CN-121998697-A - Multi-source data fusion and deep learning-based supply chain dynamic demand prediction method

CN121998697ACN 121998697 ACN121998697 ACN 121998697ACN-121998697-A

Abstract

The invention relates to the technical field of intelligent management of supply chains, in particular to a supply chain dynamic demand prediction method based on multi-source data fusion and deep learning. The prediction result with causal interpretation is calculated and output by combining knowledge driving and data driving characteristics and utilizing bidirectional causal effect quantification, and the learning proportion of the two types of characteristics is dynamically adjusted according to local optimization model parameters of a market mutation scene, so that the prediction accuracy is improved. The method can adapt to conventional and extreme scenes, provides reliable support for collaborative decision-making of a supply chain, and solves the problem that the conventional prediction method is insufficient in adaptability and interpretability.

Inventors

WANG XUESONG
LIU XIAOMEI

Assignees

北京盈齐科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260226

Claims (9)

1. The supply chain dynamic demand prediction method based on multi-source data fusion and deep learning is characterized by comprising the following steps of: Establishing an association mapping between a supply chain physical entity and a virtual node through multi-source data acquisition and physical-virtual mapping, performing multi-physical field coupling modeling to realize collaborative mapping of logistics, information flow and fund flow, and constructing an extreme scene simulation engine to support scene analysis and simulation; Forming a data driving feature by using a deep learning network, symbolizing and encoding knowledge in the field of the supply chain to form a knowledge driving feature, fusing the data driving feature and the knowledge driving feature, and constructing a neural symbol AI prediction model based on a causal inference network to realize supply chain demand prediction containing causal interpretation; And thirdly, realizing a market mutation coping mechanism, namely identifying market mutation by defining a market feature vector, carrying out virtual pressure test, screening related data samples based on mutation types, carrying out local parameter optimization on a neural symbol AI prediction model, and outputting a scene-based demand prediction.
2. The method for predicting the dynamic demand of a supply chain based on multi-source data fusion and deep learning according to claim 1, wherein the multi-physical field coupling modeling is specifically as follows: Constructing a coupling model of material flow, information flow and fund flow by adopting a data mixed modeling method, and disassembling a mapping function into coupling calculation of three sub-modules, wherein the coupling calculation is specifically as follows: S121, modeling a logistics submodule, namely constructing an inventory circulation model based on system dynamics, wherein state variables comprise on-transit inventory, on-inventory storage and ex-inventory; S122, modeling an information flow submodule, namely constructing a node communication model by adopting a graph neural network, and taking supply chain nodes including suppliers, manufacturers and distributors as graph nodes and information transfer among the nodes as edge weights; S123, modeling a fund flow submodule, namely constructing a fund flow model based on a differential equation; S124, defining state variables of a physical flow field, an information flow field and a fund field, constructing a coupling rule base, optimizing a cross-field coupling coefficient, and realizing cooperative coupling of the physical flow, the information flow and the fund flow, wherein the cross-field coupling coefficient represents influence degree or driving weight between the fields.
3. The supply chain dynamic demand prediction method based on multi-source data fusion and deep learning according to claim 1, wherein the construction of the extreme scene simulation engine is specifically as follows: S131, constructing a scene template library, wherein the scene template library comprises a preset extreme scene template, specifically comprises natural disasters, geopolitics, demand bursts, supply chain node faults, market competition mutation and policy and regulation mutation; s132, randomly designing disturbance factors, namely generating the disturbance factors based on probability distribution of historical data; s133, performing scene simulation, namely realizing scene simulation through a simulation platform based on the generated disturbance factors.
4. The method for predicting the dynamic demand of a supply chain based on multi-source data fusion and deep learning according to claim 1, wherein the forming process of the data driving features is as follows: The method comprises the steps of obtaining supply chain related data output by a supply chain digital twin virtual mirror system in the first step, wherein the data at least cover stock data of a supply chain physical entity, historical demand data in a designated time period and real-time order data synchronized by an enterprise ERP system, inputting a preprocessed data set into a preset deep learning network, and obtaining data characteristics output by the network as the data driving characteristics.
