Supply chain demand prediction and risk early warning method and system based on multi-source data

Abstract

The invention relates to a supply chain demand prediction and risk early warning method and system based on multi-source data. The method comprises the steps of collecting multi-source heterogeneous data and constructing a feature vector set; constructing a supply chain network topology structure; inputting the feature vector set and the supply chain network topology structure into a risk propagation model to obtain a risk space-time distribution diagram; inputting the feature vector set and the risk space-time distribution diagram into a demand prediction model to obtain collaborative prediction result information; based on a supply chain network topology structure, generating a node criticality distribution map, comparing, identifying high-risk nodes, obtaining risk level information, and generating graded risk early warning information; according to the method, the feature vector set and the supply chain network topology structure are constructed by integrating the multi-source heterogeneous data, combined optimization of dynamic risk simulation and demand prediction is realized by combining cooperative calculation of the risk propagation model and the demand prediction model, and the method has the advantages of improving prediction accuracy, dynamically identifying high-risk nodes and generating hierarchical early warning information.

Description

Technical Field

[0001] This application relates to the technical field of supply chain management, and in particular to a method and system for supply chain demand forecasting and risk warning based on multi-source data. Background Technology

[0002] In current supply chain management practices, demand forecasting and risk management are typically treated as two separate processes. A typical approach is to first generate demand forecasts based on historical sales data using time series analysis or machine learning methods; then, an independent risk assessment is conducted based on expert experience or monitoring of isolated risk events (such as abnormal weather in a particular location or port congestion). However, this fragmented approach has significant limitations.

[0003] First, demand forecasting often relies too heavily on internal operational data, failing to effectively incorporate and quantify the impact of external environmental dynamics (such as market sentiment fluctuations, policy and regulatory changes, and macroeconomic trends) on demand, leading to a sharp decline in forecast accuracy when facing unforeseen events. Second, in the risk assessment stage, existing methods are mostly limited to qualitative or simple quantitative alerts for single risk events, lacking the ability to dynamically simulate and quantify the propagation path, scope of impact, and ultimate chain effects on the demand side of risks in complex supply chain network structures. Summary of the Invention

[0004] To address the aforementioned shortcomings, this application provides a method and system for supply chain demand forecasting and risk early warning based on multi-source data.

[0005] This application provides a method for supply chain demand forecasting and risk early warning based on multi-source data, including the following steps: Collect multi-source heterogeneous data and construct a feature vector set based on the multi-source heterogeneous data. The multi-source heterogeneous data includes enterprise operation data and environmental situation data. The feature vector set includes statistical feature vectors, text sentiment feature vectors and event feature vectors. A supply chain network topology is constructed based on enterprise operation data. The supply chain network topology includes several nodes and their connection relationships. The nodes include supply chain entities, and the connection relationships include logistics relationships, information flow relationships, and capital flow relationships between supply chain entities. The feature vector set and the supply chain network topology are input into a pre-trained risk propagation model to simulate risk and obtain a spatiotemporal distribution map of risk. The feature vector set and the spatiotemporal distribution map of risk are input into a pre-trained demand prediction model to predict demand and obtain collaborative prediction results. Based on the supply chain network topology, a node criticality distribution map is generated, and the collaborative prediction results are compared with the node criticality distribution map to identify whether there are high-risk nodes. If high-risk nodes exist, obtain the risk level information of the high-risk nodes in the risk spatiotemporal distribution map, and generate graded risk warning information based on the high-risk nodes and risk level information.

[0006] Furthermore, this application also proposes the following steps for collecting multi-source heterogeneous data and constructing a feature vector set based on the multi-source heterogeneous data, wherein the multi-source heterogeneous data includes enterprise operation data and environmental situation data, and the feature vector set includes statistical feature vectors, text sentiment feature vectors, and event feature vectors, including the following steps: Multi-source heterogeneous data are resampled and time-series aligned to unify them to a preset minimum analysis time granularity, thus obtaining time-series unified data; Multimodal feature extraction is performed on time-series unified data to obtain a preliminary feature vector set; The mutual information values ​​between different features in the initial feature vector set are calculated, and feature cross combinations are generated based on the mutual information values. The feature cross-combination is filtered through a preset filtering strategy, and the filtered feature subset is used as the feature vector set.

[0007] Furthermore, this application also proposes the following steps for constructing a supply chain network topology based on enterprise operational data, wherein the supply chain network topology includes several nodes and their connections, the nodes include supply chain entities, and the connections include logistics relationships, information flow relationships, and capital flow relationships between supply chain entities, including the following steps: Extract entity information, order information, and logistics route information from enterprise operation data, and establish multi-level supply relationships between entity information based on order information and logistics route information to form a supply chain entity relationship map; Based on order information and logistics route information, transaction frequency information and logistics timeliness information are extracted, and dynamic weights for time-varying supply relationships are assigned based on transaction frequency information and logistics timeliness information to construct a dynamic supply chain network. Identify key hub nodes in the dynamic supply chain network, and divide the supply chain network into core layer, support layer and edge layer based on key hub nodes to form the supply chain network topology.

[0008] Furthermore, this application also proposes the following steps for extracting entity information, order information, and logistics route information from enterprise operational data, and establishing multi-level supply relationships between entity information based on order information and logistics route information to form a supply chain entity relationship graph: The logistics route information is decomposed into several transportation segment sequence information, and the transportation segment sequence information is clustered into route segments to generate high-frequency route patterns. Identify the transaction entities and transaction relationships in order information, construct a direct supply relationship network between entities based on the transaction entities and transaction relationships, and establish the logistics topology relationship between entity information in combination with high-frequency path patterns; Extract transaction time series features from order information, and identify time dynamic patterns between entity information based on transaction time series features. The time dynamic patterns include periodic supply patterns and sudden supply patterns. By integrating direct supply relationship networks, logistics topology relationships, and time dynamic patterns in multiple dimensions, a multi-level supply relationship network is constructed. Identify key supply paths in a multi-level supply relationship network and generate a supply chain entity relationship graph based on a pre-set weight allocation strategy.

[0009] Furthermore, this application also proposes that the risk propagation model includes a feature fusion layer, a spatiotemporal propagation layer, and a risk field generation layer. The step of inputting the feature vector set and the supply chain network topology into the pre-trained risk propagation model to simulate risk and obtain a risk spatiotemporal distribution map includes the following steps: The feature fusion layer aligns the feature vector set with the supply chain network topology and enhances the information, forming an enhanced supply chain network with multi-dimensional node and connection edge features. The spatiotemporal propagation layer performs risk propagation calculations based on the enhanced supply chain network, and outputs a spatiotemporal risk state sequence in units of nodes; The risk field generation layer maps the spatiotemporal risk state sequence from discrete nodes to a continuous spatiotemporal domain, generating a risk spatiotemporal distribution map.

[0010] Furthermore, this application also proposes that the demand prediction model includes a risk encoding layer, a feature fusion layer, and a prediction generation layer. The step of inputting the feature vector set and the risk spatiotemporal distribution map into the pre-trained demand prediction model to perform demand prediction and obtain collaborative prediction result information includes the following steps: The risk coding layer extracts multi-scale features from the spatiotemporal distribution map of risk and generates risk feature representations with spatiotemporal context information; The feature fusion layer performs cross-modal fusion of risk feature representations and feature vector sets, and dynamically adjusts the contribution weights of different feature vectors in the feature vector set through a gating attention mechanism to generate joint feature representations; The prediction generation layer generates and outputs collaborative prediction results based on joint feature representations.

[0011] Furthermore, this application also proposes a step in which the prediction generation layer generates and outputs collaborative prediction result information based on joint feature representation, including the following steps: Based on joint feature representation, temporal features at different time scales are extracted through parallel temporal convolutional networks. These temporal features include short-term fluctuation features, periodic features, and long-term trend features. The extracted time-series features are input into the corresponding prediction head network to generate multi-scenario prediction information, which includes baseline scenario prediction information, uplink fluctuation scenario prediction information, and downlink fluctuation scenario prediction information. The multi-scenario prediction information is calibrated based on risk feature representation, and the confidence of the multi-scenario prediction information is adjusted by risk attention weight to generate a demand prediction distribution. Based on the demand forecast distribution, key forecast indicators and their confidence intervals are calculated to generate collaborative forecast results.

[0012] Furthermore, this application also proposes the step of inputting the extracted temporal features into the corresponding prediction head network to generate multi-scenario prediction information, wherein the multi-scenario prediction information includes baseline scenario prediction information, uplink fluctuation scenario prediction information, and downlink fluctuation scenario prediction information, comprising the following steps: Historical demand data is acquired, and multi-scenario condition vectors are generated based on the quantile statistics of the historical demand data. The multi-scenario condition vectors include a baseline scenario condition vector, an upward fluctuation scenario condition vector, and a downward fluctuation scenario condition vector. The extracted temporal features are concatenated with the corresponding multi-scenario conditional vectors to generate scenario-specific conditional feature representations. The generated conditional feature representations are input into the corresponding prediction head networks to generate baseline scenario prediction information, uplink fluctuation scenario prediction information, and downlink fluctuation scenario prediction information, respectively. The prediction head network includes a baseline prediction head, an uplink prediction head, and a downlink prediction head.

