A collaborative supply chain prediction method based on a graph neural network

Abstract

This invention relates to the field of supply chain forecasting technology and provides a collaborative supply chain forecasting method based on graph neural networks. The invention utilizes the high consistency between graph neural networks and the supply chain structure to transform the supply chain structure into a graph. It also introduces attention mechanisms and time series analysis, enhancing the information density of the graph through the attention mechanism and then using graph neural networks to perform time series analysis on the supply chain data, ultimately outputting the forecast results. This invention can more accurately predict future demand, inventory, and transportation in the supply chain, providing enterprises with more accurate inventory control and order management. Furthermore, the high-complexity model employed can better adapt to the relationships between different nodes in the supply chain, thereby improving forecast accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the field of supply chain forecasting technology, and in particular to a collaborative supply chain forecasting method based on graph neural networks. Background Technology

[0002] In real-world supply chains, the complexity of collaborative supply relationships and multiple real-world factors have made supply chain forecasting increasingly important. To achieve a more efficient and cost-effective supply chain system, effective management and coordination of complex supply chain structures are necessary. Accurate supply chain forecasting enables enterprises to provide more refined inventory control and order management, thereby reducing inventory costs, improving production efficiency and customer satisfaction, and achieving efficient and stable supply chain operation. At the same time, multiple real-world factors make supply chain demand uncertain, and the contradiction between customer demand and inventory rates is constantly intensifying, making traditional forecasting methods inadequate for addressing the problems of real-world supply chains.

[0003] Traditional supply chain forecasting methods employ a range of approaches, including time series analysis and regression analysis, to predict future demand, inventory, and transportation. Time series analysis, based on historical data, analyzes historical patterns and trends to predict future demand and inventory levels. Commonly used methods include moving averages, exponential smoothing, ARIMA, and Markov models. Regression analysis, on the other hand, establishes a regression equation to model the relationship between one or more independent and dependent variables, thereby predicting future demand or inventory levels. However, these methods have limitations in addressing the complexities and heterogeneity of supply chain structures. Real-world supply chains involve multiple suppliers, customers, and complex supply coordination mechanisms, making it difficult for these methods to account for the interactions between various complex supply chain factors and thus hindering accurate predictions of future supply chain conditions. Therefore, to better address the complexity and heterogeneity of real-world supply chains, it is necessary to research and develop more advanced supply chain forecasting methods.

[0004] Compared to traditional mathematical modeling methods, deep learning models can better address the nonlinearity, non-stationarity, and heterogeneity of supply chain data. Currently, deep learning-based supply chain forecasting methods mainly include models based on recurrent neural networks (RNNs), convolutional neural networks (CNNs), and deep autoencoders (DAEs). Among these, RNNs are one of the most widely used deep learning models, capable of modeling and predicting time-series data, a characteristic well-suited to supply chain forecasting. CNNs can extract the connections between different nodes in complex supply chain structures, obtaining structural information for prediction. DAEs can autoencode and extract features from supply chain data, improving prediction accuracy and robustness. However, these deep learning models do not perfectly match the structural characteristics of supply chains, and when dealing with graph-like structures like supply chains, they are prone to generating redundant information unrelated to the supply chain structure, thus affecting prediction accuracy.

[0005] Traditional mathematical modeling and neural network methods are limited in their ability to handle dynamic supply chains in supply chain forecasting. Dynamic supply chains include the addition and disappearance of retailers or distributors, as well as changes in the relationships between them. Such issues often exist in real-world supply chain systems, limiting the application of these two types of models to static supply chain problems. Summary of the Invention

[0006] To address the aforementioned issues, this application proposes a collaborative supply chain forecasting method based on graph neural networks. The forecasting model, based on graph neural networks, leverages the high consistency between the graph neural network and the supply chain structure to transform the supply chain structure into a graph. Simultaneously, it introduces attention mechanisms and time series analysis concepts. The attention mechanism enhances the information density of the graph, and the graph neural network is then used to perform time series analysis on the supply chain data, ultimately outputting the forecast results. This invention can more accurately predict future demand, inventory, and transportation in the supply chain, providing enterprises with more accurate inventory control and order management. Furthermore, the high-complexity model employed can better adapt to the relationships between different nodes in the supply chain, thereby improving forecast accuracy and reliability.

[0007] This application provides a collaborative supply chain forecasting method based on graph neural networks, comprising the following steps:

[0008] S10. Convert supply chain information into a timeline diagram;

[0009] Each frame of the time sequence diagram corresponds to a time node. The nodes on the time sequence diagram represent nodes in the supply chain, and the edges between nodes represent the cost information for supplying goods between two nodes.

