optimization, specifically to perform comprehensive supply chain management optimization based on the market demand forecast reference data, the inventory management decision reference data and the supply chain operation decision reference data to obtain supply chain management feedback reference data.

[0054] The supply chain management optimization system based on the big model provided by the present invention includes a data processing module, a prediction optimization module, an intelligent decision-making module and a management interaction module;

[0055] The data processing module is used for data collection and processing, and obtains a supply chain management optimization data set through data collection and processing, and sends the supply chain management optimization data set to the prediction optimization module and the intelligent decision-making module;

[0056] The forecasting optimization module is used for large-model demand forecasting and inventory management optimization. Through large-model demand forecasting and inventory management optimization, market demand forecasting reference data and inventory management decision reference data are obtained, and the market demand forecasting reference data and inventory management decision reference data are sent to the management interaction module;

[0057] The intelligent decision-making module is used for supply chain operation optimization, obtains supply chain operation decision reference data through supply chain operation optimization, and sends the supply chain operation decision reference data to the management interaction module;

[0058] The management interaction module is used for supply chain management optimization, and through supply chain management optimization, supply chain management feedback reference data is obtained.

[0059] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0060] (1) In the existing supply chain management methods, supply chain data generally comes from different special processing systems, which leads to the technical problem that supply chain management needs to face greater data complexity integration and more difficult decision-making. This solution creatively adopts the process design of demand forecasting, inventory management optimization and operation optimization in sequence, which can better handle complex integrated supply chain data and execute various specific management tasks. It provides a basis for subsequent inventory management and operation management by predicting market demand, and improves the quality of supply chain management through inventory management and operation management. At the same time, the sub-item functional design is also helpful for the operation and maintenance of the system;

[0061] (2) In view of the technical problem that demand forecasting in existing supply chain management methods relies on large-scale market data, and accurate demand forecasting provides a basis for subsequent inventory adjustment and operation adjustment to a large extent, and the accuracy of existing methods needs to be improved, this solution creatively uses a long short-term memory neural network combined with a transformer model to perform large-model demand forecasting, optimizes the processing performance of time series trend data, and improves the overall flexibility of the model;

[0062] (3) In the existing inventory management optimization methods, inventory management modeling through graph structure is a common technical solution. However, when constructing graph convolutional networks, traditional methods pay more attention to node features and tend to ignore edge features. This further leads to the problem that when managing inventory, too much consideration is given to the fixed-point features of suppliers, warehouses, and retailers as nodes, but the edge features of transportation dynamics are ignored, thereby reducing the accuracy of management. This solution creatively adopts a deep reinforced hybrid linear programming method combined with graph convolution enhancement for inventory management. By adding an aggregate embedding method for node features and edge features, the extraction quality of transportation dynamic edge features is improved, thereby optimizing the dynamics, cost, and ability to cope with uncertainty problems of the inventory management process;

[0063] (4) In view of the technical problem that the supply chain operation output of traditional methods is not intuitive enough and cannot effectively provide direct help to actual staff in existing supply chain operation optimization methods, this solution creatively adopts the generative bidirectional encoder combined with the transformer model to optimize supply chain operations. By training based on text information and then training the output text reference, the referenceability of the supply chain operation output results is improved, and the intelligence and transparency of the method itself are further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A schematic diagram of the flow of the supply chain management optimization method based on a large model provided by the present invention;

[0065] Figure 2A schematic diagram of a supply chain management optimization system based on a large model provided by the present invention;

[0066] Figure 3 This is a flow chart of data collection and processing in step S1;

[0067] Figure 4 This is a flow chart of large model demand forecasting in step S2;

[0068] Figure 5 A schematic diagram of the process flow for optimizing inventory management in step S3;

[0069] Figure 6 Schematic diagram of the process of supply chain operation optimization in step S4.

