A Supply Chain Risk Propagation Prediction and Resilience Optimization Decision-Making Method

Abstract

This invention discloses a supply chain risk propagation prediction and resilience optimization decision-making method, belonging to the field of power material supply chain and risk management technology. The method first constructs an edge graph neural network containing edge node features based on a heterogeneous graph of a power grid project, and identifies weak links in the supply chain through iterative refinement. Second, it constructs a risk propagation dynamics model by combining a power grid project scenario tree and a SIR model to quantify the sequential impact of risk events on subsequent nodes. Simultaneously, it uses a time-series Transformer model to fuse multi-source heterogeneous data to predict risk scores and potential impact ranges. Finally, it achieves forward-looking resource allocation in extreme scenarios through a bibliometric optimization model combined with risk confidence interval estimation. This invention enables quantitative prediction of the entire process of risk from latency to propagation, improving the emergency response capability and resource allocation efficiency of the power supply chain in extreme scenarios.

Description

Technical Field

[0001] This invention relates to the field of power material supply chain management technology, and more specifically to a supply chain risk propagation prediction and resilience optimization decision-making method based on edge graph neural network, SIR model, time series Transformer, and distributed bar optimization. Background Technology

[0002] The power supply chain is the material foundation for ensuring the smooth construction and reliable operation of power grid projects. Compared with conventional industrial supply chains, the power supply chain is highly complex and unique: it encompasses the entire chain, from the research and development and manufacturing of large core equipment such as transformers and combined electrical appliances, to cross-regional specialized logistics transportation, and finally to on-site installation and commissioning of substations. These links are not only highly dependent on technology, but also have long production cycles and high material value. Once a risk occurs at a critical node (such as a core component supplier), it often generates a strong cascading effect through the supply chain network, leading to project delays, uncontrolled costs, and even damage to power grid stability.

[0003] Currently, the power industry still faces significant challenges in managing supply chain risks:

[0004] 1. Insufficient granularity in risk perception: Traditional methods often focus on the qualification review of individual suppliers, lacking the ability to uncover vulnerabilities that have far-reaching consequences from the perspective of complex network topology. Existing graph convolutional networks often ignore the dynamic evolution of edge features (such as logistics paths and contractual constraints) when dealing with such problems. Furthermore, the quantitative accuracy of existing risk propagation processes is not high, making it difficult to accurately simulate the sequential impact intensity of risks from latent to eruption.

[0005] 2. Data sparsity severely limits the accuracy of prediction models. In power grid engineering, extreme risk events (such as black swan events) are typical sparse samples. Conventional deep learning models often exhibit ineffective early warning when real crises occur due to the lack of negative sample training.

[0006] 3. Insufficient robustness of decision-making is the biggest bottleneck of current technology. Due to the complexity of the external environment (such as policy changes and trade frictions), the precise probability distribution of risk events is often unknown. Traditional stochastic optimization decision-making models rely heavily on accurate probability assumptions. Once the actual distribution shifts slightly, the resource allocation strategies they formulate may have huge deviations, leading to the failure of emergency resources under extreme conditions or excessive redundancy and waste of funds.

[0007] Therefore, how to solve the above-mentioned technical problems, or at least partially solve them, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] In view of the above problems, this invention proposes a supply chain risk propagation prediction and resilience optimization decision-making method, aiming to build a technical solution that can deeply integrate the heterogeneous characteristics of the supply chain, accurately predict the risk diffusion trajectory, and provide robust decision support in an environment of distribution uncertainty.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A supply chain risk propagation prediction and resilience optimization decision-making method, comprising the following steps: Construct a power supply chain relationship graph, and introduce a technology dependence operator into the node features and a supply elasticity coefficient into the edge features; The power supply chain relationship diagram is iteratively updated, and weak nodes are identified based on breakpoint scoring after convergence. For the aforementioned vulnerable nodes, the evolution of the supply chain state under external shocks is simulated based on the scenario tree and SIR model, and the propagation evolution curve of the proportion of affected nodes over time is obtained. Based on the propagation evolution curve, the expected distribution of the risk loss function is predicted by fusing multimodal time series data; Based on the expected distribution of the risk loss function, a moment-constrained uncertainty set is constructed, and an optimal configuration scheme is generated from the uncertainty set.

