Supply chain risk early warning method based on deep learning

Abstract

The invention relates to the technical field of supply chain risk early warning, in particular to a supply chain risk early warning method based on deep learning, and the method comprises the steps: obtaining supply chain data, extracting a material circulation relation between supply chain nodes, constructing a node relation graph, and employing a graph neural network to achieve the aggregation of the features of the nodes and adjacent nodes. And space correlation characteristics in the global network are extracted, and a multi-stage transmission and diffusion path of the risk is effectively modeled. And then, splicing node space features and historical time sequence features, inputting the spliced features into a long-short-term memory network, dynamically capturing the evolution trend of node risks along with time, identifying periodic fluctuations and sudden anomalies, and improving the prediction precision of the risk trend. And finally, a multi-layer perceptron is adopted to carry out nonlinear mapping and feature fusion on risk time sequence features output by the long-short-term memory network, node risk scores are generated, real-time early warning of high-risk nodes is realized accordingly, and the accuracy and timeliness of supply chain risk monitoring are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of supply chain risk early warning, and in particular to a supply chain risk early warning method based on deep learning. Background Art

[0002] The power system supply chain is complex, involving multiple links, including raw material procurement, equipment manufacturing, logistics and transportation, warehousing management, and engineering construction. Especially during grid construction and operation and maintenance, the supply chain often spans multiple regions, connecting numerous suppliers and transportation nodes. The safety and stability of its operation directly impacts the grid's power supply reliability and project progress. In recent years, the power supply chain has faced increasing uncertainty and risk due to multiple factors, including natural disasters, public emergencies, logistics bottlenecks, and fluctuations in supplier performance. Unusual fluctuations in supply chain nodes, such as delayed delivery of critical equipment, transportation disruptions, or supplier defaults, can easily lead to project delays, increased operation and maintenance costs, and even localized or systemic power risks.

[0003] Traditional supply chain risk early warning methods primarily rely on historical statistical analysis, empirical judgment, or rule-based models based on single-dimensional threshold settings. These methods generally suffer from the following shortcomings: First, they have limited capabilities for integrating and mining multi-source heterogeneous data, failing to fully utilize both structured and unstructured information within the supply chain; second, they struggle to model the complex dependencies and impact paths between nodes, failing to accurately capture the transmission and diffusion of risk within the network; and third, they inadequately monitor temporal characteristics, making it difficult to promptly identify evolving trends in risk factors, resulting in delayed early warnings and insufficient accuracy. Summary of the Invention

[0004] To solve the above problems, the present invention provides a supply chain risk early warning method based on deep learning.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A supply chain risk early warning method based on deep learning, comprising the following steps:

[0007] Access supply chain data;

[0008] Based on supply chain data, a node relationship graph containing supply chain nodes and topological structures is constructed, and the spatial correlation features between nodes are extracted through graph neural networks to obtain spatial feature representation;

[0009] Obtain historical time series data of supply chain nodes from supply chain data, input spatial feature representation and historical time series feature data of supply chain nodes into the time series deep learning model, model historical risk factors, and output risk time series feature representation;

[0010] The risk time series features are processed by a multi-layer perceptron to generate risk score data for the target node;

[0011] Output supply chain risk warning information based on risk scoring data.

[0012] Furthermore, obtaining supply chain data includes:

[0013] The system collects supplier performance information, material procurement and delivery data, logistics transportation status, equipment operation status, contract execution status, node geographical location and related meteorological environment data of supply chain nodes, and performs data cleaning, normalization and time synchronization processing to generate a structured supply chain data set.

[0014] Furthermore, the construction of a node relationship graph including supply chain nodes and a topological structure includes the following steps:

[0015] Based on the structured supply chain data set, determine the material flow relationship between supply chain nodes and generate connection data between nodes;

[0016] Based on the connection data between nodes, each supply chain node is connected to its directly associated node, and a node relationship graph is constructed that reflects the node attributes and the dependency relationship between nodes.

[0017] Furthermore, the graph neural network is constructed by the following steps:

[0018] Based on the node relationship graph, the structural features of each supply chain node are extracted as node input data, and the connection data between nodes are extracted as edge input data;

[0019] Based on historically known risk events and supply chain anomalies, a training set containing node input data, edge input data, and corresponding risk labels is constructed.

[0020] The training set is input into the graph neural network, the information of each node and its associated nodes is aggregated and propagated, a spatial feature representation containing associated nodes is generated, and the network parameters are backpropagated to obtain the trained graph neural network.

[0021] Furthermore, acquiring historical time series data of supply chain nodes from supply chain data includes:

[0022] Based on the structured supply chain data set, for each supply chain node, supplier status, material procurement and delivery progress, logistics and transportation information, equipment operating parameters, contract performance records, and related meteorological and environmental data are extracted in chronological order to generate the node's historical time series raw data;

[0023] Based on the original historical time series data of the node, missing value processing and normalization are performed to obtain the historical time series feature data of the node.

