Supply chain dynamic demand prediction method based on multi-source data fusion and deep learning

Abstract

The invention relates to the technical field of supply chain intelligent management, in particular to a supply chain dynamic demand prediction method based on multi-source data fusion and deep learning, and the method comprises three steps: constructing a supply chain digital twinning virtual mirror image system, constructing a neural symbol AI and causal inference prediction model, and realizing a market mutation coping mechanism. Knowledge-driven and data-driven features are fused, and a prediction result with causal interpretation is output through bidirectional causal effect quantitative calculation; and for market sudden change scene local optimization model parameters, the learning proportion of the two types of features is dynamically adjusted, and the prediction accuracy is improved. The method can adapt to conventional and extreme scenes, provides reliable support for supply chain collaborative decision-making, and solves the problem that a traditional prediction method is insufficient in adaptability and interpretability.

Description

Technical Field

[0001] This invention relates to the field of intelligent supply chain management technology, and more specifically, to a method for dynamic demand forecasting of the supply chain based on multi-source data fusion and deep learning. Background Technology

[0002] Supply chain demand forecasting is a core technology for enterprises to optimize inventory allocation and coordinate production and sales. It directly affects the efficiency of supply chain operation and cost control, and has key application value in modern manufacturing, retail and other fields.

[0003] Existing supply chain demand forecasting methods largely rely on single data sources or traditional machine learning models, which have significant limitations. On the one hand, some methods focus solely on data-driven modeling, neglecting the integration of knowledge such as supply chain rules and node relationships. This results in forecasts lacking causal logic support, making it difficult to explain the core driving factors behind the forecasts and hindering accurate corporate decision-making. On the other hand, existing models are mostly statically trained, and when faced with extreme scenarios such as natural disasters or sudden market changes, they cannot quickly adjust model parameters to adapt to changes in data distribution, leading to a significant drop in forecast accuracy. Furthermore, traditional methods struggle to achieve collaborative modeling of logistics, information flow, and capital flow, failing to comprehensively capture the impact of multi-dimensional dynamic relationships within the supply chain on demand.

[0004] To address this, the present invention proposes a dynamic demand prediction method that integrates multi-source data and deep learning, thereby solving the problems of poor interpretability and insufficient adaptability to extreme scenarios in existing technologies. Summary of the Invention

[0005] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a method for dynamic supply chain demand forecasting based on multi-source data fusion and deep learning.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The supply chain dynamic demand forecasting method based on multi-source data fusion and deep learning includes the following steps: Step 1: Construct a digital twin virtual mirror system for the supply chain: Establish the association mapping between physical entities and virtual nodes in the supply chain through multi-source data collection and physical-virtual mapping, perform multi-physics coupling modeling to achieve collaborative mapping of logistics, information flow, and capital flow, and build an extreme scenario simulation engine to support scenario-based analysis and simulation; Step 2: Construction of a supply chain demand forecasting model based on neural symbolic AI and causal inference: Data-driven features are formed by using deep learning networks, and knowledge in the supply chain domain is symbolically encoded to form knowledge-driven features. The data-driven features and knowledge-driven features are integrated to construct a neural symbolic AI forecasting model based on causal inference networks, so as to realize supply chain demand forecasting with causal explanation. Step 3: Implementation of the market mutation response mechanism: Identify market mutations by defining market feature vectors and conduct virtual stress tests, screen related data samples based on mutation types, optimize local parameters of the neural symbol AI prediction model, and output scenario-based demand predictions.

[0007] Furthermore, the aforementioned multiphysics coupling modeling is specifically as follows: A hybrid data modeling approach is used to construct a coupled model of logistics, information flow, and capital flow. The mapping function is decomposed into three sub-modules for coupled calculation, as follows: S121. Logistics Sub-module Modeling: Construct an inventory turnover model based on system dynamics, with state variables including inventory in transit, inventory in stock, and outbound quantity; S122. Information flow sub-module modeling: A graph neural network is used to construct a node communication model, with supply chain nodes including suppliers, manufacturers, and distributors as graph nodes, and information transmission between nodes as edge weights. S123, Funds Flow Submodule Modeling: Constructing a funds flow model based on differential equations; S124. Multi-field coupling mechanism: Define the state variables of logistics field, information flow field and capital field, construct a coupling rule base, optimize the cross-field coupling coefficient, and realize the coordinated coupling of logistics, information flow and capital flow; the cross-field coupling coefficient represents the degree of influence or driving weight between fields.

[0008] Furthermore, the construction of the extreme scenario simulation engine is detailed as follows: S131. Construction of Scenario Template Library: Pre-set extreme scenario templates, specifically including natural disasters, geopolitics, demand surges, supply chain node failures, sudden changes in market competition, and sudden changes in policies and regulations; S132, Random perturbation factor design: generating perturbation factors based on the probability distribution of historical data; S133, Scenario Simulation Execution: Scenario simulation is achieved through a simulation platform based on the generated disturbance factors.