5. The method for predicting the dynamic demand of a supply chain based on multi-source data fusion and deep learning according to claim 1, wherein the knowledge driving feature forming process is specifically as follows: Firstly, converting a supply chain core rule into a logic form which can be recognized by a machine by means of symbolized coding of knowledge in the field of a supply chain and adopting a preset logic programming language, and simultaneously, defining a typical association relation of the supply chain by combining a constructed supply chain knowledge graph; Secondly, constructing a symbol reasoning sub-module, definitely reasoning logic aiming at the typical association relation, and converting the symbol reasoning sub-module into a reasoning rule executable by the symbol reasoning sub-module; And finally, carrying out reasoning analysis on the real-time operation data of the supply chain and the history associated data through the symbol reasoning sub-module, and outputting a reasoning result in a binary vector form to form a knowledge driving characteristic.
6. The method for predicting the dynamic demand of a supply chain based on multi-source data fusion and deep learning according to claim 1, wherein the method for constructing the neural symbol AI prediction model based on a causal inference network is specifically as follows: Performing dimension expansion on the knowledge driving features through a feature fusion layer to enable the knowledge driving features to be matched with the data driving features in dimension, calculating the attention weights of the two types of features by using an attention mechanism, and weighting to generate fusion features; The method comprises the steps of taking fusion characteristics as node characteristics of supply chain nodes, constructing a causal inference network by adopting a causal graph neural network, calculating the primary causal association degree among nodes through mutual information, combining causal relation priori constraint provided by knowledge driving characteristics, improving the mutual information of known association nodes by setting penalty items, constraining the upper limit of the edge weight of the unknown association nodes, and quantifying the bidirectional causal effect among calculation nodes through Do-Calculus intervention operation to complete model construction.
7. The method for predicting the dynamic demand of a supply chain based on multi-source data fusion and deep learning according to claim 6, wherein the calculating the bidirectional causal effect between nodes is as follows: determining an upstream node and a downstream node in a supply chain, and defining a key state and a reference state of the node, wherein the key state refers to a special operation condition with obvious influence on demand prediction, and the reference state refers to a normal and stable operation condition; Calculating forward causal effect, namely adopting Do-Calculus intervention operation to respectively calculate the demand prediction result expectation of the downstream node when the upstream node is actively controlled to be in a key state and the downstream node is in a reference state, and the demand prediction result expectation of the downstream node when the upstream node is actively controlled to be in the reference state and the downstream node is in the reference state, and taking the difference value of the two expectations as the forward causal effect value of the upstream node to the downstream node; And calculating the opposite causal effect, namely adopting Do-Calculus intervention operation to calculate the demand prediction result expectation of the upstream node when the downstream node is actively controlled to be in a key state and the upstream node is in a reference state and the demand prediction result expectation of the upstream node when the downstream node is actively controlled to be in the reference state and the upstream node is in the reference state, and taking the difference value of the two expectations as the opposite causal effect value of the downstream node to the upstream node.
8. The method for predicting dynamic demand of a supply chain based on multi-source data fusion and deep learning according to claim 1, wherein the market feature vector comprises four dimensions of demand fluctuation rate, price fluctuation rate, order cancellation rate and policy change.
9. The supply chain dynamic demand prediction method based on multi-source data fusion and deep learning according to claim 1, wherein the local parameter optimization is performed on the neural symbol AI prediction model, specifically as follows: Judging the influence degree of mutation on a supply chain according to a quantization index output by a virtual pressure test, and adjusting the limit weight constraint upper limit; the punishment item optimization, namely combining the core indexes output by the virtual pressure test, and adapting punishment item values according to different risk grades; Attention mechanism weight duty ratio reinforcement, namely aiming at core features directly related to mutation, improving coefficients of trainable weight matrixes corresponding to the core features in an attention weight calculation formula of a feature fusion layer.