[0024] Furthermore, the temporal deep learning model includes a long short-term memory network model.

[0025] Furthermore, the time series deep learning model is constructed by the following steps:

[0026] Based on the node's historical time series feature data and spatial feature representation, combined with known risk event labels, an LSTM training set containing input feature sequences and risk labels is constructed;

[0027] Inputting the LSTM training set into a long short-term memory network model, jointly modeling the node's historical temporal feature data and spatial feature representation, and predicting the risk score or risk probability in a future specified time window;

[0028] Based on historical risk labels and model prediction results, the model parameters are optimized through the back-propagation algorithm to obtain the trained long-short-term memory network model.

[0029] Furthermore, the loss function of the long short-term memory network model is as follows:

[0030]

[0031] Among them, L is the loss value; is the number of training samples; N is the number of training samples; C is the total number of risk categories; The true risk label of the jth category for the i-th sample at the future prediction time point; It is the predicted risk probability of the long short-term memory network model for the i-th sample belonging to the j-th category at the future prediction time point.

[0032] Furthermore, the processing of risk time series features by a multi-layer perceptron includes the following steps:

[0033] Based on the risk time series feature representation output by the time series deep learning model, normalizing the risk time series feature representation to obtain normalized feature data;

[0034] Based on the normalized feature data, the normalized feature data is sequentially input into each hidden layer of the multi-layer perceptron for feature transformation to obtain fused feature data;

[0035] Based on the fused feature data, the fused feature data is input into the output layer of the multi-layer perceptron to generate the risk score data of the target node.

[0036] Furthermore, outputting supply chain risk warning information based on risk score data includes:

[0037] Based on the risk score data of the target node, the risk score data is compared with a preset risk warning threshold. If the risk score data is greater than or equal to the warning threshold, it is determined that the target node has a supply chain risk and corresponding risk warning information is generated;

[0038] If the risk score data is less than the warning threshold, it is determined that there is no supply chain risk at the target node.

[0039] The beneficial effects of the present invention lie in: During the supply chain data collection and preprocessing phase, the present invention cleans, normalizes, and time-synchronizes heterogeneous data from multiple sources, including supplier performance, procurement delivery, logistics and transportation, equipment operation, contract execution, geography, and meteorology, to generate a high-dimensional, structured dataset. This processing step not only ensures data consistency and integrity but also provides sufficient and high-quality input features for subsequent deep learning modeling, effectively overcoming the limitations of traditional methods, which suffer from single data and insufficient information dimensionality. Regarding network structure modeling, based on the structured dataset, the material flow relationships between supply chain nodes are automatically extracted. By constructing a node relationship graph, node attributes and inter-node connectivity are encoded as graph-structured data. Utilizing graph neural networks, this solution iteratively aggregates node features and their adjacent node features, extracting the spatial correlation characteristics of nodes in the global network through parameter sharing and neighborhood information fusion. This approach explicitly models the multi-level transmission and diffusion paths of risk in the supply chain network, dynamically reflects complex dependencies between nodes, and enables the capture and propagation of risk signals at the network level, addressing the limitations of traditional methods, which struggle to express topological risk. Node spatial features are jointly modeled with node historical temporal features. By splicing the spatial feature representation output by the graph neural network with the node's historical time series feature data and inputting it into a long short-term memory network (LSTM), dynamic time series modeling of risk factors is achieved. The LSTM network can capture the evolution trend of node risk characteristics over time, identify complex dynamic patterns such as cyclical fluctuations, sudden anomalies, and historical dependencies, and significantly improve the ability to proactively identify and predict risk trends. Compared with traditional static statistical or empirical threshold models, it can achieve more accurate risk trend analysis. Finally, in the output link, the solution uses a multi-layer perceptron (MLP) to perform nonlinear mapping and feature fusion on the risk time series features output by the LSTM, further improving the ability to discriminate node risk scores. By comparing the final risk score with the set threshold, real-time early warning of high-risk nodes is achieved, effectively improving the accuracy and timeliness of risk monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flowchart of the steps of a supply chain risk early warning method based on deep learning in the present invention.

[0041] Figure 2This is a flowchart of the steps of processing risk time series features through a multi-layer perceptron in the present invention. DETAILED DESCRIPTION

[0042] See also Figure 1-Figure 2 As shown, the present invention relates to a supply chain risk early warning method based on deep learning, comprising the following steps:

[0043] Access supply chain data;

[0044] Based on supply chain data, a node relationship graph containing supply chain nodes and topological structures is constructed, and the spatial correlation features between nodes are extracted through graph neural networks to obtain spatial feature representation;

[0045] Obtain historical time series data of supply chain nodes from supply chain data, input spatial feature representation and historical time series feature data of supply chain nodes into the time series deep learning model, model historical risk factors, and output risk time series feature representation;

[0046] The risk time series features are processed by a multi-layer perceptron to generate risk score data for the target node;

[0047] Output supply chain risk warning information based on risk scoring data.