[0070] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0071] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0072] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0073] Example 1, see Figure 1 The present invention provides a large model-based supply chain management optimization method, which includes the following steps:

[0074] Step S1: data collection and processing;

[0075] Step S2: large model demand forecast;

[0076] Step S3: Inventory management optimization;

[0077] Step S4: Supply chain operation optimization;

[0078] Step S5: Supply chain management optimization.

[0079] Example 2, see Figure 1 , Figure 2 and Figure 3 In step S1, the data collection process is used to collect the original data set required for large-model supply chain management optimization and pre-process the original data set. Specifically, the original data set for supply chain management is obtained from the enterprise supply chain association system through data collection, and the supply chain management original data set is sequentially subjected to multi-source data fusion, natural language data processing, data cleaning and standardization to obtain the supply chain management optimization data set. Specifically, the following steps are included:

[0080] Step S11: multi-source data collection, specifically collecting the supply chain management original data set from the enterprise supply chain association system;

[0081] The enterprise supply chain related system specifically includes an enterprise resource planning system, a warehouse management system and a supply chain management system;

[0082] The supply chain management original data set specifically includes enterprise resource planning original data, warehouse management original data and supply chain management original data;

[0083] The enterprise resource planning original data specifically includes financial information data, production plan data, procurement plan data, sales data and customer management data;

[0084] The warehouse management raw data specifically includes inventory quantity data, inventory location data, in-and-out warehouse record data, and cargo tracking reference data;

[0085] The supply chain management raw data specifically includes supplier information data, supply production data, logistics tracking data, inventory management data and basic customer demand data;

[0086] Step S12: natural language data processing, specifically, obtaining structured text data by performing word segmentation and word vector conversion on the text data in the supply chain management original data set;

[0087] Step S13: data cleaning, specifically, performing outlier detection and processing operations on the numerical data in the supply chain management original data set to obtain cleaned and optimized data;

[0088] Step S14: standardization processing, specifically, obtaining standardized supply chain management data by performing manual feature extraction and data standard normalization on the structured text data and the cleansing optimization data;

[0089] Step S15: preliminary preprocessing of the data set, specifically, performing the natural language data processing, the data cleaning and the standardization processing to obtain a supply chain management optimization data set;

[0090] The supply chain management optimization data set specifically includes a market demand forecasting data set, an inventory management data set, and an operation optimization data set.

[0091] By executing the above operations, in view of the fact that in the existing supply chain management methods, supply chain data generally comes from different special processing systems, which leads to the technical problem that supply chain management needs to face greater data complexity integration and more difficult decision-making, this solution creatively adopts the process design of demand forecasting, inventory management optimization and operation optimization in sequence, which can better handle complex integrated supply chain data and execute various specific management tasks, provide a basis for subsequent inventory management and operation management by predicting market demand, and improve the quality of supply chain management through inventory management and operation management. At the same time, the sub-item functional design is also helpful for the operation and maintenance of the system.

[0092] Example 3, see Figure 1 , Figure 2 and Figure 4 , this embodiment is based on the above embodiment. In step S2, the large model demand forecast is used to combine historical data and external data to predict the basic demand indicators of supply chain management. Specifically, based on the market demand forecast data set in the supply chain management optimization data set, a long short-term memory neural network combined with a transformer model is used to perform large model demand forecasting to obtain market demand forecast reference data;

[0093] The long short-term memory neural network combined with the transformer model specifically includes a long short-term memory subnet, a self-attention layer, a transformer layer, a decoding layer and a fully connected layer;

[0094] The step of using a long short-term memory neural network combined with a transformer model to perform large model demand forecasting and obtain market demand forecast reference data specifically includes:

[0095] Step S21: constructing input data, specifically extracting historical data and external information data from the market demand forecast data set, and performing time series data processing to obtain a market demand forecast input data set;

[0096] Step S22: constructing a long short-term memory subnet, specifically constructing a standard long short-term memory neural network including an input gate, an output gate and a forget gate, to obtain the long short-term memory subnet;