[0010] Preferably, in the power supply chain diagram, nodes include supplier nodes, logistics nodes, and / or engineering nodes; edges include physical supply edges, contractual association edges, and / or geographical proximity edges.

[0011] Preferably, the technology dependency operator is calculated based on the number of qualified suppliers and the hierarchy depth of the materials in the bill of materials structure tree; the calculation formula is:

[0012] In the formula, Indicates the number of qualified suppliers of materials. This indicates the level depth of resources within the Bill of Materials (BOM). Indicates the maximum level depth. and This indicates the preset weighting coefficient.

[0013] The supply elasticity coefficient is determined based on the logarithmic linear regression slope of historical raw material prices and logistics costs. This includes constructing a logarithmic linear regression model based on time series data of raw material futures price indices and logistics unit prices over the past few years, and extracting the regression slope as a side feature to quantify the sensitivity of the supply path to price fluctuations.

[0014] Preferably, the power supply chain relationship graph is iteratively updated, including calculating the edge-guided attention coefficient, and using the edge-guided attention coefficient to aggregate neighborhood information to update node features and edge features; the formula for calculating the edge-guided attention coefficient is:

[0015] In the formula, Softmax represents the normalization function, LeakyReLU represents the nonlinear activation function, and W represents the feature embedding weight matrix. and These are the features of the i-th and j-th nodes, respectively. Features of the edge between nodes i and j; For gated functions; This is the supply elasticity coefficient.

[0016] Preferably, the breakpoint score is calculated based on the node failure probability and betweenness centrality, and the expression is:

[0017] In the formula, Sigmoid is the activation function, and MLP is a multilayer perceptron classifier used to output the intrinsic failure probability based on feature learning. For the node features of the i-th node in the L-th iteration, The betweenness centrality of a node reflects its importance in the topology. This is the adjustment coefficient.

[0018] Preferably, when using the SIR model for deduction, the infection rate is:

[0019] In the formula, Basic reproduction rate, Let j be the degree of technological dependence of upstream node j on downstream node i. Let t be the market substitutability indicator at time t.

[0020] Preferably, the calculation steps of the dynamic market substitutability index include: Search for a set of alternative suppliers with the qualifications to produce similar materials; Obtain the real-time production schedule of each supplier in the set and calculate the remaining available capacity ratio; Construct a substitution effectiveness decay function, taking into account qualification verification and production preparation time, and dynamically adjust substitutability.

[0021] Preferably, based on the propagation evolution curve, the expected distribution of the risk loss function is predicted by fusing multimodal time series data, including constructing a cross-modal time series Transformer model, using the propagation evolution curve as the base sequence, fusing the multimodal time series data, and predicting the expected distribution of the risk loss function; The multimodal time-series data is first vectorized and aligned with the material supply numerical sequence from the time dimension.

[0022] The multimodal time series data includes material supply quantity sequences, raw material price data, policy text information, and meteorological information data.

[0023] Preferably, the cross-modal temporal Transformer model includes a cross-modal gating unit for dynamically adjusting the influence weights of text semantics on the numerical prediction stream in the following manner:

[0024] In the formula, and These are the hidden layer representations for text and numerical modalities, respectively. For gated activation functions, This represents the cross-modal gating weight matrix.

[0025] Preferably, the moment-constrained uncertainty set is constructed as follows:

[0026] In the formula, μ and Σ are the mean vector and covariance matrix extracted from the expected distribution of the risk loss function, respectively, and the mean vector and covariance matrix obtained by Transformer prediction, where ξ is a random risk variable. It represents the mathematical expectation under the probability distribution P.

[0027] Furthermore, based on the bibliophilic bar optimization model, the optimal configuration scheme is solved with the objective of minimizing CVaR under the worst-case distribution.