[0013] Furthermore, this application also proposes the following steps: if high-risk nodes exist, obtain the risk level information of the high-risk nodes in the spatiotemporal distribution map of the risk, and generate graded risk warning information based on the high-risk nodes and the risk level information, including the following steps: Based on the risk level information of high-risk nodes and the demand fluctuation range in the collaborative prediction results, a comprehensive risk score is calculated, and a comprehensive risk distribution is generated based on all comprehensive risk scores. Risk warning levels are divided based on comprehensive risk distribution, and risk transmission simulations are performed on each high-risk node in turn based on the supply chain network topology to determine the impact range distribution, which includes the impact range of upstream supply disruptions and the impact range of downstream delivery delays. The risk warning levels and their affected areas are integrated into a tiered risk warning information system.

[0014] Furthermore, this application also proposes a supply chain demand forecasting and risk early warning system based on multi-source data, comprising: The data acquisition module is used to collect multi-source heterogeneous data and construct a feature vector set based on the multi-source heterogeneous data. The multi-source heterogeneous data includes enterprise operation data and environmental situation data. The feature vector set includes statistical feature vectors, text sentiment feature vectors and event feature vectors. The topology construction module is used to construct a supply chain network topology based on enterprise operation data. The supply chain network topology includes several nodes and their connection relationships. The nodes include supply chain entities, and the connection relationships include logistics relationships, information flow relationships, and capital flow relationships between supply chain entities. The risk simulation module is used to input the feature vector set and the supply chain network topology into a pre-trained risk propagation model to simulate risk and obtain a risk spatiotemporal distribution map. The demand forecasting module is used to input the feature vector set and the risk spatiotemporal distribution map into the pre-trained demand forecasting model to perform demand forecasting and obtain collaborative forecasting results. The risk node identification module is used to generate a node criticality distribution map based on the supply chain network topology, and compare the collaborative prediction results with the node criticality distribution map to identify whether there are high-risk nodes. The risk information generation module is used to obtain the risk level information of high-risk nodes in the spatiotemporal distribution map of risk if high-risk nodes exist, and to generate graded risk warning information based on the high-risk nodes and risk level information.

[0015] As can be seen from the above, the supply chain demand forecasting and risk early warning method and system provided in this application integrates multi-source heterogeneous data to construct a feature vector set and establish a supply chain network topology. By combining the collaborative analysis of risk propagation model and demand forecasting model, it dynamically simulates the spatiotemporal propagation path of risk in the supply chain network, accurately identifies the impact of key node risks on demand fluctuations, and generates hierarchical risk early warning information with spatiotemporal dimensions. It has the advantages of improving the accuracy of supply chain demand forecasting, enhancing the dynamism of risk assessment and the accuracy of early warning. Attached Figure Description

[0016] Figure 1 This is a flowchart of an embodiment of a supply chain demand forecasting and risk warning method based on multi-source data according to this application; Figure 2 This is a flowchart of step S10 in an embodiment of a supply chain demand forecasting and risk warning method based on multi-source data in this application; Figure 3 This is a flowchart of step S20 in an embodiment of a supply chain demand forecasting and risk warning method based on multi-source data in this application. Detailed Implementation

[0017] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following drawings indicate similar items; therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0018] The following is in conjunction with the appendix Figures 1-3 This application will be described in further detail.

[0019] In one embodiment, such as Figure 1 As shown, this application discloses a method for supply chain demand forecasting and risk early warning based on multi-source data, which specifically includes the following steps: S10: Collect multi-source heterogeneous data and construct a feature vector set based on the multi-source heterogeneous data. The multi-source heterogeneous data includes enterprise operation data and environmental situation data. The feature vector set includes statistical feature vectors, text sentiment feature vectors and event feature vectors. In this embodiment, multi-source heterogeneous data refers to a collection of various types of data with different sources and structures in the supply chain management environment. This includes enterprise operational data and environmental situation data, specifically implemented using multi-dimensional data sources such as order records, logistics trajectories, and macroeconomic indicators, used to capture internal and external dynamics of the supply chain. Feature vector sets refer to a standardized set of numerical representations extracted from multi-source heterogeneous data through feature engineering techniques. Feature vector sets transform heterogeneous data from different sources into a unified numerical vector form, providing standardized input for subsequent machine learning models. Each feature vector represents a multi-dimensional feature snapshot of a specific point in time or entity. Enterprise operational data refers to a set of structured business data generated by an enterprise in its daily operations, including supply chain master data, transaction data, production data, logistics data, quality data, and financial data. Environmental situation data refers to a set of multi-dimensional information in the external environment that may affect supply chain operations, including market data, geographical data, industry data, and risk data. Statistical feature vectors refer to quantitative feature representations generated through statistical calculations of numerical time series data, including time-series statistical features (mean, variance, skewness, kurtosis). Statistical feature vectors include trend characteristics (moving average, linear regression slope), periodic characteristics (autocorrelation coefficient, seasonal index), distribution characteristics (quantiles, extreme value statistics), and fluctuation characteristics (coefficient of variation, volatility). These statistical feature vectors can be calculated by sliding through a time window to capture the basic statistical regularities and evolutionary trends of the data. Text sentiment feature vectors refer to the quantitative representation of sentiment tendencies extracted from unstructured text using natural language processing techniques. These include sentiment polarity features (positive / negative sentiment intensity scores), topic distribution features (topic probability distribution generated by LDA topic models), entity sentiment features (sentiment tendencies for specific entities), sentiment fluctuation features (trends and amplitudes of sentiment value changes), and crisis index features (risk warning indicators built based on a crisis lexicon). Event feature vectors refer to the feature representation generated by structured encoding of discrete event information. These include event type encoding (unique hot encoding of events such as supply disruptions and demand surges), event intensity indicators (quantitative assessment of impact range and duration), event correlation (the correlation strength between the event and supply chain nodes), event propagation chain (the propagation path characteristics of the event in the supply chain), and timeliness features (the time difference between the event occurrence time and the prediction time).

[0020] S20: Construct a supply chain network topology based on enterprise operation data. The supply chain network topology includes several nodes and their connection relationships. The nodes include supply chain entities, and the connection relationships include logistics relationships, information flow relationships, and capital flow relationships between supply chain entities. In this embodiment, the supply chain network topology is constructed by extracting information such as logistics paths, transaction frequencies, information transmission, interaction patterns of sharing and collaboration, and fund transfers between entities. For example, suppliers, manufacturers, and distributors are mapped as network nodes, and transportation routes and transaction frequencies are quantified as edge weights. The information flow relationship does not describe the transfer of physical goods or funds, but rather depicts the flow path and dependency relationship of intangible information such as knowledge, instructions, and status updates.

[0021] S30: Input the feature vector set and the supply chain network topology into the pre-trained risk propagation model to simulate risk and obtain a risk spatiotemporal distribution map; In this embodiment, the risk propagation model refers to a computational model built based on graph neural networks or infectious disease models, used to simulate the diffusion process of risk in the supply chain network. It adopts a structure combining graph neural networks and spatiotemporal convolution, for example, using graph attention mechanisms to simulate the diffusion process of risk along the supply chain path. The risk spatiotemporal distribution map refers to a two-dimensional or three-dimensional data structure that graphically displays the spatial distribution and temporal evolution of risk in the supply chain network. The change in risk intensity can be represented by changing the color depth or size of different nodes in the graph, and the direction and value of the edges represent the direction and probability of risk propagation. At the same time, it supports dynamic display of multiple time steps, intuitively presenting the spatial pattern of risk at a specific moment and the trend of risk changing over time.

[0022] S40: Input the feature vector set and the risk spatiotemporal distribution map into the pre-trained demand prediction model to predict demand and obtain collaborative prediction results. In this embodiment, the demand forecasting model refers to a computational model that uses machine learning or deep learning algorithms to quantitatively estimate the demand for products or services within a specific future timeframe based on historical data and multi-source information. In this embodiment, the core difference between the demand forecasting model and conventional models lies in quantifying the impact of risks and incorporating it into the forecasting process, thus achieving a leap from traditional statistical forecasting to intelligent forecasting based on risk perception. The collaborative forecasting result information refers to the enhanced forecasting result generated by fusing traditional demand forecasting with risk information. This collaborative forecasting result information not only includes the baseline demand forecast value but also reflects the forecast adjustments under different risk scenarios, including point forecasts, interval forecasts, and conditional forecast distributions corresponding to different risk levels, providing decision-makers with multifaceted demand prospects.

[0023] S50: Based on the supply chain network topology, generate a node criticality distribution map, and compare the collaborative prediction results with the node criticality distribution map to identify whether there are high-risk nodes; In this embodiment, the node criticality distribution map refers to the result of quantifying the importance of nodes in the supply chain network based on graph theory algorithms. By calculating topological indicators such as degree centrality, betweenness centrality, and proximity centrality of each node in the network, the relative importance of nodes in terms of network connectivity, information flow, and control is evaluated. It can be visualized in the form of heatmaps or hierarchical graphs. The node criticality distribution can be calculated using centrality algorithms, such as combining eigenvector centrality and betweenness centrality to evaluate the strategic position of nodes in the supply chain.

