CN-116341752-B - Collaborative supply chain prediction method based on graph neural network

Abstract

The invention is suitable for the technical field of supply chain prediction, and provides a collaborative supply chain prediction method based on a graph neural network. The invention is based on a prediction model of a graphic neural network, converts a supply chain structure into a graph by utilizing the high consistency of the graphic neural network and the supply chain structure, introduces the ideas of an attention mechanism and time sequence analysis, strengthens the information density of the graph by the attention mechanism, then performs time sequence analysis on the supply chain data by utilizing the graphic neural network, and finally outputs the obtained prediction result. The invention can more accurately forecast future conditions of demand, stock, transportation and the like in the supply chain, and provides more accurate stock control and order management for enterprises. In addition, the adopted high-complexity model can be better adapted to the relation between different nodes in the supply chain, so that the prediction accuracy and reliability are improved.

Inventors

• MIAO QIANG
• NIU TIANYU
• YAN XINGYOU
• ZHANG HENG
• Lei Chunrui

Assignees

• 四川大学

Dates

Publication Date

20260505

Application Date

20230404

Claims (4)

1. 1. The collaborative supply chain prediction method based on the graph neural network is characterized by comprising the following steps of: s10, converting the supply chain information into a time sequence diagram; Each frame of the time sequence diagram corresponds to a time node, the nodes on the time sequence diagram represent nodes of a supply chain, and edges between the nodes represent cost information for supplying goods between the two nodes; s20, performing feature aggregation of space dimensions on the time sequence diagram obtained in the step S10, and outputting the time sequence diagram subjected to feature reinforcement; S30, performing feature aggregation of time dimension on the time sequence diagram subjected to feature reinforcement obtained in the step S20 to obtain a supply chain information prediction diagram; s40, decoding the supply chain information prediction graph to obtain a final prediction result; Wherein, step S30 includes the following steps: S31, converting time sequence information in a time sequence diagram into a plane information diagram, and updating an edge level event and a point level event in the plane information diagram; The updating of the edge level event is as follows: , , wherein t is the t time in the time sequence, m i (t) and m j (t) are the characteristic information of the i node and the j node after the interaction event between the ij nodes respectively, mag s () and msg d () are the leachable side level information transfer functions, and mag s () and msg d () are the same or different; And Updated characteristic information of the i node and the j node at the time t-1 respectively; as the time difference between time t and time t-1, (T) edge level events, specifically representing changes in demand and transportation cost information between two nodes on the supply chain; The point level event, namely the update of the change of the cargo reserve quantity and the production capacity information of a single node on the supply chain is as follows: , Wherein, the For the feature information updated by node i event at time t, mag n () is a learnable point-level information transfer function, The characteristic information of the node i at the t moment obtained in the step S20; updating of the edge level event and the point level event is not sequenced; S32, calculating the aggregation characteristic information of each node; , Wherein, the Aggregation characteristic information representing inodes from a time t 1 to a time t n , wherein agg () is an aggregation function; s33, updating node information of each node; ; Wherein s i (t) is updated node information at the moment of the i node t, and mem () is an update function; s34, coding the updated node information; ; Where h () is the graph meaning force function, The coupling information of all relevant nodes and edges and i-nodes between 0 and t time and i-nodes; The updating function selects LSTM network as memory updating module, and/or the information transfer function selects MLPs.
2. 2. The collaborative supply chain prediction method based on a graph neural network according to claim 1, wherein in step S20, a convolution calculation is performed on each frame timing chart using the following convolution formula: , Wherein, the Represent the first The characteristic matrix of the time sequence diagram during layer convolution calculation, i=1, 2, & gt, L; in order to activate the function, Represent the first Layer convolution calculates a weight matrix; For the adjacency matrix of the timing diagram, A ij is an adjacent matrix of a node i in the time sequence diagram, and j is a neighbor node of the node i; Is that A degree matrix of (2); thus, the characteristic information of each node in each frame time sequence diagram is obtained.
3. 3. The method for predicting a collaborative supply chain based on a graphic neural network according to any one of claims 1-2, wherein the supply chain comprises at least two of raw materials, production, warehouse, transportation, and point of sale, and the cost information comprises warehouse capacity, production quantity, transportation cost, demand quantity, and sales quantity information corresponding to the supply chain.
4. 4. A coordinated supply chain prediction method based on a graph neural network according to claim 3, wherein the method comprises the steps of Or (b) Correspondingly, let the Or (b) 。