[0048] It should be noted that, taking the supply chain management of core equipment in power grid substations as the scenario, firstly, full data collection is carried out for the supply chain nodes of all key materials of the substation, including but not limited to the supplier fulfillment rate of each node, material production and delivery plan, logistics transportation status, warehouse inventory level, equipment operation status, contract fulfillment information, and the geographical location and real-time weather conditions of the node. After pre-processing operations such as missing value filling, outlier removal, normalization and time synchronization, the above multi-source heterogeneous data forms a high-dimensional structured data set to ensure that all types of input features can be directly used for subsequent modeling and analysis. In the supply chain network modeling stage, the system automatically constructs a node relationship graph containing all supply chain nodes and their topological dependencies based on the material flow path and the business relationship between nodes, and encodes the structured attributes of each node and the connection relationship between nodes into graph structure data. Subsequently, a graph neural network model is used to aggregate and embed node and neighborhood information. Through parameter training and iterative optimization, the system is able to extract the spatial correlation characteristics of each node in the entire supply chain network. In particular, it can dynamically perceive the risk transmission effects caused by upstream suppliers or logistics nodes, thereby achieving global perception and local amplification of risks, overcoming the limitations of traditional methods that cannot model network topology relationships and risk diffusion paths. Furthermore, historical time series data of each node (such as on-time delivery rate, equipment operation anomalies, inventory fluctuations, logistics interruption records, etc. over the past 12 months) is spliced ​​with the spatial features output by the graph neural network and input into a long short-term memory network (LSTM) model. LSTM can comprehensively utilize the temporal dynamic characteristics of the node's historical risk factors and the node's spatial status in the supply chain network to deeply model complex time series patterns such as risk evolution trends over time, cyclical fluctuations, and abnormal jumps, thereby providing high-precision predictions of risk scores or risk probabilities within a specified future time window. Compared with methods that rely solely on static thresholds or single feature analysis, this solution significantly improves forward-looking early warning capabilities for potential risk outbreaks, effectively avoiding missed and false alarms caused by historical myopia or spatial isolation. Finally, the system inputs the risk time-series features output by the LSTM into a multi-layer perceptron (MLP). Through nonlinear mapping and feature fusion, the system further enhances the discrimination of high-dimensional features and generates risk scores for target nodes. The risk score is compared with the warning threshold in real time. If the threshold is exceeded, a warning message containing the risk node identifier, risk level, key influencing factors, and recommended remedial measures is automatically output, enabling intelligent monitoring and early intervention of high-risk links in the power supply chain. Unlike existing technologies, this deep fusion of graph neural networks and time-series neural networks for power supply chain risk warning not only models risk diffusion across the entire network structure but also fully utilizes dynamic time-series information to achieve multidimensional, dynamic, and forward-looking comprehensive risk assessment. This overcomes the technical bottlenecks of existing methods in multi-source data fusion, risk analysis of complex network topologies, and risk trend prediction.

[0049] Furthermore, obtaining supply chain data includes:

[0050] The system collects supplier performance information, material procurement and delivery data, logistics transportation status, equipment operation status, contract execution status, node geographical location and related meteorological environment data of supply chain nodes, and performs data cleaning, normalization and time synchronization processing to generate a structured supply chain data set.

[0051] Specifically, the system first collects comprehensive data from all key nodes in the supply chain. This includes supplier performance information (such as historical contract fulfillment rates, number of defaults, and delivery timeliness), material procurement and delivery data (such as purchase order volume, actual arrival volume, and batch delivery cycle), logistics and transportation status (such as transportation routes, timeliness, and in-transit anomaly records), equipment operating status (such as key equipment start-up and shutdown frequency, fault alarms, and health index), contract execution status (such as contract fulfillment progress, contract performance objections, and compensation), node location (such as the city or transportation hub where the node is located, and transportation distances to upstream and downstream nodes), and relevant meteorological and environmental data (such as historical and real-time weather information and disaster warning information at the node location). All collected multi-source, heterogeneous data undergoes unified data cleaning, including missing value filling, duplicate value removal, and outlier detection, to ensure data accuracy and completeness. The system then applies normalization algorithms (such as Z-score and Min-Max normalization) for different feature types to achieve data scale consistency, facilitating subsequent model processing. Finally, all data are strictly synchronized according to timestamps to ensure that all types of data are collaboratively modeled under the same time series benchmark, and ultimately generate a structured supply chain data set.