[0097] Step S23: constructing a self-attention layer, specifically, using the output hidden state of the long short-term memory subnetwork as the input of the self-attention layer, and constructing a query matrix, a key matrix and a value matrix, and calculating multi-head attention in parallel to construct the self-attention layer;

[0098] Step S24: constructing a transformer layer, specifically by constructing the self-attention layer, constructing a standard transformer module as the transformer layer, and performing feature encoding to obtain demand forecast transformer encoding feature data;

[0099] Step S25: constructing a decoding layer and a fully connected layer, specifically, using a multi-layer perceptron as a decoder to perform feature decoding on the demand forecast transformer coded feature data to obtain demand forecast decoded feature data, and constructing a fully connected layer to convert the demand forecast decoded feature data into a forecast output through a linear transformation to obtain demand forecast output data;

[0100] Step S26: Demand forecasting model training, specifically, through the input data construction, the long short-term memory subnet, the self-attention layer, the transformer layer, the decoding layer and the fully connected layer, using the mean square error loss function as the model loss function, to perform model optimization training to obtain the demand forecasting model Model DP ;

[0101] Step S27: Market demand forecasting, specifically using the demand forecasting model Model DP , performing market demand forecasting based on the market demand forecasting data set to obtain market demand forecasting reference data;

[0102] The market demand forecast reference data specifically includes product category demand forecast data, product demand quantity forecast data, product regional demand forecast data, product sales trend forecast data and external market impact reference data.

[0103] By executing the above operations, in order to address the technical problem that in existing supply chain management methods, demand forecasting relies on large-scale market data, and accurate demand forecasting provides a basis for subsequent inventory adjustments and operational adjustments to a large extent, while the accuracy of existing methods needs to be improved, this solution creatively uses a long short-term memory neural network combined with a transformer model to perform large-model demand forecasting, optimizes the processing performance of time series trend data, and improves the overall flexibility of the model.

[0104] Example 4, see Figure 1 , Figure 2 and Figure 5 This embodiment is based on the above embodiment. In step S3, the inventory management optimization is used to generate replenishment and inventory optimization solutions according to inventory information. Specifically, based on the inventory management data set in the supply chain management optimization data set, a deep reinforced hybrid linear programming method combined with graph convolution enhancement is used to perform inventory management optimization to obtain inventory management decision reference data;

[0105] The deep reinforcement hybrid linear programming method combined with graph convolution enhancement specifically includes a hybrid linear programming operator, a graph convolution enhancement feature extraction subnet and a deep reinforcement learning subnet;

[0106] The hybrid linear programming operator is used to generate refined inventory management decisions;

[0107] The graph convolution enhanced feature extraction subnet is used to extract relational features in supply chain inventory data;

[0108] The deep reinforcement learning subnet is used to optimize the modeling of inventory management problems and optimize specific inventory management decision-making solutions;

[0109] The step of optimizing inventory management by using the deep reinforced hybrid linear programming method combined with graph convolution enhancement to obtain reference data for inventory management decisions includes:

[0110] Step S31: Inventory information modeling, specifically, supply chain relationship graph structure modeling, using nodes of the supply chain relationship graph structure to represent key elements in inventory management, and using edges between nodes to represent logistics relationships between key elements;

[0111] The key elements specifically include supplier nodes, warehouse nodes and retailer nodes;

[0112] The logistics relationship specifically includes the replenishment path from the supplier to the warehouse and the delivery path from the warehouse to the retailer;

[0113] The calculation formula for supply chain relationship graph structure modeling is:

[0114] G = (V, E);

[0115] In the formula, G is the supply chain relationship graph structure, V is the node set used to represent the key elements in warehouse management, and E is the edge set used to represent the logistics relationship between the key elements;