[0009] Furthermore, the formation process of the data-driven features is as follows: Obtain the supply chain-related data output by the supply chain digital twin virtual mirror system in step one. The data shall at least cover the inventory data of the physical entities in the supply chain, the historical demand data within a specified time period, and the real-time order data synchronized by the enterprise ERP system. Input the preprocessed dataset into a preset deep learning network, and the data features output by the network shall be the data-driven features.

[0010] Furthermore, the formation process of the aforementioned knowledge-driven features is as follows: First, relying on the symbolic encoding of knowledge in the supply chain field, a pre-set logic programming language is used to transform the core rules of the supply chain into a machine-recognizable logical form. At the same time, the typical relationships in the supply chain are clarified by combining the constructed supply chain knowledge graph. Secondly, a symbolic reasoning submodule is constructed to clarify the reasoning logic for the typical relationships and transform it into reasoning rules that can be executed by the symbolic reasoning submodule. Finally, the symbolic reasoning submodule performs reasoning analysis on the real-time operation data and historical correlation data of the supply chain, and outputs the reasoning results in the form of binary vectors to form knowledge-driven features.

[0011] Furthermore, the method for constructing the neural symbol AI prediction model based on causal inference networks is as follows: The knowledge-driven features are expanded in dimension by a feature fusion layer to match the dimensions of the data-driven features. Then, the attention weights of the two types of features are calculated using an attention mechanism, and the fused features are generated by weighting. Using fusion features as node features of supply chain nodes, a causal inference network is constructed using a causal graph neural network. The preliminary causal correlation between nodes is calculated through mutual information. Combined with the prior constraints of causal relationships provided by knowledge-driven features, the mutual information of known associated nodes is improved by setting penalty terms, and the upper limit of edge weights of nodes without known associations is constrained. Then, the bidirectional causal effect between nodes is quantified by Do-Calculus intervention, thus completing the model construction.

[0012] Furthermore, the aforementioned bidirectional causal effect between computing nodes is as follows: Node and State Definitions: Identify upstream and downstream nodes in the supply chain, and define the critical and baseline states of nodes. The critical state refers to a special operating condition in which a node has a significant impact on demand forecasting, while the baseline state refers to a node operating in a normal and stable state. Positive causal effect calculation: Using Do-Calculus intervention, the expected demand forecast of the downstream node is calculated when the upstream node is in a critical state and the downstream node is in a baseline state, and the expected demand forecast of the downstream node is calculated when the upstream node is in a baseline state and the downstream node is in a baseline state. The difference between the two expectations is taken as the positive causal effect value of the upstream node on the downstream node. Reverse causal effect calculation: Using the same Do-Calculus intervention operation, the expected demand forecast of the upstream node is calculated when the downstream node is in a critical state and the upstream node is in a baseline state, and the expected demand forecast of the upstream node is calculated when the downstream node is in a baseline state and the upstream node is in a baseline state. The difference between the two expectations is taken as the reverse causal effect value of the downstream node on the upstream node.

[0013] Furthermore, the aforementioned market characteristic vector specifically includes four dimensions: demand volatility, price volatility, order cancellation rate, and policy changes.

[0014] Furthermore, the local parameter optimization of the neural symbol AI prediction model is specifically as follows: Adjusting the upper limit of edge weight constraints in causal inference networks: Based on the quantitative indicators output by virtual stress tests, the impact of sudden changes on the supply chain is judged, and the upper limit of edge weight constraints is adjusted accordingly. Penalty optimization: Based on the core metrics output by virtual stress testing, the penalty values ​​are adapted to different risk levels; Attention mechanism weight enhancement: For core features directly related to mutations, the coefficients of the trainable weight matrix corresponding to such core features are increased in the attention weight calculation formula of the feature fusion layer.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This method employs neural symbolic AI and causal inference networks, integrating knowledge-driven and data-driven features. Through Do-Calculus intervention, it quantifies the bidirectional causal effects between nodes, achieving demand forecast outputs with causal explanations. This approach overcomes the black-box limitations of traditional data-driven models, making the logical connections of the forecast results clear and traceable, providing explicit guidance for collaborative decision-making across the supply chain. 2. Construct a market mutation identification and virtual stress testing mechanism. This mechanism can quickly identify market mutations based on feature vectors such as demand volatility and policy changes, and selectively filter relevant data samples to optimize model edge weight constraints, penalty terms, and attention weight ratios. By dynamically adjusting the learning ratio of knowledge and data, the prediction accuracy under mutation scenarios is significantly improved, enhancing the supply chain's responsiveness to extreme situations. Attached Figure Description

[0016] Figure 1 A flowchart of a supply chain dynamic demand forecasting method based on multi-source data fusion and deep learning; Figure 2 A flowchart illustrating the construction of the causal inference model for neural symbol fusion in this invention; Figure 3 This is a flowchart illustrating the implementation of the market mutation response mechanism of the present invention. Detailed Implementation

[0017] Example, refer to Figure 1 The supply chain dynamic demand forecasting method based on multi-source data fusion and deep learning in this embodiment specifically includes the following steps: Step 1: Construct a digital twin virtual mirror system for the supply chain.