[0028] This invention provides a method for predicting the propagation of supply chain risks and optimizing resilience, which, compared with existing technologies, offers at least the following advantages: 1. By constructing an edge graph neural network for the supply chain, the features of nodes and edges are explicitly modeled. Unlike traditional graph neural networks that only update node features, the edge graph neural network of this invention can more accurately represent the transmission pressure on the supply and demand chain through iterative refinement of edge features, thereby accurately locating weak links such as breakpoints and bottlenecks in the supply chain network and realizing intelligent resilience diagnosis from "point" to "chain". 2. By constructing a scenario tree based on power grid engineering projects and introducing an improved SIR (Susceptible-Aware-Related-Recovery) model, we simulate the evolution of risk information and physical shocks in complex supply chain systems, thereby quantifying the diffusion patterns of risks throughout the entire supply chain of materials and project construction.

[0029] 3. Utilizing a temporal Transformer model, this application efficiently integrates cross-modal, multi-source, heterogeneous data such as policy changes, market conditions, and media reports. Through temporal relationship extraction and a variable masking mechanism, it overcomes the problem of data misalignment across different time scales, enabling intelligent forward-looking projection of risk scores, completion time deviations, cost increases, and potential impact ranges. Furthermore, to alleviate the scarcity of extreme risk event samples, this application utilizes Generative Adversarial Networks (GANs) for risk sample augmentation. The generator produces virtual risk sequences with extreme volatility characteristics using random noise, while a discriminator distinguishes between real and generated samples. Through adversarial training, the model gradually approximates the real risk distribution, thereby expanding the sparse risk event database and improving the predictive ability of the cross-modal temporal Transformer model for black swan events.

[0030] 4. Building upon this foundation, a robust decision-making framework immune to distributional biases is constructed by calculating the Conditional Value at Risk (CVaR) and establishing risk confidence intervals using the Bootstrapping method. This framework does not rely on a single probability distribution but rather seeks the resource allocation strategy that minimizes the worst-case expected loss from a set of distributions satisfying known statistical characteristics. This ensures that, while guaranteeing the stability of critical paths such as substation construction, resources can be proactively allocated, reducing the additional costs caused by extreme risks and fundamentally enhancing supply chain resilience. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0032] Figure 1 This is a flowchart of the supply chain risk propagation prediction and resilience optimization decision-making method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the edge graph neural network edge point joint update structure provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of risk propagation dynamics state transitions provided in an embodiment of the present invention; Figure 4 This provides a decision framework for the split-bar optimization model in the embodiments of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] This invention discloses a supply chain risk propagation prediction and resilience optimization decision-making method, which aims to achieve full-process control of power material supply chain risks through multi-dimensional data-driven and advanced algorithm coupling.

[0035] In this embodiment, as Figure 1 As shown, the steps include: S1. Construct a power supply chain relationship graph, and introduce a technology dependence operator into the node features and a supply elasticity coefficient into the edge features; S2. Iteratively update the power material supply chain relationship diagram, and identify weak nodes based on breakpoint scoring after convergence. S3. For the weak nodes, simulate the evolution of the supply chain state under external shocks based on the scenario tree and SIR model to obtain the propagation evolution curve of the proportion of affected nodes over time. S4. Based on the propagation evolution curve, predict the expected distribution of the risk loss function by fusing multimodal time series data; S5. Construct a moment-constrained uncertainty set based on the expected distribution of the risk loss function, and generate an optimal configuration scheme in the uncertainty set.

[0036] In an optional embodiment, in S1, a digital supply chain relationship graph is first constructed for a specific power grid engineering project (such as a 330kV new substation project). This relationship graph G=(V, E) contains multiple types of heterogeneous nodes: the node set V is defined as supplier nodes (attributes include capacity and technical route), logistics nodes (attributes include transfer capacity and average time consumption), and engineering nodes (attributes include current progress and planned duration); the edge set E contains physical supply edges, contractual relationship edges, and geographical proximity edges.

[0037] Furthermore, this invention employs an embedding layer to encode heterogeneous features. For node features... Sum of edge features Mapped to a d-dimensional latent space:

[0038] in, is a non-linear activation function, W is the weight coefficient, and b is the bias term.

[0039] In some preferred implementations, a technology dependency operator is introduced into the node features to enhance the model's perception of power industry-specific attributes. Introducing a supply elasticity coefficient into edge characteristics .