[0116] Step S32: constructing a graph convolution enhanced feature extraction subnet, specifically, enhancing the graph convolution feature extraction process by aggregating node features and edge features, and constructing a node adjacency matrix and an edge adjacency matrix to aggregate and embed node features and edge features, thereby realizing the construction of an enhanced feature extraction subnet and obtaining the graph convolution enhanced feature extraction subnet, specifically comprising the following steps:

[0117] Step S321: construct a node adjacency matrix, the calculation formula is:

[0118] ;

[0119] In the formula, is the adjacency matrix of node v, D vis the degree matrix of node v, A v is the original eigenvalue of node v, is the dimension of the identity matrix corresponding to node v;

[0120] Step S322: node feature aggregation, the calculation formula is:

[0121] ;

[0122] In the formula, is the node feature output of the l+1th graph convolutional layer, R(·) is the nonlinear activation function, T is the binary transformation matrix, is the diagonalization function, is the edge feature output of the lth graph convolutional layer, P e is an auxiliary vector with dimension equal to the number of edges, T T is the transposed matrix of the binary transformation matrix, is the element-wise multiplication operator, is the adjacency matrix of node v, v is the node index, is the node feature output of the lth graph convolutional layer, W v is the node update weight matrix;

[0123] Step S323: construct an edge adjacency matrix, the calculation formula is:

[0124] ;

[0125] In the formula, is the adjacency matrix of edge e, D e is the degree matrix of edge e, A e is the original eigenvalue of edge e, is the dimension of the identity matrix corresponding to edge e;

[0126] Step S324: Edge feature aggregation, the calculation formula is:

[0127] ;

[0128] In the formula, is the edge feature output of the l+1th graph convolutional layer, R(·) is the nonlinear activation function, T is the binary transformation matrix, is the diagonalization function, is the node feature output of the lth graph convolutional layer, P v is an auxiliary vector with a dimension equal to the number of nodes, T T is the transposed matrix of the binary transformation matrix, is the adjacency matrix of edge e, where e is the edge index, is the edge feature output of the lth graph convolutional layer, W eis the edge update weight matrix;

[0129] Step S325: graph convolution enhanced feature extraction, specifically, performing aggregation embedding of node features and edge features through the node feature aggregation and the edge feature aggregation, so as to simultaneously optimize the feature refinement of inventory management on supplier nodes, warehouse nodes, retailer nodes and logistics relationships;

[0130] Step S33: construct a deep reinforcement learning subnet, specifically defining the state space, dynamic space, reward function and update strategy in sequence to construct the deep reinforcement learning subnet;

[0131] The state space specifically includes current inventory, order demand, supplier delivery time, and retailer demand time;

[0132] The action space specifically includes current replenishment decisions and inventory allocation decisions;

[0133] The reward function is specifically designed based on the current inventory level, out-of-stock losses, and order delivery status, and optimizes replenishment and inventory strategies by maximizing long-term rewards;

[0134] The calculation formula of the reward function is:

[0135] Re(s t ,a t )=-C hold (s t )-C shortage (s t )+C delivery (a t );

[0136] Where Re(·) is the reward function, s t is the state space, a t is the action space, C hold (·) is the current inventory, C shortage (·) is the out-of-stock loss, C delivery (·) is the profit from order delivery;

[0137] The updating strategy is specifically performed by constructing a policy gradient method to update the strategy;

[0138] Step S34: constructing a hybrid linear programming operator, specifically constructing a cost coefficient, a decision variable and a constraint matrix, and performing hybrid linear programming modeling for refining specific decisions of inventory management, and optimizing the decision variables according to the constraint matrix by minimizing the cost coefficient to obtain the optimal inventory management output;

[0139] The cost coefficient is specifically designed based on the current inventory, out-of-stock losses and order delivery conditions;

[0140] The decision variable specifically refers to the specific replenishment quantity of the warehouse;

[0141] The constraint matrix is ​​specifically designed based on the transportation capacity, inventory capacity and time constraints of the node;

[0142] Step S35: Inventory management optimization model training, specifically, by constructing the hybrid linear programming operator, the graph convolution enhanced feature extraction subnet and the deep reinforcement learning subnet, to perform inventory management optimization model training to obtain the inventory management optimization model Model SM ;

[0143] Step S36: Inventory management optimization, specifically using the inventory management data set to optimize inventory management and obtain inventory management decision reference data;

[0144] The inventory management decision reference data specifically includes replenishment quantity data, replenishment time data, inventory allocation plan data and replenishment cost data.