[0052] Furthermore, the construction of a node relationship graph including supply chain nodes and a topological structure includes the following steps:

[0053] Based on the structured supply chain data set, determine the material flow relationship between supply chain nodes and generate connection data between nodes;

[0054] Based on the connection data between nodes, each supply chain node is connected to its directly associated node, and a node relationship graph is constructed that reflects the node attributes and the dependency relationship between nodes.

[0055] It is important to note that the core attributes of each supply chain node (such as main equipment suppliers, regional warehouses, substations, and logistics service providers) are identified, including node type, fulfillment capabilities, storage capacity, and historical service performance. Subsequently, the system automatically determines whether a direct material flow relationship exists between each pair of nodes by analyzing material procurement, delivery, and logistics data, such as suppliers supplying goods to warehouses, warehouses shipping goods to substations, or logistics companies providing transportation services within a specific interval. For node pairs with business transactions, the system generates connection data between them, clearly defining the attributes of each connection, such as material category, historical transaction frequency, transportation cycle, and collaboration intensity, and annotating the connection direction to reflect upstream and downstream dependencies. Based on this, the system uses each node as a vertex in the graph and establishes these identified node pairs and their connection attributes as edges, forming a complete node relationship graph. This graph structure not only encodes the business attributes of the nodes themselves but also systematically expresses the dependencies and material flow paths between multiple nodes in the supply chain network, providing the fundamental input for subsequent spatial feature learning and risk transfer modeling in graph neural networks.

[0056] Furthermore, the graph neural network is constructed by the following steps:

[0057] Based on the node relationship graph, the structural features of each supply chain node are extracted as node input data, and the connection data between nodes are extracted as edge input data;

[0058] Based on historically known risk events and supply chain anomalies, a training set containing node input data, edge input data, and corresponding risk labels is constructed.

[0059] The training set is input into the graph neural network, the information of each node and its associated nodes is aggregated and propagated, a spatial feature representation containing associated nodes is generated, and the network parameters are backpropagated to obtain the trained graph neural network.

[0060] In some embodiments, first, based on the node relationship graph, the structural features of each node are systematically extracted, including the supplier's fulfillment score, node historical inventory, contract activity, and on-time arrival rate of the transportation node as node input data. At the same time, based on business flow analysis such as material flow and transportation dependence, the connection attributes between nodes such as historical transportation timeliness, collaboration frequency, one-way dependence, and probability of associated risk event propagation are extracted as edge input data. For risk event samples recorded in the system in the past, including delivery interruptions, warehouse failures, contract breaches, and link anomalies caused by natural disasters, the associated nodes and edges are marked one by one to construct a training set that integrates node attributes, edge attributes, and labels. The above training set is modeled using a graph neural network (such as GCN, GraphSAGE, etc.). Through the iterative message passing mechanism, the network not only aggregates its own features, but also integrates the features of all first-order and multi-order neighboring nodes and connecting edges to achieve high-order relationship modeling. During the training phase, historical risk labels are used to calculate cross-entropy loss, and all parameters are updated through a back-propagation mechanism. This ensures that the final node spatial feature expression not only reflects its own risk status but also embeds the topological risk influence of the entire network, enabling efficient transmission and representation of risk information in the global network. Unlike methods based solely on single-node historical thresholds or shallow aggregation, the graph neural network of this embodiment achieves innovations in key aspects such as multi-order neighborhoods, node-edge joint feature modeling, and reverse optimization of full-network risk labels. This significantly improves the accuracy and robustness of spatial risk perception and structured risk modeling in complex supply chain networks.