[0040] In this implementation plan, the calculation logic of the technology dependence operator is as follows: Read the BOM (Bill of Materials) structure tree of the power grid project, count the number of qualified suppliers corresponding to the material at that node, and the hierarchical depth of that material in the BOM; the calculation formula is:

[0041] In the formula, Indicates the number of qualified suppliers of materials. This indicates the level depth of resources within the Bill of Materials (BOM). Indicates the maximum level depth. and This indicates the preset weighting coefficient.

[0042] The larger the value of this operator, the stronger the technological monopoly of the node's materials and the lower it is in the assembly chain, making it extremely difficult to repair once the supply is cut off.

[0043] Supply elasticity coefficient Instead of static assignment, the regression slope is obtained by performing log-linear regression analysis on historical (past 3 years) raw material futures price indices (such as copper and aluminum prices) and historical logistics unit prices for this route. This coefficient represents the rate of change in logistics resistance caused by unit raw material price fluctuations; that is, the percentage increase in cost for this transportation route for every 1% increase in raw material prices, used to accurately quantify the inhibitory effect of price fluctuations on the feasibility of physical transportation.

[0044] In an optional embodiment, in S2, a deep resilience diagnosis is further performed based on the power material supply chain association graph. The core of this is the point-edge joint update mechanism, which includes calculating the edge-guided attention coefficient and using the edge-guided attention coefficient to aggregate neighborhood information, thereby updating node features and edge features.

[0045] In some specific implementation schemes, at each layer iteration First, the edge-guided attention coefficients between nodes are calculated. The calculation of this coefficient not only considers the characteristics of the connecting edges and integrates the characteristics of the source and target nodes, but also introduces a physical gating mechanism based on supply elasticity. In this embodiment, the specific calculation formula is as follows:

[0046] In the formula, Softmax represents the normalization function, LeakyReLU represents the nonlinear activation function, and W represents the feature embedding weight matrix. and These are the features of the i-th and j-th nodes, respectively. The edge features between nodes i and j The supply elasticity coefficient, For the gate function, when the supply elasticity coefficient When the system is in a period of high sensitivity and fluctuation and the cost exceeds the threshold, the attention weight is forcibly suppressed, causing the gating function output to be 0, and the attention weight of that side is directly cut off, simulating the logistics disruption effect under high cost.

[0047] This design makes the feature propagation of neural networks not just a mathematical numerical smoothing, but a real simulation of the physical phenomenon of "logistics path disruption due to excessive costs" in reality, which is something that existing general GNN models cannot achieve.

[0048] Furthermore, the neighborhood information is aggregated using the edge-guided attention coefficient, the process of which is referred to... Figure 2 Update node characteristics, including:

[0049] In the formula, This represents the node feature update weight matrix, which is used to perform a linear transformation on the aggregated neighborhood features to obtain a new node representation. Let i represent the neighborhood set of node i.

[0050] Subsequently, to simulate the feedback of chain stress due to changes in node state (such as supplier production cuts), edge features are reconstructed using the updated node features:

[0051] After multiple iterations and convergence, the breakpoint score for each node is calculated. The system will compare the score with a preset threshold. When comparing, At that time, the node This node was identified as a weak point in the supply chain. The score is obtained by a weighted fusion of node failure probability and network centrality.

[0052] In the formula, Sigmoid is the activation function, and MLP is a multilayer perceptron classifier used to output the intrinsic failure probability based on feature learning. For the node features of the i-th node in the L-th iteration, The betweenness centrality of a node reflects its importance in the topology. This is the adjustment coefficient.

[0053] The node failure probability is output by the trained multilayer perceptron classifier and is used to characterize the intrinsic risk level of a node experiencing supply disruption when subjected to external shocks. This failure probability is further weighted and fused with the node's betweenness centrality to calculate the node's breakpoint score. This allows for the identification of weak points in the supply chain.

[0054] In an optional embodiment, in S3, for the weak node, the evolution of the supply chain state under external shocks is simulated based on a scenario tree and a SIR model. Specifically, this application first constructs a branching evolution logic based on a power grid engineering scenario tree, where each branch represents a specific external shock scenario. Simultaneously, during the propagation process, an improved SIR model is introduced for state iteration. Wherein, as... Figure 3 The improved SIR model divides the supply chain node status into four types: risk-vulnerable state S, risk-aware but irrelevant state I, risk-aware and relevant state R, and risk-recovered state; the transition rate between states is dynamically determined by the bottleneck of material supply, capacity utilization rate, and logistics timeliness characteristics.