[0145] By performing the above operations, in the existing inventory management optimization methods, inventory management modeling through graph structure is a commonly used technical solution, but the traditional method pays more attention to node features and easily ignores edge features when constructing graph convolutional networks, which further leads to excessive consideration of fixed-point features of suppliers, warehouses and retailers as nodes during inventory management, but ignores the edge features of transportation dynamics, thereby reducing the accuracy of management. Technical problems, this solution creatively adopts a deep reinforced hybrid linear programming method combined with graph convolution enhancement for inventory management. By adding an aggregate embedding method of node features and edge features, the extraction quality of transportation dynamic edge features is improved, thereby optimizing the dynamics, cost and ability to cope with uncertainty problems of the inventory management process.

[0146] Example 5, see Figure 1 , Figure 2 and Figure 6 , this embodiment is based on the above embodiment. In step S4, the supply chain operation optimization is used to generate supplier selection and rating information based on text information. Specifically, based on the operation optimization data set in the supply chain management optimization data set, a generative bidirectional encoder combined with a transformer model is used to perform supply chain operation optimization to obtain supply chain operation decision reference data;

[0147] The generative bidirectional encoder combined with the transformer model specifically includes a bidirectional encoder backbone layer, a generative pre-trained encoder layer and a dense processing layer;

[0148] The bidirectional encoder backbone layer is used to serve as a backbone network of the model and extract context information from input data;

[0149] The generative pre-trained encoder layer is used to assist in recognizing grammar, semantics and contextual understanding in text;

[0150] The dense processing layer is used to integrate the outputs of the bidirectional encoder and the generative pre-trained encoder and obtain the final supply chain operation optimization result;

[0151] The step of using a bidirectional encoder combined with a transformer model to optimize supply chain operations and obtain reference data for supply chain operation decisions specifically includes:

[0152] Step S41: constructing a bidirectional encoder backbone layer, specifically constructing a bidirectional encoder including a pre-trained model, a hidden layer, and a pooling layer as the bidirectional encoder backbone layer;

[0153] The pre-training model specifically adopts the bert-base-uncased model;

[0154] Step S42: constructing a generative pre-trained encoder layer, specifically using a small generative pre-trained encoder model as the generative pre-trained encoder layer;

[0155] The small generative pre-trained encoder model specifically refers to a small GPT-2 model;

[0156] Step S43: constructing a dense processing layer, specifically constructing a first random dropout layer, a first base layer, a second random dropout layer, and a second dense layer in sequence, and using an S-type activation function in the second dense layer for model enhancement, and using a binary cross entropy loss function as a model loss function;

[0157] Step S44: Supply chain operation optimization model training, specifically, by constructing the bidirectional encoder backbone layer, the generative pre-trained encoder layer and the intensive processing layer, the supply chain operation optimization model training is performed to obtain the supply chain operation optimization model Model RM ;

[0158] Step S45: Supply chain operation optimization, specifically using the supply chain operation optimization model Model RM , based on the operation optimization data set, optimizing the supply chain operation and obtaining reference data for supply chain operation decision-making;

[0159] The supply chain operation decision reference data specifically includes supplier rating information generation data, supplier selection recommendation generation data and supply chain disruption risk analysis data.

[0160] By performing the above operations, in order to address the technical problem that the supply chain operation output of traditional methods in existing supply chain operation optimization methods is not intuitive enough and cannot effectively provide direct help to actual staff, this solution creatively adopts a generative bidirectional encoder combined with a transformer model to optimize supply chain operations. By training based on text information and then training output text references, the referenceability of the supply chain operation output results is improved, and the intelligence and transparency of the method itself are further improved.