Description

Collaborative supply chain prediction method based on graph neural network Technical Field The invention relates to the technical field of supply chain prediction, in particular to a collaborative supply chain prediction method based on a graph neural network. Background In a real supply chain, the complexity of collaborative supply relationships and multiple real factors have led to an increasing emphasis on supply chain predictions. To achieve a more efficient and cost effective supply chain system, efficient management and coordination of complex supply chain structures is required. Accurate supply chain prediction can provide more refined inventory control and order management for enterprises, thereby reducing inventory cost, improving production efficiency and customer satisfaction, and realizing efficient and stable operation of the supply chain. Meanwhile, the multiple reality factors make the demand change of the supply chain uncertain, so that the contradiction between the client demand and the stock rate is continuously increased, and the problem of the real supply chain is difficult to be solved by the traditional prediction method. Conventional supply chain prediction methods employ a series of methods in predicting future demands, inventory, transportation, etc., including time series analysis, regression analysis, etc. The time sequence analysis is a method for predicting based on historical data, and by analyzing the time sequence data, rules and trends of the historical data are mastered so as to predict future demands and stock conditions, and common analysis methods include moving average, exponential smoothing, ARIMA and Markov model. Regression analysis is to build a mathematical model of the relationship between one or more independent and dependent variables by building a regression equation, thereby predicting future demand or inventory conditions. However, these methods have limitations in coping with the problems of complicated supply chain structure and heterogeneity. Because of the existence of multiple suppliers, multiple customers, and complex supply coordination mechanisms in real supply chains, these methods often fail to take into account the interactions between the various complex supply chain factors, making it difficult to accurately predict future conditions of the supply chain. Thus, to better address the complexity and heterogeneity of real supply chains, more advanced supply chain prediction methods need to be studied and developed. Compared with the traditional mathematical modeling method, the deep learning model can better solve the problems of nonlinearity, non-stationarity and heterogeneity of the supply chain data. Currently, a supply chain prediction method based on deep learning mainly includes a cyclic neural network (RNN) -based model, a Convolutional Neural Network (CNN) -based model, a deep self-encoder (DAE) -based model, and the like. Among them, RNN model is one of the most widely used deep learning models at present, and can model and predict time series data, and the characteristics of RNN model are very suitable for the characteristics of supply chain prediction. The CNN model can be used for extracting the connection between different nodes in the complex supply chain structure and obtaining the supply chain structure information so as to realize prediction. The DAE model can perform self-coding and feature extraction on the supply chain data, and the prediction accuracy and the robustness of the model are improved. However, the deep learning model has low consistency with the structural characteristics of the supply chain, and redundant information irrelevant to the structure of the supply chain is easy to generate when the structure of the map of the supply chain is processed, so that the prediction accuracy of the model is affected. Conventional mathematical modeling and neural network methods suffer from the disadvantage of being unable to cope with the dynamic supply chain in supply chain predictions. Such dynamic supply chains include the new and lost of retailers or distributors, as well as the changing of the links between them. Such problems often exist in practical supply chain systems, such that the two types of models described above can only be used for static supply chain problems. Disclosure of Invention In order to solve the problems, the application provides a collaborative supply chain prediction method based on a graph neural network, which is based on a prediction model of the graph neural network, converts a supply chain structure into a graph by utilizing the high consistency of the graph neural network and the supply chain structure, introduces the ideas of attention mechanism and time sequence analysis, strengthens the information density of the graph by the attention mechanism, performs time sequence analysis on supply chain data by utilizing the graph neural network, and finally outputs a prediction result. The application