[0018] To achieve real-time mapping of supply chain status and simulation of extreme scenarios, and as the foundation for dynamic demand forecasting of the supply chain, a high-fidelity digital twin virtual mirror system for the supply chain is first constructed. The specific steps are as follows: S11. Multi-source data acquisition and physical-virtual mapping: Use IoT sensors to collect full-link data of physical entities in the supply chain, including but not limited to inventory data, logistics trajectory data, and production equipment status (such as operating parameters, fault codes, energy consumption data, etc.); connect to the enterprise ERP system through API interface to extract production progress and real-time order data, connect to the enterprise data warehouse to extract historical demand data for the past 12 months, and connect to third-party data platforms to obtain external data such as market trends and policies and regulations; Based on this physical data, an entity-attribute-relationship modeling framework is adopted to map physical entities (such as warehouses, transport vehicles, and suppliers) as virtual nodes. The attributes of each node (such as warehouse capacity, vehicle load capacity, and supplier capacity) are bound to the physical data in real time. The real-time transmission of physical data to the virtual system is realized through the MQTT protocol. The output real-time multi-source data is synchronized to the input layer of the neural symbol AI prediction model in step two through the twin system data interface, providing the basic data for model training and inference; at the same time, the supply chain node relationship generated by the physical-virtual mapping is stored in the knowledge graph database. S12. High-fidelity multiphysics modeling: A data hybrid modeling approach is used to construct a coupled model of logistics, information flow, and capital flow, with mapping functions... The computation is broken down into three coupled sub-modules, as follows: S121. Logistics Submodule Modeling: An inventory turnover model is constructed based on system dynamics. State variables include inventory in transit, inventory in stock, and outbound quantity. The core equation is: ; in, The rate of change of in-transit inventory reflects the dynamic turnover of in-transit inventory over time. This refers to the total amount of goods that are in transit and have not yet arrived at the target warehouse; The supplier replenishment quantity at time t is obtained in real time from the ERP system. The amount of inventory arriving in transit at time t is calculated from the GPS trajectory. An arrival event is triggered when the distance between the vehicle location and the warehouse is ≤500m. This represents the rate of change in inventory, reflecting the dynamic changes in warehouse inventory as replenishment and shipments occur. This refers to the total amount of goods that have arrived at the warehouse and are in storage. The market demand at time t is determined by the prediction results output by the neural symbolic AI prediction model in step two of this invention. S122. Information Flow Submodule Modeling: A graph neural network is used to construct a node communication model, treating supply chain nodes, including suppliers, manufacturers, and distributors, as graph nodes. Information transmission between nodes (such as order instructions and inventory alerts) is used as edge weights, and information transmission delay is considered. It is calculated from the physical distance between nodes, specifically by dividing the transmission distance by the empirical transmission speed; S123. Modeling of the Funds Flow Submodule: A funds flow model is constructed based on differential equations. The core equation is: ; in, This indicates the rate of change in the account's cash balance. It refers to the company's account balance, reflecting the total amount of disposable funds of the company at a certain point in time; The payment amount to downstream enterprises at time t is calculated by multiplying the order amount by the payment cycle. For monthly settlement orders, the payment amount is divided by 30 days and evenly distributed. The upstream procurement cost at time t is calculated by multiplying the procurement quantity by the unit price. The unit price is calculated by the weighted average of historical transaction prices and real-time quotations. S124. Multi-field coupling mechanism: Achieving coordinated coupling of logistics, information flow, and capital flow through a standardized framework of state definition, rule modeling, and coefficient optimization, as detailed below: Cross-field state variable definition: Logistics facilities: inventory levels, volume of goods in transit, and load on logistics nodes, etc. Information flow field: instruction type (such as replenishment, transfer), information transmission delay, node response status, etc.; Capital flow: account balance, income and expenditure, capital turnover rate, etc.