[0010] S20. Perform spatial dimension feature aggregation on the time series graph obtained in step S10, and output the time series graph with enhanced features;

[0011] S30. Perform time-dimensional feature aggregation on the time series graph obtained in step S20 to obtain a supply chain information prediction graph;

[0012] S40. Decode the supply chain information prediction map to obtain the final prediction result.

[0013] Preferably, in step S20, the following convolution formula is used to perform convolution calculation on each frame of the time sequence map:

[0014] ,

[0015] in, Indicates the first The feature matrix of the time series map during layer convolution calculation, l=1,2,...,L; L is the total number of convolution calculation layers; For activation function, Indicates the first The weight matrix is ​​used during layer convolution calculation; This is the adjacency matrix of the time series graph. A ij Let j be the adjacency matrix of node i in the time series graph, and j be the neighboring nodes of node i. yes The degree matrix;

[0016] This yields the feature information of each node in each frame's time sequence diagram.

[0017] Preferably, step S30 includes the following steps:

[0018] S31. Convert the timing information in the timing graph into planar information, and update the edge-level events and point-level events in the planar information graph;

[0019] The update of the edge-level event is as follows:

[0020] ,

[0021] ,

[0022] Where t is the t-th time in the time series, m i (t) and m j (t) represents the feature information of node i and node j after the interaction event between nodes j; mag s () and msg d () is a learnable edge-level information transfer function, mag s () and msgd () are the same or different; and These are the updated feature information of node i and node j at time t-1, respectively; Let be the time difference between time t and time t-1. (t) represents an edge-level event, specifically indicating changes in information such as demand and transportation costs between two nodes in the supply chain;

[0023] The aforementioned point-level events, namely the updates to information such as changes in cargo inventory and production capacity at a single node in the supply chain, are as follows:

[0024] ,

[0025] in, For the feature information of node i after event update at time t, mag n () represents a learnable point-level information transfer function. The feature information of node i at time t obtained in step S20;

[0026] The updates of edge-level events and point-level events are not sequential;

[0027] S32. Calculate the aggregated feature information for each node;

[0028] ,

[0029] in, , indicating from time t1 to t n The aggregated feature information of node i at time i, where agg() is the aggregation function;

[0030] S33. Update the node information for each node;

[0031] ;

[0032] Among them, s i (t) represents the updated node information of node i at time t, and mem() is the update function;

[0033] S34. Encode the updated node information;

[0034] ;

[0035] Where h() is the graph attention function, This represents the coupling information between node i and all nodes and edges related to node i from time 0 to time t.

[0036] The collaborative supply chain forecasting method based on graph neural networks of the present invention has at least the following advantages compared with the prior art:

[0037] This invention utilizes a graph neural network-based prediction model that can more accurately predict future demand, inventory, and transportation within the supply chain, providing enterprises with more precise inventory control and order management. Furthermore, the high-complexity model employed better adapts to the relationships between different nodes in the supply chain, thereby improving prediction accuracy and reliability. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart illustrating a collaborative supply chain forecasting method based on graph neural networks, as described in this application. Detailed Implementation

[0040] The following description provides many different embodiments or examples for implementing various features of the invention. The elements and arrangements described in the specific examples below are only for concise expression of the invention and are merely examples, not intended to limit the invention.

[0041] A collaborative supply chain forecasting method based on graph neural networks, such as Figure 1 As shown, it includes the following steps:

[0042] S10. Convert supply chain information into a timeline diagram;

[0043] Each frame of the time sequence diagram corresponds to a time node. The nodes on the time sequence diagram represent nodes in the supply chain, and the edges between nodes represent the cost information for supplying goods between two nodes.

[0044] Those skilled in the art will understand that the supply chain includes at least two of the following: raw materials, production, warehousing, transportation, and sales points; the cost information includes information such as storage capacity, production quantity, transportation costs, demand, and sales quantity corresponding to the supply chain.

[0045] In this step, the supply chain information is arranged in chronological order to form T-shaped graphs with time information.

[0046] S20. Perform spatial dimension feature aggregation on the time series graph obtained in step S10, and output the time series graph with enhanced features;

[0047] In step S20, the following convolution formula is used to perform convolution calculation on each frame of the temporal map:

[0048] ,

[0049] in, Indicates the first The feature matrix of the time series map during layer convolution calculation, l=1,2,...,L; L is the total number of convolution calculation layers; For activation function, Indicates the first The weight matrix is ​​used during layer convolution calculation; This is the adjacency matrix of the time series graph. A ij Let j be the adjacency matrix of node i in the time series graph, and j be the neighboring nodes of node i. yes The degree matrix, ;

[0050] Through continuous iteration, the feature information of each node in each frame of the time sequence graph can be obtained.