[0024] Furthermore, specifically, firstly, based on the supply chain network topology, graph theory algorithms are used to calculate the network centrality index of each node to generate a node criticality distribution map. Specifically, by calculating multi-dimensional indicators such as degree centrality, betweenness centrality, and proximity centrality, the structural importance of each node in the network is quantified, forming a probability distribution map reflecting the node criticality. Subsequently, the collaborative prediction results are compared and analyzed with the node criticality distribution map in a multi-dimensional manner. This analysis process includes: extracting the demand fluctuation characteristics and risk exposure levels corresponding to each node from the collaborative prediction results, and obtaining the structural importance score of the corresponding node from the node criticality distribution map; and calculating the criticality-risk coupling matrix for each node. The risk criticality score of a node is a weighted function of the node's structural importance and risk exposure level. During the identification and judgment phase, a dynamic threshold mechanism is set: when a node's risk criticality score exceeds the quantile threshold based on historical data statistics, and its risk exposure level consistently exceeds the business tolerance threshold, the node is marked as a high-risk node. Further verification of the risk transmission path of high-risk nodes in the risk spatiotemporal distribution map confirms the spatiotemporal persistence of their risk impact. Finally, a list of high-risk nodes and their risk profiles are output, including structured data such as node identifier, risk level, criticality score, risk exposure value, and impact scope assessment, completing the accurate identification and classification of high-risk nodes.

[0025] For example, background: the supply chain of a car manufacturer (Company M).

[0026] Step 1: Generate a node criticality distribution map; By analyzing the supply chain network, it was calculated that: Supplier B (provides dedicated chips): has extremely high betweenness centrality because Supplier B is the sole supplier of chips required for multiple vehicle platform models of Company M, and its supply path is the "must-pass" for many vehicle production lines.

[0027] Supplier A (provides general-purpose screws): low degree centrality and low betweenness centrality because there are multiple alternative suppliers and it is not on the critical path.

[0028] Conclusion: The node criticality distribution map shows that supplier B is a critical hub node.

[0029] Step 2: Obtain collaborative prediction results information; After integrating risk information, the demand forecasting model outputs a forecast: Forecasts for the region where Supplier B is located indicate that the risk of supply chain disruptions in the region is high over the next three months due to geopolitical tensions (environmental situation data). Therefore, the downside scenario forecast for Supplier B's component supply shows that demand may plummet by 70%.

[0030] Forecasts for supplier A indicate that everything is normal and demand is stable.

[0031] Step 3: Comparison and identification; Compare the results of step 1 and step 2: Supplier B: Meeting both "high structural criticality" and "high negative risk exposure" criteria, it was identified as a high-risk node.

[0032] Supplier A: Although the forecast is normal, the structural criticality is low, and it is easily replaceable even if there is an outage. It has not been identified as a high-risk node.

[0033] Step 4: Generate an alert (subsequent steps); The system will generate the highest level of warning for supplier B, a high-risk node, and will indicate: "The sole chip supplier B faces an extremely high risk of disruption due to the regional situation, which may lead to a complete production halt. It is recommended to immediately initiate an evaluation of alternative solutions."

[0034] S60: If high-risk nodes exist, obtain the risk level information of high-risk nodes in the risk spatiotemporal distribution map, and generate graded risk warning information based on the high-risk nodes and risk level information.

[0035] In this embodiment, a high-risk node refers to an entity node in the supply chain network topology that simultaneously possesses high structural criticality and high business impact potential. This is typically not a single-dimensional judgment, but rather a comprehensive identification result based on network structural attributes and dynamic risk prediction. Risk level information is a dynamic data set bound to high-risk nodes, quantifying the probability of a specific risk event occurring and its expected impact. This risk level information is obtained from a risk spatiotemporal distribution map, providing a specific threat metric for high-risk nodes. Hierarchical risk warning information refers to tiered alarm information generated based on the comprehensive risk assessment results. It is divided into different levels according to the severity of the risk, with each level corresponding to a specific handling plan and response timeframe. The content of the hierarchical risk warning information may include structured data such as risk type, affected nodes, expected occurrence time, potential business impact, and recommended countermeasures.

[0036] For example, the identification of high-risk nodes can rely on the simultaneous fulfillment of the following two conditions: 1. Structural criticality condition: A node with a high centrality index in the node criticality distribution map indicates that the node occupies a pivotal position in the network topology. For example: High degree centrality: It connects a large number of upstream and downstream nodes and is a key hub for logistics and information flow; High betweenness centrality: Many shortest paths pass through this node, and its failure will sever the network and have a wide impact; High proximity centrality: It can quickly reach other parts of the network, and its influence spreads rapidly; 2. Risk Exposure Condition: If the node shows significant negative fluctuations or anomalies in the collaborative forecast results, it indicates that the node's demand or supply is facing an uncertain shock. For example, if a single supplier that provides critical components to multiple downstream factories (meeting the high structural criticality requirement) is located in an area that is facing earthquake risk according to the risk spatiotemporal distribution map, and the collaborative prediction shows that its supply may be interrupted (meeting the high risk exposure requirement), then this supplier node is identified as a high-risk node.

[0037] For example, the risk level information includes: Risk intensity: The severity of a risk event that is about to occur or is occurring at this node, usually a quantitative value (e.g., earthquake magnitude, port congestion index, supplier financial risk score). Probability of occurrence: The estimated likelihood that the risk event will actually occur and have an impact, usually expressed as a probability value (0%-100%); Scope of impact: The estimated impact of the risk event on the node's own operations (e.g., percentage decrease in capacity, number of days of delivery delay). Time dimension: Expected occurrence time, duration, and decay period of the risk event; Continuing from the previous example: For the supplier node identified as high-risk in the previous example, its risk level information might be as follows: Risk intensity: 7.0 magnitude earthquake; Probability of occurrence: 30%; Expected impact: Complete interruption of production capacity; Duration: 7-10 days.

[0038] Specifically, the process begins by aligning the time dimension of multi-source heterogeneous data through data cleaning, constructing a fused feature set that includes statistical, sentiment, and event characteristics. Then, a dynamically weighted supply chain network is established based on order data and logistics routes, identifying core hub nodes to form a hierarchical topology. The feature set and topology are input into a risk propagation model, which uses multi-layer graph convolution and other methods to capture the transmission patterns of risk between nodes, outputting a risk field distribution that varies with time and space. The demand forecasting model simultaneously receives the risk field and feature vector set, using cross-modal attention mechanisms to adjust the weights of external risks on demand, generating forecast results that consider the risk transmission effect. Finally, by comparing the forecast results with the node criticality distribution, high-risk nodes that may trigger systemic risks are identified, and tiered early warning information is generated based on their risk propagation range.

[0039] Compared with existing technologies, traditional methods, when using linear regression to analyze demand, do not consider the nonlinear impact of supplier node interruptions on the forecast results. This solution, by constructing a dynamic supply chain network, can quantify the transmission effect of single node risk on overall demand forecasting. Existing risk assessment tools mostly use static weight analysis, while this application, through spatiotemporal risk field modeling, can dynamically reflect the diffusion path and intensity changes of risk in the supply chain network. Conventional early warning systems only focus on the risk value of the node itself, while this method, by combining the criticality of the node and the scope of risk transmission, can identify the key risk points that have the greatest impact on supply chain stability.

[0040] Through the above technical solutions, this application effectively solves the problem of the disconnect between external environmental factors and internal operational data, enabling multi-dimensional attribution analysis of demand fluctuations; the established dynamic risk propagation model can predict in advance the impact of regional events on different levels of the supply chain, such as the probability of delays in port congestion risk being transmitted to manufacturing nodes through logistics topology; the collaborative forecasting mechanism enables demand forecasting results to include risk transmission effects, helping enterprises to accurately adjust inventory strategies in events such as supplier shutdowns; at the same time, by identifying key risk nodes and conducting tiered early warnings, risk monitoring in the supply chain can be strengthened in a targeted manner, reducing the probability and impact of local risks evolving into systemic crises.

[0041] In one embodiment, such as Figure 2 As shown, step S10 includes: S11: Resample and time-series align multi-source heterogeneous data to unify them to the preset minimum analysis time granularity, thereby obtaining time-series unified data; In this embodiment, resampling refers to adjusting raw data from different collection frequencies to a uniform time interval. This can be achieved using linear interpolation or nearest neighbor interpolation algorithms to eliminate time scale differences between different data sources. Time alignment refers to mapping timestamps from multiple data sources to the same time base. This can be achieved using sliding window matching or dynamic time warping algorithms to ensure data comparability in the time dimension. The minimum analytical time granularity refers to the smallest time unit used in data analysis, determined by business needs and data availability. This minimum analytical time granularity determines the finest time scale pattern that the model can capture. For example, if "day" is chosen as the minimum granularity, the model can identify daily fluctuations but cannot capture intraday patterns. Therefore, the granularity selection needs to strike a balance between computational efficiency and pattern capture accuracy.

[0042] S12: Perform multimodal feature extraction on the time-series unified data to obtain a preliminary feature vector set; In this embodiment, multimodal feature extraction refers to extracting statistical features, text sentiment features, and event features from structured and unstructured data, respectively. Specifically, convolutional neural networks can be used to extract text features, and sliding window statistical methods can be used to extract numerical features.