[0019] Construction of the coupling rule base: The rule form is uniformly described as follows: If a certain state variable of field A is within a preset threshold range, then field B is triggered to execute the specified action, and field C is triggered to execute the specified action at the same time. The rules in the rule base cover all scenarios of inventory management, order processing, and financial settlement. For example, if the inventory in the logistics field is lower than the safety stock threshold, the information flow field will be triggered to send a replenishment instruction to the upstream supplier, and the capital field will be triggered to calculate the funds required for procurement and reserve the budget. Coupling coefficient optimization: Gradient descent method is used to optimize the cross-field coupling coefficient. Using the root mean square error (RMSE) between the virtual model and the physical reality as the objective function, iteratively adjust to... The cross-field coupling coefficient represents the influence weight of field i on field j. For example, if i = logistics field and j = information flow field, then... This represents the degree to which changes in the state of the logistics field (such as insufficient inventory) affect the triggering of actions in the information flow field (such as sending replenishment instructions), where i = information flow field and j = capital field. The driving weight of instructions in the information flow field (such as replenishment instructions) on operations in the cash flow field (such as reserving procurement funds); The supply chain node state data output through multiphysics modeling is directly used as the node feature input for the causal graph neural network in step two, supporting the calculation of causal effects between nodes; simultaneously, the mapping function... The modeling logic also provides a scenario calculation basis for the virtual stress test of the fourth step market change response mechanism - when a market change is detected, the fourth step can call the multiphysics model of this step to simulate the chain impact of the change on logistics, information flow and capital flow. S13, Extreme Scene Simulation Engine Setup: S131. Scene Template Library Construction: Six categories of extreme scene templates are preset, each containing 3-5 parameterized scenes, as detailed below: Natural disasters: such as typhoons causing logistics disruptions, earthquakes causing production base shutdowns, and floods causing damage to warehousing facilities; Geopolitical factors include tariff increases, trade sanctions leading to raw material supply disruptions, and significant exchange rate fluctuations. Explosive demand: such as a surge in orders during major sales events, a surge in popularity of trending products, and seasonal demand peaks; Supply chain node failures: such as core supplier shutdowns, logistics hub disruptions, and critical equipment failures; Sudden changes in market competition: such as competitors drastically reducing prices, competitors launching disruptive new products, and distributors switching partners on a large scale; Sudden changes in policies and regulations: such as upgrades to environmental regulations, increases in food safety standards, and tightening of data privacy regulations; S132. Random Disturbance Factor Design: The Monte Carlo method is used to generate disturbance factors for 3-5 core indicators in the calculation scenario template. For example, in the logistics disruption scenario, random disturbances affect vehicle traffic efficiency (disturbance range 0-50%); in the demand surge scenario, random disturbances affect the demand growth rate (disturbance range 100%-300%). The probability distribution of the disturbance factors is fitted based on historical extreme event data (such as logistics disruption data caused by typhoons in the past 5 years). S133, Scene Simulation Execution: Scene simulation is implemented based on the AnyLogic simulation platform, supporting single-scene simulation and multi-scene superposition simulation. The simulation step size is set to 1 hour, and the time taken for a single scene simulation is ≤10 minutes, meeting the real-time decision-making requirements. The constructed extreme scenario template library and simulation engine, on the one hand, generate mutation trigger signals by monitoring market data characteristics in real time, and push them to the fourth step of the market mutation response mechanism to start the adjustment process; on the other hand, the results of the scenario simulation output, such as inventory depletion time and logistics congestion nodes, provide scenario-based datasets for the optimization of local parameters of the model in the fourth step, ensuring the pertinence of the mutation adjustment.

[0020] Step 2: Construction of a supply chain demand forecasting model based on neural symbolic AI and causal inference.