[0061] In some specific embodiments, the system collects structured data for each node from January 2023 to January 2024: the input of supplier A node is the monthly contract fulfillment rate (such as 0.95 in March 2023 and 0.85 in April 2023), the average delivery cycle (such as 10 days), and the number of production failures (such as 3 times); the input of warehouse B node is the monthly inventory turnover rate (such as 1.8), inventory utilization rate (such as 90%), and storage equipment failure rate (such as 0.03); the input of logistics company C node is the average transportation time (such as 2.2 days), the frequency of transportation interruption (such as 1 time / quarter), and the proportion of congested lines (such as 8%); the input of substation D node is the number of maintenance plans (such as 4 times / year) and the number of equipment operation abnormalities (such as 2 times / half year). The input data of the edge between nodes: the edge between A and B is the transformer category "500kV", the historical delivery frequency is 10 times / year, and the maximum single batch delivery volume is 2 units; the edge between B and C is the historical transportation time of 2 days, the transportation risk zone pass rate is 15%; the edge between C and D is the line risk zone pass rate of 20%, the annual interruption record is 2 times, etc. The historical risk event label collection: For example, in April 2023, the A→B→C→D link was delayed in the arrival of the main transformer of substation D due to the sudden drop in the fulfillment rate of supplier A to 0.85 and the congestion rate of logistics company C in that month increased to 15%, causing the maintenance plan to be postponed. Based on this, the system marks the nodes A, B, C, D and related edges of this cycle as high risk labels (1), while other cycle data without abnormal events are marked as normal (0). The above dataset containing node features, edge features and risk labels is input into the graph neural network (such as GCN), and the network adopts a two-layer structure. The first layer aggregates information about each node's immediate neighbors and edge features. For example, the input for node B in April 2023 is: [inventory turnover rate 1.8, inventory utilization 90%, failure rate 0.03] (node ​​features), along with the edge features from A to B [shipping frequency 10, single batch 2] and from B to C [transportation time 2, risk approval rate 15]. The output of the first layer of the GCN is activated with a ReLU and then enters the second layer, where global neighborhood information is further aggregated. During training, monthly risk labels are used as supervisory signals, and cross-entropy loss and backpropagation are used to optimize parameters. Ultimately, the trained graph neural network can be fed with data for a new cycle in February 2024, such as supplier A's fulfillment rate of 0.88 and logistics company C's congestion rate of 12%, and outputs a real-time spatial risk score (e.g., 0.76) for each node (e.g., substation D) in that cycle. If the score is higher than the system warning threshold (such as 0.7), the system will automatically mark substation D and related links as high-risk, and generate a detailed warning report including major influencing factors (such as "supplier A's fulfillment rate has decreased" and "logistics company C's congestion rate has increased") to provide specific decision-making support for the operation and maintenance management team.This specific embodiment fully demonstrates the complete data flow from multi-source structured data collection, feature extraction, historical label generation, to algorithm model training and reasoning, highlighting the spatial risk identification and tracing capabilities of this solution based on graph neural networks, which is an important technological innovation that distinguishes it from traditional static risk thresholds and isolated node analysis methods.

[0062] Furthermore, acquiring historical time series data of supply chain nodes from supply chain data includes:

[0063] Based on the structured supply chain data set, for each supply chain node, supplier status, material procurement and delivery progress, logistics and transportation information, equipment operating parameters, contract performance records, and related meteorological and environmental data are extracted in chronological order to generate the node's historical time series raw data;

[0064] Based on the original historical time series data of the node, missing value processing and normalization are performed to obtain the historical time series feature data of the node.

[0065] In some embodiments, based on a structured supply chain dataset, historical operational data for various nodes, including main equipment suppliers, warehousing nodes, transportation companies, and terminal substations, is sequentially extracted for corresponding time periods. Specifically, for main equipment suppliers, their contract fulfillment status, order completion rate, batch delivery cycle, and delay records are extracted for each time period. For warehousing nodes, time series information such as inventory fluctuations, material inbound and outbound timeliness, and storage equipment anomalies is collected. For transportation nodes, the focus is on logistics completion rate, number of transportation interruptions, and transportation route risk index within each time period. The system also combines auxiliary features such as contract fulfillment progress and environmental meteorological anomalies for each node to form a complete node historical time series raw data set. This raw time series data first undergoes data cleaning steps such as missing value interpolation, extreme value removal, and noise filtering to improve overall data quality. Subsequently, standardization or normalization methods are applied to different types of feature data, such as Min-Max normalization for numerical data such as delivery cycle and inventory changes, and one-hot encoding for categorical features. Ultimately, historical time series feature data for the nodes is generated with consistent feature scales and is convenient for model input.

[0066] Furthermore, the temporal deep learning model includes a long short-term memory network model.

[0067] It should be noted that the time series deep learning model uses a long short-term memory (LSTM) network. Specifically, the LSTM uses preprocessed historical time series features from each supply chain node as input. At each time step, the model retains and updates key node status information, such as fulfillment rate fluctuations, inventory level changes, logistics disruption history, and environmental risk dynamics. By incorporating forget gates, input gates, and output gates, the LSTM architecture automatically learns the short-term mutations and long-term dependencies of node features over time, effectively identifying complex risk evolution patterns such as cyclical risk increases, sudden fulfillment anomalies, or continuous delivery bottlenecks. During model training, the network parameters are continuously optimized using methods such as cross-entropy loss by comparing historically known risk labels with the LSTM predicted outputs. This allows prediction of risk scores or high-risk probabilities for supply chain nodes within a specified future period. Compared to traditional static rules or simple time window statistical methods, the LSTM-based time series deep learning model has the ability to adaptively model nonlinear dynamic changes, significantly improving the accuracy and real-time performance of supply chain risk warnings.