[0161] Example 6, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S5, the supply chain management optimization is used to perform comprehensive supply chain management according to the results of demand forecasting, inventory optimization and operation optimization. Specifically, the supply chain management is comprehensively optimized based on the market demand forecast reference data, the inventory management decision reference data and the supply chain operation decision reference data to obtain supply chain management feedback reference data.

[0162] Embodiment 7, see Figure 1 and Figure 2 , this embodiment is based on the above embodiment, and the supply chain management optimization system based on the big model provided by the present invention includes a data processing module, a prediction optimization module, an intelligent decision-making module and a management interaction module;

[0163] The data processing module is used for data collection and processing, and obtains a supply chain management optimization data set through data collection and processing, and sends the supply chain management optimization data set to the prediction optimization module and the intelligent decision-making module;

[0164] The forecasting optimization module is used for large-model demand forecasting and inventory management optimization. Through large-model demand forecasting and inventory management optimization, market demand forecasting reference data and inventory management decision reference data are obtained, and the market demand forecasting reference data and inventory management decision reference data are sent to the management interaction module;

[0165] The intelligent decision-making module is used for supply chain operation optimization, obtains supply chain operation decision reference data through supply chain operation optimization, and sends the supply chain operation decision reference data to the management interaction module;

[0166] The management interaction module is used for supply chain management optimization, and through supply chain management optimization, supply chain management feedback reference data is obtained.

[0167] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0168] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and alterations may be made to the embodiments without departing from the principles and spirit of the invention.

[0169] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A supply chain management optimization method based on a large model, characterized by: The method comprises the following steps: Step S1: data collection and processing to obtain a supply chain management optimization data set; Step S2: large model demand forecasting, which is used to forecast the basic demand indicators of supply chain management by combining historical data and external data. Specifically, based on the market demand forecasting dataset in the supply chain management optimization dataset, a long short-term memory neural network combined with a transformer model is used to conduct large model demand forecasting to obtain market demand forecasting reference data; Step S3: Inventory management optimization, used to generate replenishment and inventory optimization plans based on inventory information, specifically based on the inventory management data set in the supply chain management optimization data set, using a deep reinforcement hybrid linear programming method combined with graph convolution enhancement to perform inventory management optimization, and obtain inventory management decision reference data, including the following steps: Step S31: Inventory information modeling; Step S32: Constructing a graph convolution enhanced feature extraction subnet; Step S33: Constructing a deep reinforcement learning subnet; Step S34: Constructing a hybrid linear programming operator; Step S35: Inventory management optimization model training; Step S36: Inventory management optimization; Step S4: supply chain operation optimization, used to generate supplier selection and rating information based on text information, specifically, based on the operation optimization data set in the supply chain management optimization data set, using a generative bidirectional encoder combined with a transformer model to perform supply chain operation optimization and obtain supply chain operation decision reference data; Step S5: Optimize supply chain management and obtain supply chain management feedback reference data.

2. The supply chain management optimization method based on a large model according to claim 1, characterized in that: In step S1, the data collection process is used to collect the original data set required for large-scale supply chain management optimization and pre-process the original data set, specifically, to obtain the supply chain management original data set from the enterprise supply chain association system through data collection, and to obtain the supply chain management optimization data set by sequentially performing multi-source data fusion, natural language data processing, data cleaning and standardization on the supply chain management original data set; The supply chain management optimization data set specifically includes a market demand forecasting data set, an inventory management data set, and an operation optimization data set.