[0043] S13: Calculate the mutual information values ​​between different features in the preliminary feature vector set, and generate feature cross combinations based on the mutual information values; In this embodiment, mutual information value calculation refers to quantifying the nonlinear correlation between different features. Specifically, kernel density estimation or entropy-based calculation methods can be used to identify feature combinations with synergistic effects. Feature cross combination refers to combining features whose mutual information values ​​exceed a preset threshold. Specifically, Cartesian product or tensor concatenation methods can be used to generate higher-order features.

[0044] S14: Select the cross-combination of features using a preset selection strategy, and use the selected feature subset as the feature vector set.

[0045] In this embodiment, the filtering strategy refers to filtering based on the information gain or redundancy of feature combinations. Specifically, recursive feature elimination or regularization-based feature selection algorithms can be used to retain the most discriminative feature subset.

[0046] Specifically, multi-source heterogeneous data, such as time-series data like order volume and inventory levels from enterprise operational data, and unstructured data from environmental situation data, are first processed through resampling and time-series alignment to unify them to a minimum analytical time granularity, such as hourly or daily. Then, multimodal feature extraction methods are used to extract statistical features such as mean and variance from structured data, textual features such as sentiment polarity and keyword frequency from text data, and event features such as event type and scope of impact from event data, forming a preliminary feature vector set. Further, mutual information values ​​are used to identify the correlation between different features; for example, a strong correlation is found between order volume fluctuations and changes in public opinion sentiment polarity, thus generating a cross-combination feature of order volume and sentiment polarity. Finally, redundant features are removed through a screening strategy, such as eliminating feature combinations with a correlation below a preset threshold with the target variable, retaining the effective feature subset as the final feature vector set input to the prediction model.

[0047] Through the above technical solutions, this application can effectively integrate the temporal and semantic information of multi-source heterogeneous data, eliminate noise interference caused by differences in data collection frequency, and enhance the model's ability to capture complex relationships through feature cross-combination. For example, in the scenario of supply chain demand forecasting, after cross-combining order volume statistical features and public opinion sentiment features, the model can more accurately identify demand fluctuations caused by sudden events, thereby improving the robustness of the forecast results. In addition, the application of screening strategies reduces the interference of irrelevant features on the model, reduces computational complexity, and enables the forecasting model to have higher operating efficiency while ensuring accuracy.

[0048] In one embodiment, such as Figure 3 As shown, step S20 includes: S21: Extract entity information, order information, and logistics route information from enterprise operation data, and establish multi-level supply relationships between entity information based on order information and logistics route information to form a supply chain entity relationship graph; In this embodiment, the multi-level supply relationship refers to a composite relationship network that includes direct supply, logistics topology, and time dynamic patterns, established through order transaction entities, logistics paths, and time series characteristics. Specifically, it can be implemented by high-frequency path pattern clustering and transaction time series feature identification to reflect the interaction characteristics of different dimensions between supply chain entities. The supply chain entity relationship graph is a data model that uses a graph structure to model and express the knowledge of the supply chain system. Through the topological structure of nodes and edges, it integrates the scattered supply chain information into a machine-understandable and computable knowledge base.

[0049] S22: Extract transaction frequency information and logistics timeliness information based on order information and logistics route information, and assign dynamic weights based on time changes to multi-level supply relationships based on transaction frequency information and logistics timeliness information to construct a dynamic supply chain network; In this embodiment, transaction frequency information is a dynamic indicator that quantifies the activity of business interactions between entities. It does not refer to a single number of transactions, but rather a processed and standardized feature used to measure the strength of a relationship. Logistics timeliness information is an indicator that quantifies the efficiency and stability of logistics fulfillment between entities. It measures the performance characteristics of the time required to "deliver goods from point A to point B". Logistics timeliness information includes: average logistics time: the average time from "shipment" to "receipt" for historical orders; logistics time standard deviation / coefficient of variation: measures the volatility or reliability of logistics time. The smaller the standard deviation, the more stable the fulfillment; on-time delivery rate: the percentage of orders delivered within the promised time. Time-varying dynamic weight refers to the numerical attribute assigned to the connecting edges in the supply chain network that changes over time. It is used to quantify the strength or activity of the relationship at a specific point in time. The time-varying dynamic weight is not fixed, but is calculated based on dynamic indicators such as recent transaction frequency (e.g., weekly order count) and logistics timeliness (e.g., average delivery delay). For example, supply relationships with frequent recent transactions and on-time delivery will be given higher weights, and vice versa.

[0050] S23: Identify key hub nodes in the dynamic supply chain network, and divide the supply chain network into core layer, support layer and edge layer based on key hub nodes to form the supply chain network topology.

[0051] In this embodiment, key hub nodes refer to network center nodes selected based on transaction volume and connection density. Specifically, they can be identified using a betweenness centrality algorithm combined with transaction volume thresholds, and are used to divide the network hierarchy. The core layer, support layer, and edge layer refer to the three layers formed after strategically dividing the supply chain network based on node criticality. The core layer consists of a few key hub nodes, which have extremely high centrality and are the "anchors" of the network. Their failure will cause a large-scale network paralysis, requiring the highest level of risk monitoring and protection measures. The support layer consists of nodes directly connected to the core layer nodes and providing important support. Although these nodes are not global hubs, they are crucial to maintaining the normal operation of the core layer. The edge layer consists of nodes with relatively low importance and fewer connections in the network. The failure of these nodes has a limited impact on the overall stability of the network, and the risk is relatively controllable.

[0052] Specifically, after structured processing, entity information, order information, and logistics route information are clustered using high-frequency path patterns to establish logistics topology relationships. Combined with transaction time series characteristics, periodic or sudden supply patterns are identified to form a multi-level supply relationship network. Based on dynamic weight calculation of transaction frequency and logistics timeliness, the static supply relationship is transformed into a dynamic network that changes over time. Through the identification of key hub nodes, the dynamic network is divided into a core layer, a support layer, and an edge layer. The core layer contains entities with high-frequency transactions and dense connections, the support layer contains less important entities, and the edge layer contains entities with occasional transactions. This hierarchical structure allows risk propagation analysis to focus on the critical paths of the core layer.

[0053] Through the above technical solutions, this application effectively solves the problems of missing dynamic characteristics and ambiguous hierarchical structure in traditional supply chain network modeling; the introduction of dynamic weights enables the network model to adapt to fluctuations in transaction frequency and changes in logistics timeliness, such as automatically increasing the weight of high-frequency transaction paths during promotional seasons; the construction of a hierarchical topology structure enables risk transmission simulation to prioritize core layer nodes, such as quickly locating the upstream and downstream critical paths affected when a core supplier experiences a risk; the integrated modeling of multi-level supply relationships improves the completeness of network structure representation, such as simultaneously considering the combined impact of periodic purchase orders and sudden logistics disruptions on network stability.

[0054] In one embodiment, step S21 includes: S211: Decompose the logistics route information into several transportation segment sequence information, and perform route segment clustering on the transportation segment sequence information to generate high-frequency route patterns; In this embodiment, the transport segment sequence information refers to the segmented path data formed by breaking down the complete logistics path according to transport nodes. Specifically, it can be implemented using a logistics node coordinate analysis algorithm to reflect the actual flow trajectory of goods in the supply chain. Path segment clustering refers to the technique of pattern mining of transport segment sequences using unsupervised machine learning algorithms. By calculating the similarity of transport segments in spatiotemporal features, frequently occurring continuous transport segment sequences are identified as high-frequency path patterns. Path segment clustering usually adopts a density clustering algorithm based on dynamic time warping, which can effectively handle path sequences of different lengths and temporal deformations. High-frequency path patterns refer to the combinations of transport paths with a recurrence probability exceeding a preset threshold extracted through cluster analysis. Specifically, it can be implemented using a density-based spatial clustering algorithm to identify core logistics channels in the supply chain.

[0055] S212: Identify the transaction entities and transaction relationships in the order information, construct a direct supply relationship network between entities based on the transaction entities and transaction relationships, and establish the logistics topology relationship between entity information in combination with high-frequency path patterns; In this embodiment, the direct supply relationship network refers to the supplier-customer relationship topology directly formed by the two parties in an order transaction. Specifically, it can be implemented using the node-edge modeling method of a graph database to represent the explicit transaction relationship between enterprises. The logistics topology refers to the spatial connection relationship established by introducing high-frequency path patterns on the basis of the direct supply relationship network. This logistics topology not only reflects the business connections between entities, but also depicts the actual spatial connection path formed between entities through the logistics network, providing a physical path basis for risk propagation.

[0056] S213: Extract transaction time series features from order information, and identify time dynamic patterns between entity information based on transaction time series features, wherein the time dynamic patterns include periodic supply patterns and sudden supply patterns; In this embodiment, transaction time series features refer to time-dimensional statistics extracted from order transaction records, specifically implemented using a sliding window statistical method, to capture the periodic or sudden patterns of supply behavior; periodic supply patterns refer to regular fluctuation patterns identified from transaction time series through spectrum analysis. Periodic supply patterns reflect regular demand changes based on time scales such as seasons, months, and weeks, such as quarterly promotions and monthly replenishment cycles, and are the basis for predicting stable demand; sudden supply patterns refer to unexpected fluctuation patterns identified from transaction time series through anomaly detection algorithms. Sudden supply patterns usually correspond to demand surges or supply interruptions caused by sudden events, such as emergency orders and sudden stockouts, and are important indicators for risk warning.