[0051] Preferably, in this step, the timing diagram obtained in step S10 is input into the convolutional neural network in step S20 for calculation using a sliding window.

[0052] S30. Perform time-dimensional feature aggregation on the feature-enhanced time series graph obtained in step S20 to obtain a supply chain information prediction graph; step S30 includes the following steps:

[0053] S31. Convert the time sequence information in the time sequence diagram into planar information, and update the edge-level events and node-level events in the planar information diagram; edge-level events are changes in demand between nodes, changes in transportation costs, etc., and node-level events are increases or decreases in production capacity, increases or decreases in sales volume, increases or decreases in warehouse quantity, etc.

[0054] The update of the edge-level event is as follows:

[0055] ,,

[0056] ,

[0057] Where t is the t-th time in the time series, m i (t) and m j (t) represents the feature information of node i and node j after the interaction event between nodes j; mag s () and msg d () is a learnable edge-level information transfer function, mag s () and msg d () are the same or different; and These are the updated feature information of node i and node j at time t-1 (see step S33). Let be the time difference between time t and time t-1, i.e., the time step. (t) represents an edge-level event;

[0058] in, and These are the updated feature information for nodes i and j at time t-1, respectively. However, nodes ij at the initial time do not have updated feature information. Those skilled in the art can choose the following method to incorporate the feature information of nodes ij at the initial time, depending on the circumstances:

[0059] make ,or Accordingly, let ,or In other words, the feature information of node ij at the initial time is set to 0, or the feature information data of node ij obtained by convolution calculation in step S20 is directly used.

[0060] The update of the point-level event is as follows:

[0061] ,

[0062] in, For the feature information of node i after event update at time t, mag n () represents a learnable point-level information transfer function. The feature information of node i at time t obtained in step S20;

[0063] Among them, mag s (), msg d () and mag n All of them can be learned information transfer functions, such as MLPs.

[0064] The updates of edge-level events and point-level events are not sequential;

[0065] S32. Calculate the aggregated feature information for each node;

[0066] Considering that batch processing may require updating nodes simultaneously at any given moment, an aggregation method is used to combine the features of each node to reduce the amount of computational information.

[0067] ,

[0068] in, , indicating from time t1 to t n The aggregated feature information of node i at time i, where agg() is the aggregation function.

[0069] S33. Update the node information for each node;

[0070] For longer time-series data, networks tend to "forget" earlier time-series data during processing. Therefore, a memory module is needed to enhance the connection between features of different time series. The output of this module is used to update the memory data already stored in this module, and also to update the prediction data of the embedded modules.

[0071] In time-graph networks, node features need to be dynamically updated through corresponding events, edges need to interact with events to update the features of associated node pairs, and node events also need to update their corresponding node information. The specific update formula is as follows:

[0072] ;

[0073] Among them, s i (t) represents the updated node information of node i at time t, and mem() is the update function, which uses an LSTM network as the memory update module.

[0074] S34. Encode the updated node information;

[0075] To update the state of the nodes mentioned above, the features of node i at time t need to be encoded to obtain the embedded representation:

[0076] ;

[0077] Where h() is the graph attention function, that is, the graph attention layer is selected for feature focusing; This represents the coupling information between node i and all nodes and edges related to node i from time 0 to time t.

[0078] S40. Decode the supply chain information prediction graph to obtain the final prediction result. For the model predicting the supply and production of goods in the supply chain, the nodes of the input graph represent the production and sales capacity of manufacturers, producers, and distributors in the supply chain; the edges of the graph represent the demand for goods and transportation costs between any two nodes. The input data consists of changes in the production and sales data of manufacturers, producers, and distributors in the supply chain over a period of time, as well as changes in the supply volume between different supply chain levels. The final output result is the production capacity of different supply chain nodes and their demand for goods at the next moment.

[0079] The data calculated above is essentially a graph encoded by a graph encoder, and therefore needs to be decoded to restore the graph data. Therefore, the aforementioned output is input into a decoder composed of a multilayer perceptron (MLP), which decodes the data and outputs the final prediction result. This step is a conventional technique and will not be described in detail in this application.

[0080] Traditional mathematical modeling methods require extensive parameter adjustments and calculations, a process that is often tedious and time-consuming, and demands specialized mathematical knowledge and skills, posing a significant challenge for non-professionals. However, this application presents a collaborative supply chain forecasting method based on graph neural networks, eliminating this process. The selection and calculation of parameters are determined automatically by the network model, thus reducing the difficulty of modeling complex supply chains.

[0081] Furthermore, compared to conventional neural network methods, the graph neural network employed in this method possesses stronger expressive power, can handle data in non-Euclidean space, and exhibits greater interpretability. Therefore, it can effectively extract structural features of the supply chain, thereby achieving higher prediction accuracy. Simultaneously, the model used in this method can handle dynamic supply chain forecasting problems, thus addressing more complex supply chain forecasting challenges.