[0055] In this embodiment, the probability that node i is infected by its affected neighbor node j at time t is defined as follows: This probability is not a fixed constant, but depends on the technological dependence of upstream node j on downstream node i. and market substitutability indicators Dynamic correction is performed and defined as the time-varying coupling infectivity rate. The expression is:

[0056] In the formula, Basic reproduction rate, The degree of technical dependence of upstream node j on downstream node i (determined by the uniqueness of the BOM list). Let t be the market substitutability indicator at time t.

[0057] This improvement enables the model to distinguish between "general goods" (low dissemination rate) and "core proprietary equipment" (high dissemination rate), thereby addressing the technical deficiency of the traditional SIR model in failing to characterize differences in supply chain levels.

[0058] Furthermore, the calculation steps for the dynamic market substitutability indicator include: Alternative supplier search: When node i shows signs of risk, the State Grid e-commerce platform database is searched in real time to identify all alternative suppliers Ω with the qualifications to produce similar materials; Dynamic capacity assessment: Obtain the real-time production schedule of each manufacturer in the set Ω and calculate their remaining available capacity ratio; Time-sensitivity decay modeling: Considering the qualification verification and production preparation time required for introducing new suppliers, a substitution effectiveness decay function is defined:

[0059] In the formula, express t The market substitutability metric at any given time describes the actual substitutability level that emerges as alternative suppliers complete their qualification reviews and production preparations over time. Indicates the first The qualification weighting coefficient of each alternative supplier, and the access rate coefficient for substitution capability. It refers to the time when a risk event is triggered or the time when an alternative supply mechanism is activated.

[0060] This formula describes the process by which alternative manufacturers gradually switch production and integrate into the supply network as the crisis duration t progresses. In the early stages of the risk (when t is relatively small). Low transmission rate The infection rate is extremely high; however, as substitution increases over time, the infection rate gradually decreases. This mechanism accurately describes the "time-varying characteristics" of supply chain risk transmission.

[0061] The most preferred option is the risk recovery rate of the nodes in this application. Mapped to:

[0062] in, Let i be the resource reserve of node i. Represents a node Supply complexity index This represents the resilience adjustment coefficient. As shown in the formula above, the risk recovery rate of a node is positively correlated with its resource reserves and negatively correlated with the node's supply complexity, thus reflecting the differences in resilience among different nodes after supply chain shocks.

[0063] Finally, by solving a system of differential equations containing dS / dt, dI / dt, and dR / dt, the time-varying R(t) function curve is obtained. The value on this curve represents the proportion of project nodes in a disaster-affected state (Recovered / Removed, referring to those halted or requiring repair due to risk shocks) at time t. This indicator directly maps to the average expected delay and total material shortage of the project, providing a quantitative basis for forward-looking early warning.

[0064] Furthermore, in an optional embodiment, S4 includes constructing a cross-modal temporal Transformer model, using the propagation evolution curve as the base sequence, fusing the multimodal temporal data, and predicting the expected distribution of the risk loss function.

[0065] In this embodiment, the multimodal time series data includes: Material supply quantity sequence Raw material price data Policy text information Meteorological information data.

[0066] Numerical data (such as material supply, raw material prices and meteorological indicators) are directly normalized and then input into the numerical sequence encoding module; textual data (such as policy announcements and industrial policies) are semantically vectorized through a pre-trained language model and then mapped to a unified time axis through a time alignment mechanism to extract semantic feature vectors of keywords such as "production restriction", "tariff", and "force majeure". A time alignment mapping function is constructed to map the features of discontinuously released policy texts to a continuous material supply timeline, including identifying the effective time interval [t1, t2] of text-based policy information; and generating a mask vector Mpolicy of the same length as the material supply numerical sequence, which is filled only at the time steps corresponding to [t1, t2]. The remaining time steps are filled with zero vectors.