[0057] S214: Integrate direct supply relationship networks, logistics topology relationships, and time dynamic patterns in multiple dimensions to construct a multi-level supply relationship network; In this embodiment, multi-dimensional fusion refers to the technology of integrating network relationships from different sources and types. It typically employs the message passing mechanism of graph neural networks to allow direct supply relationships, logistics topology relationships, and time pattern features to interact in a unified graph structure, generating node and edge representations rich in multi-dimensional features. Multi-level supply relationship network refers to a composite network structure that integrates direct transactions, logistics paths, and time dynamic features. Specifically, it can be implemented using multi-graph fusion algorithms to comprehensively characterize the interaction relationships between supply chain entities.

[0058] S215: Identify key supply paths in a multi-level supply relationship network and generate a supply chain entity relationship graph based on a pre-set weight allocation strategy.

[0059] In this embodiment, the critical supply path refers to the important material transmission channel identified based on network traffic analysis and node centrality indicators. Specifically, it can be implemented using the maximum flow minimum cut algorithm to locate the vulnerable links in the supply chain network. The weight allocation strategy refers to the relationship strength quantification method based on multi-objective optimization theory. Typically, the weight allocation strategy comprehensively considers multiple dimensions such as transaction frequency, logistics timeliness, and cooperation stability, and calculates the weights of each dimension through the entropy weight method or the analytic hierarchy process.

[0060] Specifically, after logistics route information is decomposed into a sequence of transportation segments, high-frequency transportation patterns can be identified through route segment clustering. For example, cross-border logistics can be decomposed into port transshipment segments, land transportation segments, and customs clearance segments, and the frequency of occurrence of each segment can be counted. After the order transaction data is parsed for entity relationships, a direct supply relationship network is constructed, for example, using the purchaser and supplier as nodes and the transaction amount as the edge weight. In the process of extracting transaction time series features, Fourier transform is used to detect periodic supply patterns, such as quarterly procurement peaks, while anomaly detection algorithms are used to identify sudden supply events. In the multi-dimensional fusion stage, the direct transaction network and logistics topology are spatially superimposed, and time dynamic patterns are introduced as edge attributes, such as marking the average delivery cycle on the transportation route. In the process of identifying key supply routes, the transportation capacity, node connectivity, and time sensitivity indicators of the route are combined, for example, screening out transportation routes that simultaneously carry more than 60% of the cargo volume and connect core hub nodes.

[0061] Through the above technical solutions, this application can accurately capture the spatial interaction characteristics and temporal dynamic patterns between supply chain entities, solving the problems of simple network structure and lack of dynamic characteristics in traditional methods; by identifying high-frequency path patterns, it can effectively distinguish between regular logistics channels and temporary alternative routes, improving the risk resistance of the supply chain network; combined with the multi-dimensional fusion of temporal dynamic patterns, the constructed entity relationship map can simultaneously reflect transaction relationships, logistics paths and temporal fluctuation characteristics, providing an accurate topological foundation for subsequent risk propagation simulation; the key supply path identification mechanism helps to quickly locate network bottlenecks and provides data support for hierarchical early warning.

[0062] In one embodiment, the risk propagation model includes a feature fusion layer, a spatiotemporal propagation layer, and a risk field generation layer. Step S30 includes: S31: The feature fusion layer aligns the feature vector set with the supply chain network topology and enhances the information to form an enhanced supply chain network with multi-dimensional node and connection edge features. In this embodiment, the feature fusion layer refers to the module that integrates feature vectors from different data sources with the network topology. Specifically, it can use cross-modal attention mechanisms or graph neural networks for feature alignment, and achieve information fusion by enhancing the multidimensional attributes of nodes and connecting edges. Feature alignment refers to the technical process of mapping features from different sources and scales to a unified vector space. This is achieved by normalizing numerical features, embedding categorical features, and semantically encoding text features to ensure that all features have the same dimension and comparability, providing a foundation for subsequent feature fusion. Information enhancement refers to the technique of enriching the original feature representation by introducing external knowledge or feature interactions, including feature cross (generating feature interaction terms), attention mechanisms (highlighting important features), and graph structure information injection (integrating network neighbor features), aiming to improve the representational ability of features. The enhanced supply chain network refers to a rich network representation on the basis of the original supply chain network topology, with each node and edge having a multidimensional feature vector added. The enhanced supply chain network not only includes topological connections but also integrates operational features, environmental features, and risk features, providing a multi-level and multi-faceted data foundation for accurate risk propagation simulation.

[0063] S32: The spatiotemporal propagation layer performs risk propagation calculations based on the enhanced supply chain network, and outputs a spatiotemporal risk state sequence in units of nodes; In this embodiment, the spatiotemporal propagation layer refers to the computational unit that simulates the dynamic diffusion of risks in the supply chain network. Specifically, it can adopt a spatiotemporal graph convolutional network or a propagation dynamics model, and combine the connection relationship between nodes with the time series changes to perform iterative calculation of risk states. The spatiotemporal risk state sequence refers to the multidimensional time series data formed by arranging the risk state probabilities of each node at different time points in chronological order. The state at each time step includes quantitative indicators of multiple dimensions such as the probability of risk occurrence, risk intensity estimation, and impact range assessment, which fully describes the dynamic evolution process of the node's risk state.

[0064] S33: The risk field generation layer maps the spatiotemporal risk state sequence from discrete nodes to a continuous spatiotemporal domain, generating a risk spatiotemporal distribution map.

[0065] In this embodiment, the risk field generation layer refers to a visualization module that transforms the risk status of discrete nodes into a continuous spatial distribution. Specifically, it can use spatial interpolation algorithms or kernel density estimation methods to map node risk values ​​into a two-dimensional or three-dimensional risk field covering the geographical scope of the supply chain network. Discrete nodes refer to individual entities that are considered independent, distinguishable, and discontinuous in the supply chain network topology. The continuous spatiotemporal domain refers to a theoretical framework that transforms discrete node locations and risk observations into continuous function representations using mathematical methods. By using spatial location and time points as independent variables and risk intensity as a dependent variable, an interpolation estimation model for risk at any spatiotemporal point is established to achieve full coverage.

[0066] Specifically, the feature fusion layer first performs dimensional unification and standardization on the feature vector set, for example, converting statistical feature vectors and text sentiment feature vectors into a vector space of the same dimension through an embedding layer; then, it uses the node attributes and edge weights in the supply chain network topology as graph structure feature inputs, and dynamically adjusts the correlation strength between different features between nodes using graph attention mechanisms and other methods; after receiving the enhanced supply chain network, the spatiotemporal propagation layer performs risk propagation simulation according to a preset time step, for example, calculating the risk transmission coefficient between nodes in each time step, and superimposing external influencing factors from the environmental situation data; the risk field generation layer expands the risk values ​​of discrete nodes into a continuous distribution map covering the entire supply chain activity area through spatial coordinate transformation and interpolation algorithms, for example, using the Kriging interpolation method to generate a risk heat map with spatial continuity.

[0067] Through the above technical solutions, this application can accurately capture the spatiotemporal evolution of risks in the supply chain network, providing dynamic risk environment input for subsequent demand forecasting; through the generation of continuous risk fields, high-risk areas and their diffusion trends can be intuitively identified, supporting targeted early warning and emergency resource scheduling decisions for key nodes in the supply chain.

[0068] In one embodiment, the demand prediction model includes a risk encoding layer, a feature fusion layer, and a prediction generation layer. Step S40 includes: S41: The risk coding layer performs multi-scale feature extraction on the spatiotemporal distribution map of risk and generates risk feature representations with spatiotemporal context information; In this embodiment, the risk coding layer refers to a neural network structure that performs multi-scale feature extraction on the spatiotemporal distribution map of risk. Specifically, it can be implemented using a multi-branch convolutional neural network, capturing local and global risk patterns through convolutional kernels of different scales. Multi-scale feature extraction refers to a feature learning technique that simultaneously captures different granularity patterns in the input data. In risk coding, multi-scale feature extraction extracts risk patterns at the microscale (single node / local path), mesoscale (functional module / regional cluster), and macroscale (network topology) from the spatiotemporal distribution map of risk through convolutional kernels or attention mechanisms with different receptive fields, ensuring that risk features contain both details and global context. Spatiotemporal context information refers to the correlation information of risk features in the time and space dimensions, which makes the risk value of each node or region not exist independently, but closely related to the risk status of its neighboring nodes and the historical evolution trend of its own risk, providing a dynamic and correlated perspective for risk impact assessment.