[0021] S21. Symbolic encoding of knowledge in the supply chain domain: S211. Symbolization Tools and Rule Transformation: Knowledge symbolization is achieved using the Prolog programming language. For core supply chain rules, logical relationships are clearly defined using natural language, and then transformed into a machine-recognizable logical form using symbolization tools. For example: Rule 1 (Replenishment Trigger Rule): If there is a warehouse X1 with an actual inventory value of Y1 and a safety stock value of Z1, and the actual inventory value Y1 is less than the safety stock value Z1, then a replenishment operation for warehouse X1 is triggered. Rule 2 (Order Priority Rule): If there is an order Q1 (corresponding to fresh produce) and an order Q2 (corresponding to general daily necessities), and the delivery deadline of order Q1 is less than 24 hours from the current time, while the delivery deadline of order Q2 is more than 72 hours from the current time, then the processing priority of order Q1 is higher than that of order Q2. The rule base contains 20-30 core rules developed by a group of experts in the field based on their knowledge, covering scenarios such as inventory management, order processing, and logistics scheduling. Each rule clearly defines the conditions and the resulting actions. S212 Knowledge Graph Construction and Update: The knowledge graph is constructed using the Neo4j graph database. Node types include enterprises, warehouses, and products, and edge types include supply, storage, and sales. Dynamic updates of the knowledge graph are achieved through scheduled tasks. During updates, an incremental merging strategy is adopted—only new enterprise partnerships and product categories are added, or the attributes of existing nodes or edges are modified, to avoid overwriting historical related data. The constructed symbolic rule base and knowledge graph serve as the core knowledge input for the neural symbolic AI prediction model. The symbolic reasoning submodule performs reasoning by calling the logical relationships in the rule base, while the knowledge graph provides the supply chain node association logic for the causal inference network, ensuring the knowledge-driven rationality of subsequent model training. S22. Construction of a causal inference model based on neural symbol fusion (referred to as a neural symbol AI prediction model): For example... Figure 2 As shown, the steps are as follows: S221. Deep Learning Sub-Network Parameters: A Transformer network is used to extract data features. The network has 6 layers, 8 multi-head attention heads, and 512 hidden layer dimensions. Input data includes inventory data, historical demand data from the past 12 months, and real-time order data. Data preprocessing uses Z-score normalization. All input data originates from the digital twin system data interface in Step 1. Output data consists of data-driven features extracted by the deep learning sub-network. ; S222, Symbolic Reasoning Submodule Implementation: Logical reasoning is implemented using Answer Set Programming (ASP). For typical supply chain relationships (originating from), the reasoning logic is explicitly stated in natural language and then transformed into ASP-executable reasoning rules. For example: Inference Rule 1 (Insufficient Inventory Attribution): If there is an upstream company L1 that has a delivery delay (e.g., the actual delivery time is more than 48 hours later than the agreed delivery time), and the upstream company L1 and the downstream company L2 have a stable supply relationship (e.g., there are at least 3 supply records per month in the past 12 months), then it can be inferred that the insufficient inventory problem of the downstream company L2 is related to the delivery delay of the upstream company L1. Inference Rule 2 (Demand Fluctuation Correlation): If there exists a regional market A1, and the monthly demand for a certain type of product in it increases by more than 50% month-on-month, and a large-scale promotional event (such as a regional shopping festival) is held in regional market A1 during the same period, then it can be inferred that the demand fluctuation of this type of product is related to the promotional event.

[0022] The reasoning results are output in binary vector form, denoted as knowledge-driven features. The vector dimension is consistent with the number of core relationship types in the supply chain. In the vector, "1" indicates that the corresponding relationship is valid, and "0" indicates that it is not valid. Its core function is to transform abstract domain knowledge into numerical features that the model can recognize, providing knowledge-level input for subsequent feature fusion. The core relationship types in the supply chain are derived from the knowledge graph described in step S212. S223, Feature Fusion Layer: This layer integrates the knowledge-driven features output from the symbolic reasoning submodule. Data-driven features extracted by deep learning sub-networks Fusion features are generated through attention mechanisms. The specific steps are as follows: right To expand the dimensions, a fully connected layer is used. Mapped to 512 dimensions, and With consistent dimensions, we obtain expanded knowledge features. ; Calculate attention weights : ,in, This is a trainable weight matrix with dimensions 512×512, where b is the bias term and the weights are... Used to measure the importance of two types of features; Generate fusion features: ; The core function of this fusion process is to enable the model to utilize both the trend patterns at the data level and the logical connections at the knowledge level, thereby avoiding the problem of false associations that occur when relying solely on data. S224. Causal Inference Network Construction: A causal inference network is constructed using a causal graph neural network (Causal GNN), with supply chain nodes as graph nodes, and node features being the aforementioned fused features. The edge weights represent the initial causal correlation degree, calculated from mutual information: ,in For nodes State x and node The joint probability of state y, , They are nodes , Marginal probability of a state; Provide a priori constraints on causality: if If a known association exists between a pair of nodes, then when calculating the causal effect of that pair of nodes, a penalty term is used. Incorporating the total loss function of the causal inference network, i.e., total loss = causal prediction loss + Reg, improves the mutual information between nodes. This ensures that the network prioritizes learning causal effects between nodes with clear connections; if If there is no known association between nodes, the upper limit of edge weights (set to the default of 0.3 in this embodiment) is constrained to prevent the network from learning false causality. based on The provided causal prior constraints, through Do-Calculus intervention, enable the quantitative calculation of bidirectional causal effects between supply chain nodes, as detailed below: Calculation of positive causal effect: Calculation of upstream nodes For downstream nodes Demand-driven effect The formula is as follows: ; Reverse causality calculation: Calculating downstream nodes For upstream nodes Demand-driven effect The formula is as follows: ; in, The mathematical expectation symbol represents the average of the values ​​of a specified variable. For causal interference econometrics, it means actively controlling a variable to be in a specified state; Indicates an upstream node in the supply chain; Its critical state, Its baseline state; Indicates a downstream node in the supply chain; Its critical state, Its baseline state; and Each is an upstream node in the supply chain and downstream nodes Demand forecast results; The critical state refers to a special operational situation in the supply chain that significantly impacts demand forecasting. This is typically an extreme or core business scenario that triggers significant changes (e.g., full capacity operation, order surge, large-scale promotional activities). This state is simulated through proactive intervention to quantify the driving or feedback effect of the node on upstream and downstream demand under special scenarios. The baseline state refers to a normal, stable operational situation in the supply chain, representing a normal, undisturbed business state (e.g., normal capacity operation, stable order volume, no special marketing activities). This state serves as a benchmark, used to calculate the difference between the demand forecast results of the critical state and the actual situation, thereby quantifying the differences in causal effects of the node under different states. For example: This refers to factories that supply raw materials. This refers to the manufacturer that uses the raw materials. To ensure the factory operates at 100% full capacity, To ensure stable order volumes for manufacturers; Indicates active control Full load When stable, Raw material demand forecast The average value (the average value after multiple trials).