3. The supply chain management optimization method based on a large model according to claim 2 is characterized in that: In step S2, the long short-term memory neural network combined with the transformer model specifically includes a long short-term memory subnet, a self-attention layer, a transformer layer, a decoding layer and a fully connected layer; The step of using a long short-term memory neural network combined with a transformer model to perform large model demand forecasting and obtain market demand forecast reference data specifically includes: Step S21: constructing input data, specifically extracting historical data and external information data from the market demand forecast data set, and performing time series data processing to obtain a market demand forecast input data set; Step S22: constructing a long short-term memory subnet, specifically constructing a standard long short-term memory neural network including an input gate, an output gate and a forget gate, to obtain the long short-term memory subnet; Step S23: constructing a self-attention layer, specifically, using the output hidden state of the long short-term memory subnetwork as the input of the self-attention layer, and constructing a query matrix, a key matrix and a value matrix, and calculating multi-head attention in parallel to construct the self-attention layer; Step S24: constructing a transformer layer, specifically by constructing the self-attention layer, constructing a standard transformer module as the transformer layer, and performing feature encoding to obtain demand forecast transformer encoding feature data; Step S25: constructing a decoding layer and a fully connected layer, specifically, using a multi-layer perceptron as a decoder to perform feature decoding on the demand forecast transformer coded feature data to obtain demand forecast decoded feature data, and constructing a fully connected layer to convert the demand forecast decoded feature data into a forecast output through a linear transformation to obtain demand forecast output data; Step S26: Demand forecasting model training, specifically, through the input data construction, the long short-term memory subnet, the self-attention layer, the transformer layer, the decoding layer and the fully connected layer, using the mean square error loss function as the model loss function, to perform model optimization training to obtain the demand forecasting model Model DP ; Step S27: Market demand forecasting, specifically using the demand forecasting model Model DP , performing market demand forecasting based on the market demand forecasting data set to obtain market demand forecasting reference data; The market demand forecast reference data specifically includes product category demand forecast data, product demand quantity forecast data, product regional demand forecast data, product sales trend forecast data and external market impact reference data.

4. The supply chain management optimization method based on a large model according to claim 3 is characterized in that: In step S3, the deep reinforcement hybrid linear programming method combined with graph convolution enhancement specifically includes a hybrid linear programming operator, a graph convolution enhancement feature extraction subnet and a deep reinforcement learning subnet; The hybrid linear programming operator is used to generate refined inventory management decisions; The graph convolution enhanced feature extraction subnet is used to extract relational features in supply chain inventory data; The deep reinforcement learning subnet is used to optimize the modeling of inventory management problems and optimize specific inventory management decision-making solutions.

5. The supply chain management optimization method based on a large model according to claim 4 is characterized in that: In step S3, the step of optimizing inventory management by using the deep reinforced hybrid linear programming method combined with graph convolution enhancement to obtain reference data for inventory management decisions includes: Step S31: Inventory information modeling, specifically, supply chain relationship graph structure modeling, using nodes of the supply chain relationship graph structure to represent key elements in inventory management, and using edges between nodes to represent logistics relationships between key elements; The key elements specifically include supplier nodes, warehouse nodes and retailer nodes; The logistics relationship specifically includes the replenishment path from the supplier to the warehouse and the delivery path from the warehouse to the retailer; Step S32: constructing a graph convolution enhanced feature extraction subnet, specifically, enhancing the graph convolution feature extraction process by aggregating node features and edge features, and constructing a node adjacency matrix and an edge adjacency matrix to aggregate and embed node features and edge features, thereby realizing the construction of an enhanced feature extraction subnet and obtaining the graph convolution enhanced feature extraction subnet; Step S33: construct a deep reinforcement learning subnet, specifically defining the state space, dynamic space, reward function and update strategy in sequence to construct the deep reinforcement learning subnet; Step S34: constructing a hybrid linear programming operator, specifically constructing a cost coefficient, a decision variable and a constraint matrix, and performing hybrid linear programming modeling for refining specific decisions of inventory management, and optimizing the decision variables according to the constraint matrix by minimizing the cost coefficient to obtain the optimal inventory management output; Step S35: Inventory management optimization model training, specifically, by constructing the hybrid linear programming operator, the graph convolution enhanced feature extraction subnet and the deep reinforcement learning subnet, to perform inventory management optimization model training to obtain the inventory management optimization model Model SM ; Step S36: Inventory management optimization, specifically using the inventory management data set to optimize inventory management and obtain inventory management decision reference data; The inventory management decision reference data specifically includes replenishment quantity data, replenishment time data, inventory allocation plan data and replenishment cost data.