[0068] Furthermore, the time series deep learning model is constructed by the following steps:

[0069] Based on the node's historical time series feature data and spatial feature representation, combined with known risk event labels, an LSTM training set containing input feature sequences and risk labels is constructed;

[0070] Inputting the LSTM training set into a long short-term memory network model, jointly modeling the node's historical temporal feature data and spatial feature representation, and predicting the risk score or risk probability in a future specified time window;

[0071] Based on historical risk labels and model prediction results, the model parameters are optimized through the back-propagation algorithm to obtain the trained long-short-term memory network model.

[0072] In some embodiments, each node's historical time-series feature data (such as fulfillment rate, inventory fluctuations, and logistics disruption frequency) over the past 24 months is first concatenated with the spatial feature representation output by the graph neural network to form a unified multidimensional input sequence. Combined with the risk event labels annotated during business processes, such as delayed transformer delivery or warehouse explosions, the concatenated feature sequence and the corresponding time-series risk labels are used as training samples to construct an LSTM training set. The LSTM model uses the joint time-series and spatial features of each node as input. The network processes the data at each time step sequentially, leveraging its gating mechanism to automatically memorize long-term risk trends and suppress irrelevant short-term fluctuations, achieving adaptive modeling of supply chain node risk dynamics. The model output is a risk score or high-risk probability for the node within a specified future time window, specifically predicting the probability of a high-risk event due to substandard performance of a main device within the next month. During the training phase, the error is calculated using a cross-entropy loss function by comparing the risk labels with the actual risk labels, and the backpropagation algorithm is used to update the parameters of each LSTM layer. The resulting trained model not only accurately reflects the spatiotemporal coupling of node risks, but also possesses strong trend prediction capabilities and sensitivity to abnormal changes, providing a solid algorithmic foundation for subsequent intelligent identification and early warning of high-risk nodes. This dynamic, deep modeling approach, combining temporal and spatial features, significantly outperforms traditional shallow methods that rely on single time series analysis or independent feature inputs.

[0073] In some specific embodiments, the input data include: the monthly fulfillment rate of supplier A (such as 0.98 in January 2022, 0.92 in February 2022, 0.85 in March 2022, and so on to December 2023), the actual delivery cycle of monthly purchase orders (such as 12 days in June 2023, 15 days in July 2023), warehouse inventory dynamics (such as inventory rate: 87% in August 2023, 68% in September 2023), the number of logistics interruptions (such as 2 logistics abnormalities in April 2022), the number of extreme weather events in the area where the substation is located (such as 2 heavy rain warnings in July 2023), and contract performance compensation events (such as a contract breach compensation in March 2023).

[0074] These node historical time-series feature data are concatenated with the spatial features output by the previous graph neural network (e.g., a node's network risk exposure of 0.34 and a strong dependency weight of 0.28 with upstream key nodes) to form the input training samples for the LSTM. The actual monthly risk labels serve as supervisory signals. For example, in February 2022, due to a sudden drop in Supplier A's fulfillment rate to 0.92 and a logistics disruption, the main transformer delivery was delayed. The system would label that month as high risk (label 1). However, in January and March 2022, months without major anomalies, the labels would be 0. These feature sequences and labels are fed into the LSTM model, which automatically learns the impact of each feature's temporal changes on future risk. After LSTM training is complete, the actual inference phase uses the latest node features from January 2024. For example, if Supplier A's fulfillment rate is 0.88, its lead time is 15 days, its inventory ratio is 60%, it has one logistics disruption, and its network risk exposure is 0.37, the model outputting a high-risk probability of 0.79 for that node in February 2024. If it is higher than the system threshold (such as 0.7), a "high risk" warning will be output for the substation main transformer node, and the report will state that the main driving factors are "decline in supplier A's performance" and "continuous decline in inventory rate."

[0075] Furthermore, the loss function of the long short-term memory network model is as follows:

[0076]

[0077] Among them, L is the loss value; is the number of training samples; N is the number of training samples; C is the total number of risk categories; The true risk label of the jth category for the i-th sample at the future prediction time point; It is the predicted risk probability of the long short-term memory network model for the i-th sample belonging to the j-th category at the future prediction time point.

[0078] It should be noted that the loss function uses cross-entropy loss, which measures the difference between the model's predicted risk probability distribution and the true risk label. During model training, for each supply chain node sample, the LSTM model outputs the predicted probability of belonging to each risk category (e.g., high risk, low risk) at a specified future time point. The true risk label is given in one-hot encoding format. If the model's predicted probability for the high-risk category closely matches the actual risk event, the loss value is low; if the predicted probability deviates from the actual risk category, the loss value is high. The overall loss function value is obtained by averaging the losses across all training samples. The goal of model training is to continuously adjust and optimize the network parameters to minimize the value of this loss function, thereby improving the model's risk prediction accuracy when exposed to new data. This loss function design is not only applicable to binary and multi-class risk prediction scenarios, but also effectively supports the model's fine-grained modeling of future risk probabilities, ensuring the reliability and discriminatory nature of supply chain risk warning results.