[0023] The trained neural symbolic AI prediction model ultimately outputs a supply chain demand forecast package with causal explanations, which includes two parts: Core forecast values: Demand forecasts for each node in the supply chain; Causal Explanation Report: Based on the results of bidirectional causal effect analysis, it outputs the logical relationship behind the predicted values ​​to guide the collaborative decision-making direction of upstream and downstream enterprises.

[0024] Step 3: Implementation of market change response mechanisms.

[0025] like Figure 3 As shown, the specific steps are as follows: S31. Mutation Identification and Virtual Stress Testing: S311, Definition of Market Feature Vector: ,in: Demand volatility ; t and t-1 are adjacent hours, This is the real-time demand forecast output from step two; Price volatility ; The price data at time t is derived from the external data interface of the digital twin system in step one. Order cancellation rate ; Let t be the number of orders cancelled within hour t. This represents the total number of orders within hour t; order data is synchronized from the enterprise ERP interface in step one. Policy Change This is a binary variable; 1 indicates that there are policy changes that affect the supply chain on that day, and 0 indicates that there are no policy changes. The policy data comes from the third-party data platform in step one. Normal feature distribution determination: Based on non-mutation data from the past 3 months (excluding periods of major sales events, natural disasters, etc., filtered by the scene labels of the extreme scenario simulation engine in step one—only data with normal scene labels are retained), the mean of each feature is calculated. with standard deviation , Mahalanobis distance is used to measure the distance between the distribution of market feature vectors and the normal feature distribution. A distance threshold of 3 is set, corresponding to a 99.7% confidence interval, to ensure that the false alarm rate of mutation identification is ≤0.3%. S312, Virtual stress test execution: After the mutation is triggered, the knowledge-driven features in step two are first called to make a preliminary determination of the mutation type. Then, according to the mutation type, the corresponding extreme scenario template is loaded in the digital twin system. The simulation duration is set to the next 72 hours. The output simulation results include inventory depletion time, logistics congestion nodes, and funding gap amount, providing scenario support for subsequent model adjustments. S32. Optimization of local parameters of the model: S321. Screening Samples Strongly Correlated with Mutation Types: Based on the mutation types initially determined through knowledge-driven features in S312, identify the core supply chain indicators associated with the mutation, such as inventory turnover, order fluctuations, logistics efficiency, and cash inflows and outflows, all derived from the multiphysics modeling output of the digital twin system; combined with the normal feature distribution threshold determined in S311, screen samples where the above core indicators deviate from the normal distribution mean by 2. Real-time data samples are used to ensure that the correlation between the samples and the mutation indicators is matched, and that the deviation of the indicators is consistent with the scope of the mutation's impact. S322, Local parameter adjustment: Adjustment of the upper limit of edge weight constraints in the causal inference network: Based on quantitative indicators such as inventory depletion time, logistics node load, and funding gap amount output by virtual stress tests, the upper limit of edge weight constraints is adjusted. If a sudden change results in inventory depletion time ≤ 48 hours, logistics node load ≥ 90%, or funding gap ≥ 20%, it indicates a significant impact, and the upper limit of edge weight constraints is appropriately increased to strengthen the causal relationship calculation between nodes directly related to the sudden change; if the above indicators deviate from the normal range by less than 10%, it indicates a minor impact, and the upper limit is maintained or slightly adjusted to avoid over-amplifying non-core relationships. The overall adjustment range is dynamically adapted based on the basic upper limit under normal scenarios. Penalty items Optimization: Combining the three core metrics output from the virtual stress test—inventory depletion time (Tr), logistics node load (Lr), and funding gap ratio (Rr)—the following ranges are mapped. value: High risk (meeting any of the following: Tr≤48 hours, Lr≥90%, Rr≥20%): Setting it to 0.3 forces the model to prioritize domain knowledge, ensuring the determinism of causal inferences in mutation scenarios; Medium risk (meeting any of the following: 48 hours < Tr ≤ 72 hours, 70% ≤ Lr < 90%, 10% ≤ Rr < 20%): We set it to 0.2 to balance domain knowledge with data dynamism; Low risk (Tr > 72 hours, Lr < 70%, Rr < 10%): The base value of 0.1 is set to allow the model to flexibly learn new data-driven causal patterns within the knowledge framework; Attention mechanism enhances weight ratio: For core features directly related to mutation (including core rules corresponding to mutation scenarios in knowledge-driven features and key indicators affected by mutation in data-driven features), in the attention weight calculation formula of S223, the weight coefficients corresponding to these core features in the trainable weight matrix are increased to ensure that the model prioritizes focusing on the key influencing factors of mutation when feature fusion. A neural symbolic AI prediction model based on local parameter optimization outputs scenario-based demand predictions that align with current market changes.