Description

Multi-source data fusion and deep learning-based supply chain dynamic demand prediction method Technical Field The invention relates to the technical field of intelligent management of supply chains, in particular to a supply chain dynamic demand prediction method based on multi-source data fusion and deep learning. Background Supply chain demand prediction is a core technology for optimizing inventory configuration and coordinating production and marketing engagement of enterprises, directly influences the operation efficiency and cost control of a supply chain, and has key application value in the fields of modern manufacturing industry, retail industry and the like. The existing supply chain demand prediction method is mostly dependent on a single data source or a traditional machine learning model, and has obvious limitations that on one hand, part of the method only focuses on data driving modeling, ignores integration of knowledge such as rules, node association and the like in the field of a supply chain, causes a prediction result to lack of causal logic support, is difficult to explain core driving factors behind the prediction and is unfavorable for accurate decision of enterprises, and on the other hand, the existing model is mostly in a static training mode, model parameters cannot be quickly adjusted to adapt to data distribution changes when facing extreme scenes such as natural disasters, market competition mutation and the like, and prediction accuracy is greatly reduced. Meanwhile, the traditional method is difficult to realize collaborative modeling of logistics, information flow and fund flow, and cannot comprehensively capture the influence of multi-dimensional dynamic association of a supply chain on requirements. Therefore, the invention provides a dynamic demand prediction method integrating multi-source data and deep learning, which solves the problems of poor interpretability and insufficient extreme scene adaptability in the prior art. Disclosure of Invention Aiming at the defects existing in the prior art, the invention aims to provide a supply chain dynamic demand prediction method based on multi-source data fusion and deep learning. In order to achieve the above purpose, the present invention provides the following technical solutions: The supply chain dynamic demand prediction method based on multi-source data fusion and deep learning specifically comprises the following steps: Establishing an association mapping between a supply chain physical entity and a virtual node through multi-source data acquisition and physical-virtual mapping, performing multi-physical field coupling modeling to realize collaborative mapping of logistics, information flow and fund flow, and constructing an extreme scene simulation engine to support scene analysis and simulation; Forming a data driving feature by using a deep learning network, symbolizing and encoding knowledge in the field of the supply chain to form a knowledge driving feature, fusing the data driving feature and the knowledge driving feature, and constructing a neural symbol AI prediction model based on a causal inference network to realize supply chain demand prediction containing causal interpretation; And thirdly, realizing a market mutation coping mechanism, namely identifying market mutation by defining a market feature vector, carrying out virtual pressure test, screening related data samples based on mutation types, carrying out local parameter optimization on a neural symbol AI prediction model, and outputting a scene-based demand prediction. Further, the multi-physical field coupling modeling is specifically as follows: Constructing a coupling model of material flow, information flow and fund flow by adopting a data mixed modeling method, and disassembling a mapping function into coupling calculation of three sub-modules, wherein the coupling calculation is specifically as follows: S121, modeling a logistics submodule, namely constructing an inventory circulation model based on system dynamics, wherein state variables comprise on-transit inventory, on-inventory storage and ex-inventory; S122, modeling an information flow submodule, namely constructing a node communication model by adopting a graph neural network, and taking supply chain nodes including suppliers, manufacturers and distributors as graph nodes and information transfer among the nodes as edge weights; S123, modeling a fund flow submodule, namely constructing a fund flow model based on a differential equation; S124, defining state variables of a physical flow field, an information flow field and a fund field, constructing a coupling rule base, optimizing a cross-field coupling coefficient, and realizing cooperative coupling of the physical flow, the information flow and the fund flow, wherein the cross-field coupling coefficient represents influence degree or driving weight between the fields. Furthermore, the construction of the extreme scene simulation engine