[0079] Furthermore, the processing of risk time series features by a multi-layer perceptron includes the following steps:

[0080] Based on the risk time series feature representation output by the time series deep learning model, normalizing the risk time series feature representation to obtain normalized feature data;

[0081] Based on the normalized feature data, the normalized feature data is sequentially input into each hidden layer of the multi-layer perceptron for feature transformation to obtain fused feature data;

[0082] Based on the fused feature data, the fused feature data is input into the output layer of the multi-layer perceptron to generate the risk score data of the target node.

[0083] In some embodiments, the feature vector is first normalized. For example, assume that the time series features output by the LSTM are a set of real number vectors, such as [0.56, -1.04, 2.35, 0.88], where each component represents a comprehensive expression of risk factors such as the abnormal amplitude of the fulfillment rate, the abnormal inventory dynamic score, the fluctuation of the frequency of logistics interruptions, and the fluctuation of contract fulfillment. To prevent the numerical scales between different features from affecting model training, the system uses Z-score normalization. The feature value of each dimension is subtracted from the mean of the corresponding feature in the training set and divided by the standard deviation. The above feature vectors are uniformly mapped to a standard normal distribution with a mean of 0 and a variance of 1, thereby obtaining normalized feature data such as [0.72, -0.89, 1.15, 0.36]. This step ensures that all features are comparable in the subsequent network learning process, improving the model convergence speed and generalization ability. Subsequently, the normalized feature data is sent to each hidden layer in turn as the input of the multi-layer perceptron (MLP). Assume an MLP has two hidden layers. The weight matrix of the first hidden layer maps the 4-dimensional input features into an 8-dimensional latent space. The ReLU activation function is applied to the linear transformation results, performing nonlinear processing to enhance the feature space's expressive power. For example, the output of the first hidden layer might be [0, 0.53, 0, 1.22, 0, 0, 0.31, 0]. This is then passed to the second hidden layer, where the 8-dimensional features are mapped to 16 dimensions and further nonlinear activation is applied. Through layer-by-layer transformations, the MLP automatically learns the high-order interactions between feature components and their contributions to risk discrimination, achieving deep integration of complex risk representations and capturing, for example, the nonlinear discrimination boundary where the combined effects of fulfillment rate and logistics fluctuations significantly increase risk. Ultimately, the high-dimensional output, after fusion of all hidden layer features, is input to the MLP's output layer. This output layer uses a Sigmoid activation function (for binary classification) or a Softmax activation function (for multi-risk classification) to map the fused features into a risk score between 0 and 1. For example, an output of 0.81 indicates an 81% probability of a high-risk event occurring at that node within a specified future timeframe. For multi-classification, the system outputs probability distributions for each risk level, such as [low risk: 0.12, medium risk: 0.34, high risk: 0.54]. The system automatically determines the risk level based on the risk score and preset thresholds, and outputs detailed risk warning information, including the node's unique identifier, risk score, and key drivers. The key innovations of this process lie in the normalization of LSTM dynamic features, the deep extraction and fusion of high-order interactive risk features using a multi-layer perceptron, and a flexible risk identification mechanism based on probabilistic output. This effectively improves the accuracy, discrimination, and interpretability of risk scores for nodes in complex supply chain networks, meeting the power industry's stringent requirements for accurate identification and intelligent early warning of high-risk nodes.

[0084] Furthermore, outputting supply chain risk warning information based on risk score data includes:

[0085] Based on the risk score data of the target node, the risk score data is compared with a preset risk warning threshold. If the risk score data is greater than or equal to the warning threshold, it is determined that the target node has a supply chain risk and corresponding risk warning information is generated;

[0086] If the risk score data is less than the warning threshold, it is determined that there is no supply chain risk at the target node.

[0087] Specifically, the system first receives the target node risk score data output by the MLP. For example, a supply node has a risk score of 0.82. Based on industry operational experience and historical risk tolerance, the system pre-sets a risk warning threshold, such as 0.75. The system then compares the node's risk score of 0.82 with the threshold of 0.75. If the risk score exceeds the threshold, it automatically determines that the node has significant supply chain risk. The system then generates a risk warning message containing details such as the node's unique identifier, predicted risk level, key drivers (such as decreased supplier performance, poor logistics), and recommended intervention measures. This information is then pushed to the operations management platform, prompting relevant personnel to initiate risk response or optimize resource scheduling. Conversely, for nodes with risk scores below the threshold, such as a storage node with a score of 0.48, the system determines that there is currently no significant supply chain risk and does not generate a risk warning, effectively reducing false positives and unnecessary intervention. This effectively ensures the real-time nature of risk warnings and improves the efficiency of intelligent monitoring and management of key links in the power supply chain.