[0026] Through the detailed description of the above embodiments, the supply chain dynamic demand forecasting method of the present invention, based on multi-source data fusion and deep learning, constructs a digital twin virtual mirror system of the supply chain that integrates logistics, information flow, and capital flow. This system enables real-time mapping between physical entities and virtual nodes, as well as simulation of extreme scenarios, providing comprehensive data support and a foundation for scenario calculation for forecasting. Relying on neural symbolic AI technology, it combines the symbolic encoding of domain knowledge with deep learning data feature extraction. By quantifying bidirectional causal effects through causal inference networks, it provides a clear logical explanation for the forecast results. Simultaneously, it establishes a market mutation response mechanism, using feature vectors to identify mutations, virtual stress tests to deduce impacts, and local parameter optimization to adapt to scenarios, ensuring that the forecasting model maintains high accuracy under both normal and extreme conditions. This method comprehensively improves the interpretability, dynamic adaptability, and accuracy of supply chain demand forecasting, providing an efficient technical solution for the refined management of enterprise supply chains.

[0027] The above formulas are all dimensionless calculations, and the preset parameters in the formulas should be set by those skilled in the art according to the actual situation.

[0028] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0029] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0030] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0031] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0032] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0033] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0034] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A supply chain dynamic demand forecasting method based on multi-source data fusion and deep learning, characterized in that, The method flow is as follows: Step 1: Construct a digital twin virtual mirror system for the supply chain: Establish the association mapping between physical entities and virtual nodes in the supply chain through multi-source data collection and physical-virtual mapping, perform multi-physics coupling modeling to achieve collaborative mapping of logistics, information flow, and capital flow, and build an extreme scenario simulation engine to support scenario-based analysis and simulation; Step 2: Construction of a supply chain demand forecasting model based on neural symbolic AI and causal inference: Data-driven features are formed by using deep learning networks, and knowledge in the supply chain domain is symbolically encoded to form knowledge-driven features. The data-driven features and knowledge-driven features are integrated to construct a neural symbolic AI forecasting model based on causal inference networks, so as to realize supply chain demand forecasting with causal explanation. Step 3: Implementation of the market mutation response mechanism: Identify market mutations by defining market feature vectors and conduct virtual stress tests, screen related data samples based on mutation types, optimize local parameters of the neural symbol AI prediction model, and output scenario-based demand predictions.

2. The supply chain dynamic demand forecasting method based on multi-source data fusion and deep learning according to claim 1, characterized in that, The multiphysics coupling modeling described above is as follows: A hybrid data modeling approach is used to construct a coupled model of logistics, information flow, and capital flow. The mapping function is decomposed into three sub-modules for coupled calculation, as follows: S121. Logistics Sub-module Modeling: Construct an inventory turnover model based on system dynamics, with state variables including inventory in transit, inventory in stock, and outbound quantity; S122. Information flow sub-module modeling: A graph neural network is used to construct a node communication model, with supply chain nodes including suppliers, manufacturers, and distributors as graph nodes, and information transmission between nodes as edge weights. S123, Funds Flow Submodule Modeling: Constructing a funds flow model based on differential equations; S124. Multi-field coupling mechanism: Define the state variables of logistics field, information flow field and capital field, construct a coupling rule base, optimize the cross-field coupling coefficient, and realize the coordinated coupling of logistics, information flow and capital flow; the cross-field coupling coefficient represents the degree of influence or driving weight between fields.