[0082] This method is used to predict the supply chain in a three-tier model comprising suppliers, manufacturers, and distributors, where goods flow from the supplier, through the manufacturer, and to the distributor. First, the supply chain relationships are transformed into a one-way graph, where nodes represent suppliers, manufacturers, and distributors, and node data represents the production or sales capacity of each entity. Edges represent the existence of a goods flow relationship between two nodes, and edge data represents the quantity of goods supplied between the two nodes. The model's input data includes two dimensions: time and space. The spatial dimension represents the supply chain topology, and the temporal dimension represents changes in production and sales capacity and the quantity of goods supplied within the supply chain.

[0083] During prediction, firstly, a graph convolutional network is used to encode each frame of the input time series graph (step S10), fusing node and edge data to obtain new node feature representations. The output is a time series graph with the same topology as the input graph, but with aggregated and enhanced data (step S20). Then, the aggregated and enhanced time series graph is input into a time graph network, which models historical cargo flow data and predicts the production or sales volume of goods at each node in the future time period, as well as the supply of goods between two nodes. The output of the time graph neural network is a graph with the same topology as the input graph, but with aggregated and enhanced data (step S30).

[0084] Finally, the output graph of the time-series neural network is decoded by MLP to restore the aggregated and enhanced data features into data with real-world significance, thus obtaining the graph of the next moment of the input time-series graph, which is the demand state of the supply chain at the next moment (step S40).

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A collaborative supply chain forecasting method based on graph neural networks, characterized in that, Includes the following steps: S10. Convert supply chain information into a timeline diagram; Each frame of the time sequence diagram corresponds to a time node. The nodes on the time sequence diagram represent nodes in the supply chain, and the edges between nodes represent the cost information for supplying goods between two nodes. S20. Perform spatial dimension feature aggregation on the time series graph obtained in step S10, and output the time series graph with enhanced features; S30. Perform time-dimensional feature aggregation on the time series graph obtained in step S20 to obtain a supply chain information prediction graph; S40. Decode the supply chain information prediction map to obtain the final prediction result; Step S30 includes the following steps: S31. Convert the timing information in the timing graph into a planar information graph, and update the edge-level events and point-level events in the planar information graph; The update of the edge-level event is as follows: , ， Where t is the t-th time in the time series, m i (t) and m j (t) represents the feature information of node i and node j after the interaction event between nodes j; mag s () and msg d () is a learnable edge-level information transfer function, mag s () and msg d () are the same or different; and These are the updated feature information of node i and node j at time t-1, respectively; Let be the time difference between time t and time t-1. (t) represents an edge-level event, specifically indicating changes in demand and transportation cost information between two nodes in the supply chain; The point-level event, namely the update of changes in the inventory and production capacity information of a single node in the supply chain, is as follows: , in, For the feature information of node i after event update at time t, mag n () represents a learnable point-level information transfer function. The feature information of node i at time t obtained in step S20; The updates of edge-level events and point-level events are not sequential; S32. Calculate the aggregated feature information for each node; ， in, , indicating from time t1 to t n The aggregated feature information of node i at time i, where agg() is the aggregation function; S33. Update the node information for each node; ； Among them, s i (t) represents the updated node information of node i at time t, and mem() is the update function; S34. Encode the updated node information; ； Where h() is the graph attention function, This represents the coupling information between node i and all nodes and edges related to node i from time 0 to time t. The update function uses an LSTM network as the memory update module; and / or, the information transfer function uses MLPs.

2. The collaborative supply chain forecasting method based on graph neural networks according to claim 1, characterized in that, In step S20, the following convolution formula is used to perform convolution calculation on each frame of the temporal map: ， in, Indicates the first The feature matrix of the time series map during layer convolution calculation, l=1,2,...,L; L is the total number of convolution calculation layers; For activation function, Indicates the first The weight matrix is ​​used during layer convolution calculation; This is the adjacency matrix of the time series graph. A ij Let j be the adjacency matrix of node i in the time series graph, and j be the neighboring nodes of node i. yes The degree matrix; This yields the feature information of each node in each frame's time sequence diagram.

3. A collaborative supply chain forecasting method based on graph neural networks according to any one of claims 1-2, characterized in that, The supply chain includes at least two of the following: raw materials, production, warehousing, transportation, and sales points; the cost information includes information on warehousing capacity, production quantity, transportation cost, demand, and sales quantity corresponding to the supply chain.

4. The collaborative supply chain forecasting method based on graph neural networks according to claim 3, characterized in that, make ,or Accordingly, let ,or .

Images