[0088] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A supply chain risk early warning method based on deep learning, characterized by: The following steps are involved: Access supply chain data; Based on supply chain data, a node relationship graph containing supply chain nodes and topological structures is constructed, and the spatial correlation features between nodes are extracted through graph neural networks to obtain spatial feature representation; Obtain historical time series data of supply chain nodes from supply chain data, input spatial feature representation and historical time series feature data of supply chain nodes into the time series deep learning model, model historical risk factors, and output risk time series feature representation; The risk time series features are processed by a multi-layer perceptron to generate risk score data for the target node; Output supply chain risk warning information based on risk scoring data.

2. A supply chain risk early warning method based on deep learning according to claim 1, characterized in that: The acquisition of supply chain data includes: The system collects supplier performance information, material procurement and delivery data, logistics transportation status, equipment operation status, contract execution status, node geographical location and related meteorological environment data of supply chain nodes, and performs data cleaning, normalization and time synchronization processing to generate a structured supply chain data set.

3. A supply chain risk early warning method based on deep learning according to claim 2, characterized in that: The construction of a node relationship graph including supply chain nodes and topology structure comprises the following steps: Based on the structured supply chain data set, determine the material flow relationship between supply chain nodes and generate connection data between nodes; Based on the connection data between nodes, each supply chain node is connected to its directly associated node, and a node relationship graph is constructed that reflects the node attributes and the dependency relationship between nodes.

4. The supply chain risk early warning method based on deep learning according to claim 1 is characterized in that: The graph neural network is constructed through the following steps: Based on the node relationship graph, the structural features of each supply chain node are extracted as node input data, and the connection data between nodes are extracted as edge input data; Based on historically known risk events and supply chain anomalies, a training set containing node input data, edge input data, and corresponding risk labels is constructed. The training set is input into the graph neural network, the information of each node and its associated nodes is aggregated and propagated, a spatial feature representation containing associated nodes is generated, and the network parameters are backpropagated to obtain the trained graph neural network.

5. The supply chain risk early warning method based on deep learning according to claim 1 is characterized in that: Acquiring historical time series data of supply chain nodes from supply chain data includes: Based on the structured supply chain data set, for each supply chain node, supplier status, material procurement and delivery progress, logistics and transportation information, equipment operating parameters, contract performance records, and related meteorological and environmental data are extracted in chronological order to generate the node's historical time series raw data; Based on the original historical time series data of the node, missing value processing and normalization are performed to obtain the historical time series feature data of the node.

6. A supply chain risk early warning method based on deep learning according to claim 5, characterized in that: The temporal deep learning model includes a long short-term memory network model.

7. A supply chain risk early warning method based on deep learning according to claim 6, characterized in that: The time series deep learning model is constructed through the following steps: Based on the node's historical time series feature data and spatial feature representation, combined with known risk event labels, an LSTM training set containing input feature sequences and risk labels is constructed; Inputting the LSTM training set into a long short-term memory network model, jointly modeling the node's historical temporal feature data and spatial feature representation, and predicting the risk score or risk probability in a future specified time window; Based on historical risk labels and model prediction results, the model parameters are optimized through the back-propagation algorithm to obtain the trained long-short-term memory network model.

8. The supply chain risk early warning method based on deep learning according to claim 7 is characterized in that: The loss function of the long short-term memory network model is as follows: Among them, L is the loss value; is the number of training samples; N is the number of training samples; C is the total number of risk categories; The true risk label of the jth category for the i-th sample at the future prediction time point; It is the predicted risk probability of the long short-term memory network model for the i-th sample belonging to the j-th category at the future prediction time point.

9. The supply chain risk early warning method based on deep learning according to claim 1, characterized in that: The processing of risk time series features by a multi-layer perceptron includes the following steps: Based on the risk time series feature representation output by the time series deep learning model, normalizing the risk time series feature representation to obtain normalized feature data; Based on the normalized feature data, the normalized feature data is sequentially input into each hidden layer of the multi-layer perceptron for feature transformation to obtain fused feature data; Based on the fused feature data, the fused feature data is input into the output layer of the multi-layer perceptron to generate the risk score data of the target node.

10. The supply chain risk early warning method based on deep learning according to claim 1, characterized in that: Outputting supply chain risk warning information based on risk scoring data includes: Based on the risk score data of the target node, the risk score data is compared with a preset risk warning threshold. If the risk score data is greater than or equal to the warning threshold, it is determined that the target node has a supply chain risk and corresponding risk warning information is generated; If the risk score data is less than the warning threshold, it is determined that there is no supply chain risk at the target node.

Images