6. The supply chain management optimization method based on a large model according to claim 5 is characterized by: In step S4, the bidirectional encoder combined with the transformer model is generated, specifically including a bidirectional encoder backbone layer, a generative pre-trained encoder layer and a dense processing layer; The bidirectional encoder backbone layer is used to serve as a backbone network of the model and extract context information from input data; The generative pre-trained encoder layer is used to assist in recognizing grammar, semantics and contextual understanding in text; The dense processing layer is used to integrate the outputs of the bidirectional encoder and the generative pre-trained encoder and obtain the final supply chain operation optimization result; The step of using a bidirectional encoder combined with a transformer model to optimize supply chain operations and obtain reference data for supply chain operation decisions specifically includes: Step S41: constructing a bidirectional encoder backbone layer, specifically constructing a bidirectional encoder including a pre-trained model, a hidden layer, and a pooling layer as the bidirectional encoder backbone layer; The pre-training model specifically adopts the bert-base-uncased model; Step S42: constructing a generative pre-trained encoder layer, specifically using a small generative pre-trained encoder model as the generative pre-trained encoder layer; The small generative pre-trained encoder model specifically refers to a small GPT-2 model; Step S43: constructing a dense processing layer, specifically constructing a first random dropout layer, a first base layer, a second random dropout layer, and a second dense layer in sequence, and using an S-type activation function in the second dense layer for model enhancement, and using a binary cross entropy loss function as a model loss function; Step S44: Supply chain operation optimization model training, specifically, by constructing the bidirectional encoder backbone layer, the generative pre-trained encoder layer and the intensive processing layer, the supply chain operation optimization model training is performed to obtain the supply chain operation optimization model Model RM ; Step S45: Supply chain operation optimization, specifically using the supply chain operation optimization model Model RM , based on the operation optimization data set, optimizing the supply chain operation and obtaining reference data for supply chain operation decision-making; The supply chain operation decision reference data specifically includes supplier rating information generation data, supplier selection recommendation generation data and supply chain disruption risk analysis data.

7. The supply chain management optimization method based on a large model according to claim 6 is characterized in that: In step S5, the supply chain management optimization is used to perform comprehensive supply chain management based on the results of demand forecasting, inventory optimization and operation optimization. Specifically, the supply chain management is comprehensively optimized based on the market demand forecast reference data, the inventory management decision reference data and the supply chain operation decision reference data to obtain supply chain management feedback reference data.

8. A supply chain management optimization system based on a large model, used to implement the supply chain management optimization method based on a large model as described in any one of claims 1 to 7, characterized in that: It includes data processing module, prediction and optimization module, intelligent decision-making module and management interaction module.

9. The supply chain management optimization system based on a large model according to claim 8, characterized in that: The data processing module is used for data collection and processing, and obtains a supply chain management optimization data set through data collection and processing, and sends the supply chain management optimization data set to the prediction optimization module and the intelligent decision-making module; The forecasting optimization module is used for large-model demand forecasting and inventory management optimization. Through large-model demand forecasting and inventory management optimization, market demand forecasting reference data and inventory management decision reference data are obtained, and the market demand forecasting reference data and inventory management decision reference data are sent to the management interaction module; The intelligent decision-making module is used for supply chain operation optimization, obtains supply chain operation decision reference data through supply chain operation optimization, and sends the supply chain operation decision reference data to the management interaction module; The management interaction module is used for supply chain management optimization, and through supply chain management optimization, supply chain management feedback reference data is obtained.

Images