[0067] In one implementation, the multi-head attention mechanism within the Transformer employs a cross-modal gating unit to process the input. Unlike the standard Transformer's direct addition, this invention dynamically adjusts the weights of the numerical stream based on the strength of the text vector.

[0068] In the formula, For gated activation functions, and These are the hidden layer representations for text and numerical modalities, respectively. This represents the cross-modal gating weight matrix.

[0069] This structure ensures that the attention bias of text features on numerical prediction is only activated when the magnitude of the text feature vector exceeds a set threshold (in the event of a major policy emergency), significantly altering the original trajectory of material supply prediction and avoiding noise interference from irrelevant daily news on the prediction results.

[0070] In a preferred embodiment, this application utilizes Generative Adversarial Networks (GANs) for negative sample sampling to address the difficulty of modeling sparse risk events. During sampling, a generator simulates supply chain indicator fluctuations under extreme risk conditions, and a discriminator performs a true / false game to expand the sparse risk event database. In this embodiment, the generator G aims to learn the distribution of historical risk samples, transforming random noise into a virtual risk sequence with extreme fluctuations through a mapping function. The discriminator D is optimized using the following objective function:

[0071] Among them, E x ~ p_data(x) [log] D ( x ] represents the expected accuracy of the discriminator in identifying real risk samples x, while E z ~ p_z(z) [ log(1 D ( G ( z This reflects how the discriminator maps the generator's virtual samples, which are derived from noise z. G ( z The ability to identify fabricated data. A well-trained G can generate a large number of negative samples, improving the model's ability to capture black swan events. The model ultimately outputs a revised future risk score time-series curve and the potential impact range.

[0072] In an alternative embodiment, after obtaining the risk score, a decision is made through Distributed Robust Optimization (DRO). In this embodiment, reference is made to... Figure 4 First, we define Conditional Value at Risk (CVaR) as the core metric for decision variables. (e.g., the amount of reserve materials) and random risk ξ, CVaR characterizes the worst-case scenario up to (1- Expected loss of proportion:

[0073] Since the true probability distribution P is unknown, this invention constructs an uncertainty set based on moment constraints. :

[0074] In the formula, ξ is a random risk variable. Let μ and Σ represent the mathematical expectation under the probability distribution P, and μ and Σ be the mean vector and covariance matrix extracted from the expected distribution of the risk loss function, respectively. The mean vector and covariance matrix are obtained through Transformer prediction, specifically by collecting the prediction error sequence ε of the Transformer model on the validation set. ytrue - y pred The empirical distribution of the error is calculated; based on the expected distribution of the risk loss function predicted by the Transformer model, its mean vector μ and covariance matrix Σ are extracted, and a moment-constrained uncertainty set is constructed accordingly to characterize the possible distribution range of the random risk variable under the condition that the statistical characteristics are known but the probability distribution is unknown, so as to ensure that the decision scheme is robust to the prediction blind spot of the model.

[0075] Furthermore, based on the bibliophilic bar optimization model, the optimal configuration scheme is solved with the objective of minimizing CVaR under the worst-case distribution.

[0076] The Bruker bar optimization model is expressed as follows:

[0077] in, For emergency resource allocation decision variables, c is the allocation cost vector, P is the probability distribution uncertainty set, ξ is the risk impact vector, and L(x, ξ) is the loss function under risk ξ, specifically characterized by CVaR.

[0078] This application employs a column constraint generation algorithm (C&CG) for a two-level iterative solution. During the solution process, the optimization problem is broken down into a main problem and sub-problems. The main problem seeks the optimal configuration under the current scenario, while the sub-problems search for the extreme risk distribution that maximizes the loss within the uncertainty set. Convergence is achieved through alternating iterations of these two approaches. In the iteration process, the sub-problems are responsible for finding the extreme distribution scenario that maximizes the loss of the current decision; that is, finding the probability distribution P that maximizes the CVaR value under the moment constraint conditions. The main problem is to adjust the configuration strategy based on the feedback of extreme scenarios, and finally obtain the layout solution.

[0079] This application further illustrates the point by taking a newly built 330kV substation in the Ningxia power grid as an example.