3. The supply chain dynamic demand forecasting method based on multi-source data fusion and deep learning according to claim 1, characterized in that, The construction of the extreme scenario simulation engine is described in detail below: S131. Construction of Scenario Template Library: Pre-set extreme scenario templates, specifically including natural disasters, geopolitics, demand surges, supply chain node failures, sudden changes in market competition, and sudden changes in policies and regulations; S132, Random perturbation factor design: generating perturbation factors based on the probability distribution of historical data; S133, Scenario Simulation Execution: Scenario simulation is achieved through a simulation platform based on the generated disturbance factors.

4. The supply chain dynamic demand forecasting method based on multi-source data fusion and deep learning according to claim 1, characterized in that, The formation process of the aforementioned data-driven features is as follows: Obtain the supply chain-related data output by the supply chain digital twin virtual mirror system in step one. The data shall at least cover the inventory data of the physical entities in the supply chain, historical demand data within a specified time period, and real-time order data synchronized by the enterprise's ERP system. The preprocessed dataset is input into a preset deep learning network, and the data features output by the network are the data-driven features.

5. The supply chain dynamic demand forecasting method based on multi-source data fusion and deep learning according to claim 1, characterized in that, The formation process of the knowledge-driven features is as follows: First, relying on the symbolic encoding of knowledge in the supply chain field, a pre-set logic programming language is used to transform the core rules of the supply chain into a machine-recognizable logical form. At the same time, the typical relationships in the supply chain are clarified by combining the constructed supply chain knowledge graph. Secondly, a symbolic reasoning submodule is constructed to clarify the reasoning logic for the typical relationships and transform it into reasoning rules that can be executed by the symbolic reasoning submodule. Finally, the symbolic reasoning submodule performs reasoning analysis on the real-time operation data and historical correlation data of the supply chain, and outputs the reasoning results in the form of binary vectors to form knowledge-driven features.

6. The supply chain dynamic demand forecasting method based on multi-source data fusion and deep learning according to claim 1, characterized in that, The method for constructing the neural symbol AI prediction model based on causal inference networks is as follows: The knowledge-driven features are expanded in dimension by a feature fusion layer to match the dimensions of the data-driven features. Then, the attention weights of the two types of features are calculated using an attention mechanism, and the fused features are generated by weighting. Using fusion features as node features of supply chain nodes, a causal inference network is constructed using a causal graph neural network. The preliminary causal correlation between nodes is calculated through mutual information. Combined with the prior constraints of causal relationships provided by knowledge-driven features, the mutual information of known associated nodes is improved by setting penalty terms, and the upper limit of edge weights of nodes without known associations is constrained. Then, the bidirectional causal effect between nodes is quantified by Do-Calculus intervention, thus completing the model construction.

7. The supply chain dynamic demand forecasting method based on multi-source data fusion and deep learning according to claim 6, characterized in that, The aforementioned bidirectional causal effect between computing nodes is as follows: Node and State Definitions: Identify upstream and downstream nodes in the supply chain, and define the critical and baseline states of nodes. The critical state refers to a special operating condition in which a node has a significant impact on demand forecasting, while the baseline state refers to a node operating in a normal and stable state. Positive causal effect calculation: Using Do-Calculus intervention, the expected demand forecast of the downstream node is calculated when the upstream node is in a critical state and the downstream node is in a baseline state, and the expected demand forecast of the downstream node is calculated when the upstream node is in a baseline state and the downstream node is in a baseline state. The difference between the two expectations is taken as the positive causal effect value of the upstream node on the downstream node. Reverse causal effect calculation: Using the same Do-Calculus intervention operation, the expected demand forecast of the upstream node is calculated when the downstream node is in a critical state and the upstream node is in a baseline state, and the expected demand forecast of the upstream node is calculated when the downstream node is in a baseline state and the upstream node is in a baseline state. The difference between the two expectations is taken as the reverse causal effect value of the downstream node on the upstream node.

8. The supply chain dynamic demand forecasting method based on multi-source data fusion and deep learning according to claim 1, characterized in that, The market characteristic vector specifically includes four dimensions: demand volatility, price volatility, order cancellation rate, and policy changes.

9. The supply chain dynamic demand forecasting method based on multi-source data fusion and deep learning according to claim 1, characterized in that, The local parameter optimization of the neural symbol AI prediction model is described in detail below: Adjusting the upper limit of edge weight constraints in causal inference networks: Based on the quantitative indicators output by virtual stress tests, the impact of sudden changes on the supply chain is judged, and the upper limit of edge weight constraints is adjusted accordingly. Penalty optimization: Based on the core metrics output by virtual stress testing, the penalty values ​​are adapted to different risk levels; Attention mechanism weight enhancement: For core features directly related to mutations, the coefficients of the trainable weight matrix corresponding to such core features are increased in the attention weight calculation formula of the feature fusion layer.

Images