[0069] S42: The feature fusion layer performs cross-modal fusion of risk feature representation and feature vector set, and dynamically adjusts the contribution weights of different feature vectors in the feature vector set through a gating attention mechanism to generate joint feature representation; In this embodiment, cross-modal fusion refers to jointly modeling features from different data modalities. Specifically, it can be implemented using a fusion network based on an attention mechanism, achieving information complementarity by calculating the correlation between features. Gated attention mechanism refers to a calculation module that dynamically adjusts the weights of different features. Specifically, it can be implemented by combining a gated recurrent unit with attention scores, adaptively adjusting feature importance based on the current risk state. Contribution weight refers to a quantitative indicator calculated through the attention mechanism during feature fusion, reflecting the relative importance of each input feature to the final prediction task. The dynamic adjustment of contribution weight enables the model to flexibly focus on the most relevant features based on the characteristics of the current input sample, improving the adaptability of the prediction. Joint feature representation refers to the unified feature vector generated after cross-modal fusion and weight adjustment. The joint feature representation integrates the statistical regularity of the original feature vector set, text semantics, and risk situation information contained in the risk spatiotemporal distribution map, forming a comprehensive and context-aware feature foundation for the final prediction.

[0070] S43: The prediction generation layer generates and outputs collaborative prediction results information based on joint feature representation.

[0071] In this embodiment, the collaborative prediction result information refers to the enhanced prediction result output by the prediction generation layer, which integrates the baseline demand prediction and risk impact assessment. The collaborative prediction result information is usually not a single numerical value, but a structured output, which may include point prediction values, prediction ranges under different risk scenarios, and prediction uncertainty measures, providing decision-makers with comprehensive and actionable information.

[0072] Specifically, the risk encoding layer first performs spatial pyramid pooling and temporal sliding window processing on the spatiotemporal distribution map of risk to extract risk features at different spatiotemporal resolutions; the feature fusion layer concatenates the risk features with the statistical features of the feature vector set and the text sentiment features, calculates the attention score of each feature through a gated attention network, and generates a joint feature representation by weighted summation based on the scores; the prediction generation layer adopts a multi-task learning framework to generate demand prediction results for different time spans in parallel based on the joint feature representation, and outputs collaborative prediction result information with confidence intervals through a probability calibration module.

[0073] Through the above technical solution, this application can achieve dynamic coupling of risk factors and demand forecasting, automatically adjust the feature focus of the forecasting model when there are sudden changes in the external environment or accumulation of internal risks, effectively reduce the forecasting deviation caused by risk transmission, and generate collaborative forecasting results containing multi-dimensional uncertainties, providing more comprehensive information support for supply chain decision-making.

[0074] In one embodiment, step S43 includes: S431: Based on joint feature representation, temporal features at different time scales are extracted through parallel temporal convolutional networks. The temporal features include short-term fluctuation features, periodic features, and long-term trend features. In this embodiment, the parallel temporal convolutional network refers to a neural network that uses a parallel convolutional structure to process input data with different time spans. Specifically, it can be implemented using parallel convolutional layers with different kernel sizes. For example, setting the kernel width to 3, 7, and 15 can capture short-term, periodic, and long-term features respectively. Short-term fluctuation features are feature representations reflecting local changes and immediate responses in a sequence. They are extracted using small-scale convolutional kernels (such as sizes 3-7) to capture random fluctuations, the impact of sudden events, and short-term market reactions. Their feature change period is usually short, corresponding to high-frequency noise and instantaneous disturbance patterns. Periodic... Regularity features refer to the recurring patterns in a sequence. They are extracted by matching convolutional kernels of seasonal cycle length (e.g., a convolutional kernel with a period of 7 to capture weekly patterns) to reflect fixed cyclical patterns such as weekend effects, monthly settlement cycles, and seasonal consumption. Their feature cycles are usually fixed and predictable. Long-term trend features refer to the features that depict the overall direction of change and macroeconomic trends in a sequence. They are extracted by large-scale convolutional kernels (e.g., larger than 30) or dilated convolutions to filter out short-term noise and cyclical fluctuations, reflecting trend changes driven by slow-changing factors such as economic cycles, technological iterations, and market structure changes.

[0075] S432: The extracted time-series features are input into the corresponding prediction head network to generate multi-scenario prediction information, which includes baseline scenario prediction information, uplink fluctuation scenario prediction information and downlink fluctuation scenario prediction information. In this embodiment, the prediction head network refers to a neural network branch trained independently for different prediction scenarios. Specifically, it can be implemented using a combination of fully connected layers and activation functions. For example, the baseline prediction head uses a linear activation function, while the uplink and downlink prediction heads use nonlinear functions with constraints. The baseline scenario prediction information refers to the demand forecast result generated based on historical patterns and trends after excluding known major risk events and special market fluctuations under normal operating conditions. The uplink fluctuation scenario prediction information refers to the optimistic prediction scenario where demand may exceed the baseline level under the drive of positive factors. The downlink fluctuation scenario prediction information refers to the pessimistic prediction scenario where demand may fall below the baseline level under the impact of risk factors.

[0076] S433: Based on risk feature representation, calibrate multi-scenario prediction information and adjust the confidence of multi-scenario prediction information through risk attention weight to generate demand prediction distribution; In this embodiment, risk attention weight refers to the weight coefficient dynamically generated based on risk characteristics. Specifically, a multi-head attention mechanism can be used to calculate the correlation between different scenario predictions and risk states. For example, a query-key value matching mechanism can be used to quantify the degree of impact of risk on prediction volatility. Demand prediction distribution refers to the probability density function of the predicted value generated through probability modeling. The demand prediction distribution represents the uncertainty of the prediction result. It is usually output by a mixed density network or quantile regression, providing statistical information including mean, variance and complete distribution shape, rather than just point estimation.

[0077] S434: Calculate key forecasting indicators and their confidence intervals based on the demand forecasting distribution, and generate collaborative forecasting results information.

[0078] In this embodiment, key forecast indicators refer to decision support indicators extracted from the forecast distribution, including point forecast values ​​(such as expected value, median), value at risk, conditional value at risk, forecast intervals (such as 90% confidence intervals), etc., to provide customized information for decision-makers with different risk preferences; confidence intervals refer to the probability range that characterizes the uncertainty of the forecast result, which is obtained by calculating the quantiles of the forecast distribution (such as the 5% and 95% quantiles forming the 90% confidence interval). The width of the interval reflects the forecast stability, with a narrow interval indicating high confidence and a wide interval indicating high uncertainty.

[0079] Specifically, for example, the joint feature representation is first fed into three parallel temporal convolutional networks, such as network branches with kernel widths of 3, 7, and 15, to extract demand change patterns at the weekly, monthly, quarterly, and annual levels, respectively. The temporal features output by each branch are then fed into the corresponding prediction head network. For example, the baseline scenario prediction head uses a linear regression structure, while the uplink and downlink prediction heads use neural networks with boundary constraints to generate prediction results covering different fluctuation ranges. Subsequently, the risk feature representation calculates the confidence weight of each scenario prediction through an attention mechanism, for example, assigning higher weights to downlink predictions during high-risk periods. Finally, the weighted multi-scenario predictions are fused into a probability distribution, and their mean and confidence interval are calculated as the collaborative prediction result.

[0080] Through the above technical solution, this application solves the problem that the prediction results in the prior art rely too much on a single time dimension and cannot cover sudden risk scenarios. It realizes the effective decoupling of features at multiple time scales and the dynamic generation of predictions for multiple scenarios. At the same time, by calibrating the prediction confidence through a risk attention mechanism, the final output collaborative prediction results can quantify the range of demand fluctuations under different risk levels, providing more comprehensive data support for supply chain decision-making.

[0081] In one embodiment, step S432 includes: S4321: Obtain historical demand data and generate a multi-scenario condition vector based on the quantile statistics of the historical demand data. The multi-scenario condition vector includes a baseline scenario condition vector, an upward fluctuation scenario condition vector, and a downward fluctuation scenario condition vector. In this embodiment, historical demand data refers to a set of actual demand records arranged in chronological order that are related to the prediction target; multi-scenario condition vectors refer to feature vectors corresponding to different fluctuation amplitudes extracted from historical demand data using quantile statistics methods. Specifically, quantile regression algorithms can be used to calculate the 10%, 50%, and 90% quantile values ​​of historical demand data as the basis for generating condition vectors, which are used to quantify the demand change boundaries under different market fluctuation scenarios; the baseline scenario condition vector represents the vectorized representation of typical demand patterns under normal operating conditions. The baseline scenario condition vector typically corresponds to the feature patterns near the 50th quantile (median) of the historical demand distribution, reflecting the exclusion of extremes. The stable demand state after the fluctuation serves as the benchmark anchor for forecast calibration. The upward fluctuation scenario condition vector represents the vectorized representation of the demand expansion pattern under positive market conditions. The upward fluctuation scenario condition vector usually corresponds to the characteristic pattern of the upper tail part of the historical demand distribution (such as above the 75th percentile), capturing typical characteristic combinations when demand grows abnormally, and is used to generate optimistic scenario forecasts. The downward fluctuation scenario condition vector represents the vectorized representation of the demand contraction pattern under negative market conditions. The downward fluctuation scenario condition vector usually corresponds to the characteristic pattern of the lower tail part of the historical demand distribution (such as below the 25th percentile), extracting typical characteristic combinations when demand falls abnormally, and is used to generate pessimistic scenario forecasts.