[0080] In this embodiment, a proposed 330kV substation expansion project in the Ningxia power grid is selected as the technology verification vehicle. The core equipment involved in this project includes 330kV ultra-high voltage circuit breakers, integrated switchgear (GIS), and large power transformers. Its supply chain covers a logistics path of thousands of kilometers from the production base in East China to the construction site in the Northwest Plateau, and the construction period is strictly limited by seasonal climate and power grid dispatching windows.

[0081] In the implementation phase, the system first retrieved the heterogeneous graph data associated with the project, which included 14 core Tier 1 supplier nodes and 28 Tier 2 raw material supplier nodes. In the first phase of resilience diagnosis, the edge graph neural network extracted features from the production data of the core circuit breaker suppliers. This was done by analyzing the edge features of the suppliers' raw material (such as insulating gas and contact alloy) supply edges. Through multi-level iterative refinement, the model captures the attention weights of its upstream key raw material supply paths. A sustained period of fluctuation occurred. Combining the node's capacity utilization characteristics, Edge GNN ultimately output a failure probability of 0.85 for the circuit breaker supplier in the breakpoint classifier. This high score not only reflects the supplier's own unstable capacity but also explicitly reveals the high vulnerability of the logistics channel under specific climatic conditions through edge features.

[0082] Subsequently, the system enters the risk propagation simulation phase based on scenario trees and an improved SIR model. In this embodiment, the initial risk is set as "delivery of the core circuit breaker is delayed due to a shortage of critical components." In the discrete-time simulation of the SIR model, the risk spreads from the "manufacturer node" to the "logistics transit node" and then to the "engineering installation node." The calculation shows the contagion rate of this risk within the supply chain network. Due to the high uniqueness of this 330kV circuit breaker, its impact coefficient on downstream construction progress reaches 0.92, adjusted by the weighting of technological dependence. Solving the state transition equations, the prediction results show that without intervention, the project will reach the peak of the disaster state (R state) on day 45, meaning the critical path will face a 45-day absolute delay. This will directly prevent the substation from being put into operation before the scheduled peak summer power demand window, resulting in significant power shortages.

[0083] Upon sensing this severe consequence, the system triggered the fourth stage of cross-modal time-series Transformer prediction. By integrating recent industry policies, raw material market conditions, and logistics and meteorological data from Ningxia, the risk score curve was finely revised, resulting in the expected distribution of the risk loss function. Subsequently, the system enters the most crucial fifth stage—distributed bar optimization decision-making. Addressing the challenge of uncertain equipment distribution in 330kV substations, the model constructs an uncertainty set based on moment constraints. This takes all possible probabilistic deviations into account. By executing the Column Constraint Generation Algorithm (C&CG), the main problem is configuration cost. Finding a balance between CVaR risk and loss.

[0084] Ultimately, the Bruker optimization model provides a forward-looking layout solution that balances resilience and cost. This solution explicitly states that instead of stockpiling all spare equipment as in traditional solutions (which would tie up tens of millions of yuan in capital), it only requires pre-stocking 15% of critical vulnerable parts and core components in a central warehouse in Ningxia, while simultaneously activating a "cloud production capacity" linkage mechanism. "Cloud production capacity" refers to a contractual agreement that, in the event of the worst-case scenario, grants priority production rights to another alternative manufacturer located in North China.

[0085] In subsequent actual project monitoring, the original circuit breaker supplier did experience a small-scale production interruption due to fluctuations in the external environment. Because this invention pre-implemented a 15% reserve of critical spare parts, on-site installation and commissioning work did not completely halt; instead, a flexible construction plan of "installation first, commissioning later" was achieved through component replacement. Final statistics showed that the emergency response cost for this project was only about 10% of the expected loss. Compared to traditional, experience-based static reserve solutions (which typically require an inventory ratio of over 30%), this invention reduced the total configuration cost by 12% and demonstrated strong decision-making robustness under extreme probability deviations, effectively ensuring the timely commissioning of the Ningxia power grid's 330kV project. This shift from "passive emergency response" to "proactive resilience optimization" provides rigorous and practical technical support for the intelligent management of the power material supply chain.