[0082] S4322: Concatenate the extracted temporal features with the corresponding multi-scenario conditional vectors to generate scenario-specific conditional feature representations; In this embodiment, context-specific conditional feature representation refers to a composite feature obtained by fusing temporal features with contextual conditional vectors. Specifically, a feature concatenation layer can be used to connect temporal feature vectors and conditional vectors in the feature dimension to form a feature expression with contextual constraints.

[0083] S4323: The generated conditional feature representations are input into the corresponding prediction head networks to generate baseline scenario prediction information, uplink fluctuation scenario prediction information and downlink fluctuation scenario prediction information, respectively. The prediction head network includes a baseline prediction head, an uplink prediction head and a downlink prediction head.

[0084] In this embodiment, the prediction head network refers to a prediction model independently trained for different volatility scenarios. Specifically, it can adopt a fully connected network structure with different activation function thresholds. For example, the uplink prediction head can be configured with the ReLU activation function to capture positive volatility features, while the downlink prediction head can be configured with the LeakyReLU function to handle negative volatility. The baseline prediction head is a neural network module specifically used to generate baseline scenario predictions. The baseline prediction head usually adopts conservative activation functions and regularization strategies, focusing on the stability and interpretability of the predictions, and outputting smooth baseline prediction results that conform to historical trends. The uplink prediction head is a prediction module specifically used to capture and amplify positive volatility signals. The uplink prediction head usually includes an asymmetric activation function (such as ReLU) and a loss function that focuses on positive examples, enhancing the model's sensitivity to uplink signals and accurately capturing growth opportunities. The downlink prediction head is a prediction module specifically used to identify and warn of negative volatility risks. The downlink prediction head usually introduces a risk weighting mechanism and downlink protection constraints to strengthen the model's ability to identify risk signals and provide early warnings for risk prevention and control.

[0085] Specifically, historical demand data is processed using quantile statistics to generate three sets of condition vectors, corresponding to the scenario boundaries of normal market demand fluctuation, positive expansion, and negative contraction, respectively. The extracted time-series features are concatenated with the corresponding multi-scenario condition vectors to form a feature representation with scenario labels. The fused feature representations are then fed into three independently trained prediction head networks, each optimized for a specific fluctuation range. For example, the baseline prediction head uses only samples near the 50th percentile of historical data for parameter updates during training, the uplink prediction head uses a sample set above that quantile, and the downlink prediction head uses a sample set below that quantile. During the prediction phase, the three prediction heads output demand prediction values ​​for the corresponding scenarios in parallel based on the current feature representations, forming a set of prediction results covering different fluctuation ranges.

[0086] Through the above technical solution, this application realizes the scenario-based segmentation output of demand forecast results. The multi-scenario forecasting framework can provide supply chain managers with references to demand changes under different risk levels. For example, when the risk of rising raw material prices is detected, the demand suppression effect assessment output by the upward forecast head can be referenced. When the risk of logistics disruption occurs, the supply shortage impact forecast output by the downward forecast head can be used to predict, thereby supporting the dynamic adjustment of inventory strategies and supply plans.

[0087] In one embodiment, step S60 includes: S61: Based on the risk level information of high-risk nodes and the demand fluctuation range in the collaborative prediction results, calculate the comprehensive risk score and generate a comprehensive risk distribution based on all comprehensive risk scores; In this embodiment, demand fluctuation range refers to the relative or absolute difference between the predicted demand value and the benchmark demand value under a specific forecast scenario, used to quantify the degree of positive or negative change that may occur in future demand; comprehensive risk score refers to an assessment indicator that quantitatively integrates the risk level of a node with the demand fluctuation range, which can be implemented by weighted summation or nonlinear function mapping, used to reflect the vulnerability of a node under the dual pressure of risk and demand; comprehensive risk distribution refers to the probability distribution formed based on the comprehensive risk scores of all high-risk nodes. The comprehensive risk distribution reflects the statistical characteristics of risk values ​​in the entire supply chain network, including the central tendency, dispersion, and distribution pattern of risk values, providing a data basis for the dynamic classification of risk warning levels.

[0088] S62: Based on the comprehensive risk distribution, risk warning levels are divided, and risk transmission simulation is performed on each high-risk node in turn based on the supply chain network topology to determine the impact range distribution, which includes the impact range of upstream supply interruption and the impact range of downstream delivery delay. In this embodiment, the risk warning level refers to the warning level divided according to the distribution range of the comprehensive risk score. Specifically, it can be implemented using quantile thresholds or clustering algorithms to distinguish warning signals of different urgency levels. Risk transmission simulation refers to the calculation process of simulating the propagation and diffusion of risk in the supply chain network based on graph theory and system dynamics principles. By establishing node state transition equations and risk propagation probability models, it predicts the path, speed, and impact intensity of risk transmission from the source node to upstream and downstream nodes along network connections. Impact range distribution refers to the spatial impact assessment result output by the risk transmission simulation. This impact range distribution quantitatively describes the impact of risk events on different areas of the supply chain network. The extent of the impact on the domain includes information on the geographical distribution of affected nodes, the scope of business function impairment, and estimated recovery time. The impact of upstream supply disruptions refers to the risk impact assessment transmitted in reverse along the supply chain network. By simulating the propagation of risk towards suppliers, it identifies potentially affected raw material supply, component production, and primary product manufacturing links, and assesses the depth and breadth of supply disruptions. The impact of downstream delivery delays refers to the risk impact assessment transmitted in forward along the supply chain network. By simulating the propagation of risk towards customers, it identifies potentially affected finished product distribution, distribution networks, and end-customer delivery links, and assesses the scope and timeliness impact of delivery delays.

[0089] S63: Integrate the risk warning levels and the distribution of the scope of impact into a tiered risk warning information system.

[0090] Specifically, upon identifying a high-risk node, the risk intensity value corresponding to that node in the spatiotemporal risk distribution map is first extracted. This could be a probability value or standardized score within the 0-1 range. Simultaneously, the demand fluctuation magnitude associated with that node in the collaborative forecast results is obtained, such as the standard deviation or quantile difference of the demand forecast values. These two values ​​are then merged into a comprehensive risk score using pre-defined calculation rules, such as a non-linear combination of the risk intensity value and the demand fluctuation coefficient. The comprehensive risk scores of all nodes form a network-level risk distribution map, which is then divided into different levels using pre-defined grading rules. For example, the top 10% of nodes are marked as red alerts, and the middle 20% as orange alerts. Subsequently, based on the supply chain network topology, risk transmission path simulation is performed for each high-risk node. For example, a breadth-first search algorithm is used to traverse its upstream suppliers and downstream customers, counting the number of affected entities and their hierarchical depth to form impact range distribution data. Finally, the warning level label is bound to the impact range data to generate a tiered warning report containing the risk level, transmission path, and response recommendations.

[0091] Through the above technical solutions, this application realizes dynamic quantitative assessment of the risk status of supply chain nodes, solving the problem of misjudgment caused by the separation of risk warning and demand forecast in the prior art; it accurately defines the scope of impact through risk transmission simulation, solving the defect of overly general risk impact range estimation in traditional methods; the generation of graded early warning information enables enterprises to allocate emergency resources according to the priority of risk level and scope of impact, such as implementing inventory allocation or alternative supplier activation plans for red warning nodes, and initiating monitoring enhancement measures for orange warning nodes.

[0092] In one embodiment, a supply chain demand forecasting and risk warning system based on multi-source data is provided. This system corresponds one-to-one with the supply chain demand forecasting and risk warning method based on multi-source data described in the previous embodiment. The supply chain demand forecasting and risk warning system based on multi-source data includes: The data acquisition module is used to collect multi-source heterogeneous data and construct a feature vector set based on the multi-source heterogeneous data. The multi-source heterogeneous data includes enterprise operation data and environmental situation data. The feature vector set includes statistical feature vectors, text sentiment feature vectors and event feature vectors. The topology construction module is used to construct a supply chain network topology based on enterprise operation data. The supply chain network topology includes several nodes and their connection relationships. The nodes include supply chain entities, and the connection relationships include logistics relationships, information flow relationships, and capital flow relationships between supply chain entities. The risk simulation module is used to input the feature vector set and the supply chain network topology into a pre-trained risk propagation model to simulate risk and obtain a risk spatiotemporal distribution map. The demand forecasting module is used to input the feature vector set and the risk spatiotemporal distribution map into the pre-trained demand forecasting model to perform demand forecasting and obtain collaborative forecasting results. The risk node identification module is used to generate a node criticality distribution map based on the supply chain network topology, and compare the collaborative prediction results with the node criticality distribution map to identify whether there are high-risk nodes. The risk information generation module is used to obtain the risk level information of high-risk nodes in the spatiotemporal distribution map of risk if high-risk nodes exist, and to generate graded risk warning information based on the high-risk nodes and risk level information.

[0093] For specific limitations regarding a supply chain demand forecasting and risk warning system based on multi-source data, please refer to the limitations of a supply chain demand forecasting and risk warning method based on multi-source data mentioned above, which will not be repeated here. Each module in the aforementioned supply chain demand forecasting and risk warning system based on multi-source data can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0094] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for supply chain demand forecasting and risk warning based on multi-source data, characterized in that: The method comprises the steps of: collecting multi-source heterogeneous data, and constructing a feature vector set based on the multi-source heterogeneous data, wherein the multi-source heterogeneous data comprises enterprise operation data and environmental situation data, and the feature vector set comprises a statistical feature vector, a text sentiment feature vector, and an event feature vector; constructing a supply chain network topology based on the enterprise operation data, wherein the supply chain network topology comprises a plurality of nodes and connection relationships therebetween, the nodes comprise supply chain entities, and the connection relationships comprise logistics relationships, information flow relationships, and fund flow relationships between the supply chain entities; inputting the feature vector set and the supply chain network topology into a pre-trained risk propagation model to perform risk simulation, and obtaining a risk spatiotemporal distribution map; inputting the feature vector set and the risk spatiotemporal distribution map into a pre-trained demand prediction model to perform demand prediction, and obtaining collaborative prediction result information; generating a node criticality distribution map based on the supply chain network topology, and comparing the collaborative prediction result information with the node criticality distribution map to identify whether there is a high-risk node; if there is a high-risk node, obtaining risk level information of the high-risk node in the risk spatiotemporal distribution map, and generating hierarchical risk early warning information based on the high-risk node and the risk level information.

2. The method of claim 1, wherein the method is characterized by: The step of collecting multi-source heterogeneous data, and constructing a feature vector set based on the multi-source heterogeneous data, wherein the multi-source heterogeneous data comprises enterprise operation data and environmental situation data, and the feature vector set comprises a statistical feature vector, a text sentiment feature vector, and an event feature vector, comprises the steps of: resampling and time aligning the multi-source heterogeneous data to a preset minimum analysis time granularity to obtain time-aligned data; extracting multi-modal features from the time-aligned data to obtain a preliminary feature vector set; calculating mutual information values between different features in the preliminary feature vector set, and generating feature cross combinations based on the mutual information values; screening the feature cross combinations by a preset screening strategy, and taking a screened feature subset as the feature vector set.

3. The method of claim 1, wherein the method further comprises: The step of constructing a supply chain network topology based on the enterprise operation data, wherein the supply chain network topology comprises a plurality of nodes and connection relationships therebetween, the nodes comprise supply chain entities, and the connection relationships comprise logistics relationships, information flow relationships, and fund flow relationships between the supply chain entities, comprises the steps of: extracting entity information, order information, and logistics path information from the enterprise operation data, establishing multi-level supply relationships between the entity information based on the order information and the logistics path information, and forming a supply chain entity relationship graph; extracting transaction frequency information and logistics time efficiency information based on the order information and the logistics path information, giving time-varying dynamic weights to the multi-level supply relationships based on the transaction frequency information and the logistics time efficiency information, and constructing a dynamic supply chain network; identifying key hub nodes in the dynamic supply chain network, dividing the supply chain network into a core layer, a support layer, and an edge layer based on the key hub nodes, and forming the supply chain network topology.

4. The method of claim 3, wherein the method further comprises: The step of extracting entity information, order information and logistics path information from enterprise operation data, establishing multi-level supply relationship between entities based on order information and logistics path information, and forming a supply chain entity relationship graph includes the steps of: decomposing the logistics path information into a plurality of transportation segment sequence information, and clustering the transportation segment sequence information to generate high-frequency path patterns; identifying transaction subjects and transaction relationships in the order information, and constructing a direct supply relationship network between entities based on the transaction subjects and transaction relationships, and establishing a logistics topology relationship between entity information in combination with the high-frequency path patterns; extracting transaction time sequence features in the order information, and identifying time dynamic patterns between entity information based on the transaction time sequence features, the time dynamic patterns including periodic supply patterns and sudden supply patterns; fusing the direct supply relationship network, the logistics topology relationship and the time dynamic patterns in multiple dimensions to construct a multi-level supply relationship network; identifying key supply paths in the multi-level supply relationship network, and generating a supply chain entity relationship graph based on a pre-set weight distribution strategy.

5. The method of claim 1, wherein the method further comprises: The risk propagation model includes a feature fusion layer, a space-time propagation layer, and a risk field generation layer. The step of inputting the feature vector set and the supply chain network topology structure into the pre-trained risk propagation model for risk simulation to obtain a risk space-time distribution map includes the steps of: The feature fusion layer aligns and enhances the feature vector set and the supply chain network topology structure to form an enhanced supply chain network with multi-dimensional node and connection edge features; The space-time propagation layer performs risk propagation calculation based on the enhanced supply chain network, and outputs a space-time risk state sequence in node units; The risk field generation layer maps the space-time risk state sequence from discrete nodes to continuous space-time domain to generate a risk space-time distribution map.

6. The method of claim 1, wherein the method further comprises: The demand prediction model includes a risk encoding layer, a feature fusion layer and a prediction generation layer. The step of inputting the feature vector set and the risk space-time distribution map into the pre-trained demand prediction model for demand prediction to obtain collaborative prediction result information includes the steps of: The risk encoding layer extracts multi-scale features from the risk space-time distribution map and generates a risk feature representation with space-time context information; The feature fusion layer performs cross-modal fusion of the risk feature representation and the feature vector set, and dynamically adjusts the contribution weight of different feature vectors in the feature vector set through a gating attention mechanism to generate a joint feature representation; The prediction generation layer generates and outputs collaborative prediction result information based on the joint feature representation.

7. The method of claim 6, wherein the method further comprises: The step of generating and outputting collaborative prediction result information based on the joint feature representation by the prediction generation layer includes the steps of: Based on the joint feature representation, extract time series features at different time scales through a parallel time convolution network, the time series features including short-term fluctuation features, periodic features and long-term trend features; input the extracted time series features into corresponding prediction head networks respectively to generate multi-scenario prediction information, the multi-scenario prediction information including baseline scenario prediction information, uplink fluctuation scenario prediction information and downlink fluctuation scenario prediction information; The multi-scenario prediction information is calibrated based on the risk feature representation, and confidence of the multi-scenario prediction information is adjusted through a risk attention weight, to generate a demand prediction distribution; Key prediction indexes and confidence intervals thereof are calculated based on the demand prediction distribution, to generate collaborative prediction result information.

8. The method of claim 7, wherein the method further comprises: The step of inputting the extracted time sequence features into corresponding prediction head networks respectively to generate multi-scenario prediction information, the multi-scenario prediction information including benchmark scenario prediction information, uplink fluctuation scenario prediction information, and downlink fluctuation scenario prediction information, comprises the steps of: Historical demand data is acquired, and multi-scenario condition vectors are generated based on quantile statistics of the historical demand data, the multi-scenario condition vectors including benchmark scenario condition vectors, uplink fluctuation scenario condition vectors, and downlink fluctuation scenario condition vectors; The extracted time sequence features are spliced with corresponding multi-scenario condition vectors, and scenario-specific conditioned feature representations are generated; The generated conditioned feature representations are input into corresponding prediction head networks respectively to generate benchmark scenario prediction information, uplink fluctuation scenario prediction information, and downlink fluctuation scenario prediction information respectively, the prediction head networks including a benchmark prediction head, an uplink prediction head, and a downlink prediction head.

9. The method of claim 1, wherein the method further comprises: If there is a high-risk node, risk level information of the high-risk node in a risk spatiotemporal distribution map is acquired, and hierarchical risk early warning information is generated based on the high-risk node and the risk level information, the step comprising the steps of: Based on the risk level information of the high-risk node and demand fluctuation amplitudes in the collaborative prediction result information, comprehensive risk scores are calculated, and a comprehensive risk distribution is generated based on all the comprehensive risk scores; Based on the comprehensive risk distribution, risk early warning levels are divided, and risk conduction simulation is sequentially performed on each high-risk node based on the supply chain network topology to determine an influence range distribution, the influence range distribution including an upstream supply interruption influence range and a downstream delivery delay influence range; The risk early warning levels and the influence range distribution are integrated into the hierarchical risk early warning information.

10. A multi-source data-based supply chain demand forecasting and risk warning system, characterized in that: It comprises: A data acquisition module is configured to collect multi-source heterogeneous data, and construct a feature vector set based on the multi-source heterogeneous data, the multi-source heterogeneous data including enterprise operation data and environment situation data, and the feature vector set including statistical feature vectors, text sentiment feature vectors, and event feature vectors; A topology structure construction module is configured to construct a supply chain network topology based on the enterprise operation data, the supply chain network topology containing a plurality of nodes and connection relationships therebetween, the nodes including supply chain entities, and the connection relationships including material flow relationships, information flow relationships, and fund flow relationships between the supply chain entities; A risk simulation module is configured to input the feature vector set and the supply chain network topology into a pre-trained risk propagation model to perform risk simulation, and obtain a risk spatiotemporal distribution map; A demand prediction module is configured to input the feature vector set and the risk spatiotemporal distribution map into a pre-trained demand prediction model to perform demand prediction, and obtain collaborative prediction result information; The risk node identification module is configured to generate a node criticality distribution diagram based on the supply chain network topology, compare the collaborative prediction result information with the node criticality distribution diagram, and identify whether there is a high-risk node. The risk information generation module is configured to, if there is a high-risk node, acquire risk level information of the high-risk node in a risk space-time distribution diagram, and generate hierarchical risk early warning information based on the high-risk node and the risk level information.

Images