[0086] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0087] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A supply chain risk propagation prediction and resilience optimization decision-making method, characterized in that, include: Construct a power supply chain relationship graph, and introduce a technology dependence operator into the node features and a supply elasticity coefficient into the edge features; The power supply chain relationship diagram is iteratively updated, and weak nodes are identified based on breakpoint scoring after convergence. For the aforementioned vulnerable nodes, the evolution of the supply chain state under external shocks is simulated based on the scenario tree and SIR model, and the propagation evolution curve of the proportion of affected nodes over time is obtained. Based on the propagation evolution curve, the expected distribution of the risk loss function is predicted by fusing multimodal time series data; Based on the expected distribution of the risk loss function, a moment-constrained uncertainty set is constructed, and an optimal configuration scheme is generated from the uncertainty set.

2. The method as described in claim 1, characterized in that, In the power supply chain diagram, nodes include supplier nodes, logistics nodes, and / or engineering nodes; edges include physical supply edges, contractual relationship edges, and / or geographical proximity edges.

3. The method as described in claim 1, characterized in that, The technology dependency operator is calculated based on the number of qualified suppliers and the hierarchy depth of the materials in the bill of materials structure tree; The supply elasticity coefficient is determined based on the slope of the logarithmic linear regression of historical raw material prices and logistics costs.

4. The method as described in claim 1, characterized in that, The power supply chain association graph is iteratively updated, including calculating the edge-guided attention coefficient and using the edge-guided attention coefficient to aggregate neighborhood information to update node features and edge features; The formula for calculating the edge-guided attention coefficient is as follows: In the formula, Softmax represents the normalization function, LeakyReLU represents the nonlinear activation function, and W represents the feature embedding weight matrix, which is used to map the original node features or edge features to the d-dimensional latent representation space. and These are the features of the i-th and j-th nodes, respectively. Features of the edge between nodes i and j; For gated functions; This is the supply elasticity coefficient.

5. The method as described in claim 1, characterized in that, The breakpoint score is calculated based on the node failure probability and betweenness centrality, and the expression is: In the formula, Sigmoid is the activation function, and MLP is a multilayer perceptron classifier used to output the intrinsic failure probability based on feature learning. For the node features of the i-th node in the L-th iteration, The betweenness centrality of a node reflects its importance in the topology. This is the adjustment coefficient.

6. The method as described in claim 1, characterized in that, When using the SIR model for extrapolation, the infection rate is: In the formula, Basic reproduction rate, Let j be the degree of technological dependence of upstream node j on downstream node i. Let t be the market substitutability indicator at time t.

7. The method as described in claim 6, characterized in that, The calculation steps for the market substitutability indicator include: Search for a set of alternative suppliers with the qualifications to produce similar materials; Obtain the real-time production schedule of each supplier in the set and calculate the remaining available capacity ratio; Construct a substitution effectiveness decay function, taking into account qualification verification and production preparation time, and dynamically adjust substitutability.

8. The method as described in claim 1, characterized in that, Based on the propagation evolution curve, the expected distribution of the risk loss function is predicted by fusing multimodal time series data, including constructing a cross-modal time series Transformer model, using the propagation evolution curve as the base sequence, fusing the multimodal time series data, and predicting the expected distribution of the risk loss function. The multimodal time-series data is first vectorized and aligned with the material supply numerical sequence from the time dimension.

9. The method as described in claim 8, characterized in that, The cross-modal temporal Transformer model includes a cross-modal gating unit, used to dynamically adjust the weights of the influence of text semantics on the numerical prediction stream in the following manner: In the formula, and These are the hidden layer representations for text and numerical modalities, respectively. For gated activation functions, This represents the cross-modal gating weight matrix, used to calculate the modulation weights of text modal features on numerical modal features, in order to control the degree of influence of different modal information in the fusion process.

10. The method as described in claim 1, characterized in that, Construct the moment-constrained uncertainty set as follows: In the formula, μ and Σ are the mean vector and covariance matrix extracted from the expected distribution of the risk loss function, respectively, and are obtained by Transformer prediction. ξ is a random risk variable. This represents the mathematical expectation under probability distribution P; Furthermore, based on the bibliophilic bar optimization model, the optimal configuration scheme is solved with the objective of minimizing CVaR under the worst-case distribution.

Images