Big data prediction method and device based on machine learning, equipment and medium

Abstract

The invention relates to a big data prediction method and device based on machine learning, equipment and a medium, and the method comprises the steps: converting millisecond-level power grid real-time monitoring data into frequency domain features through wavelet packet transformation, and solving the timeliness conflict between high-frequency data and low-frequency commercial data; a tensor fusion technology is utilized to align multi-source heterogeneous information such as power grid topology, enterprise carbon emission and financial data, and unified space-time representation is constructed; structured constraint features are generated in combination with semantic analysis of policies and regulations, and dynamic association between environmental policies and commercial activities is quantified through cross-modal embedding; modeling a risk conduction path based on a dynamic graph neural network, and optimizing a prediction model according to an asymmetric loss mechanism; and finally, outputting a supply chain risk control strategy, and realizing closed-loop response from power grid fluctuation monitoring to supply chain decision. According to the method, the limitation of a traditional model in the aspects of cross-scale data fusion and policy conduction quantification is broken through, and the prediction precision and the response timeliness in an emergency risk scene are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a big data prediction method and device based on machine learning, equipment and medium. BACKGROUND

[0002] In the field of supply chain financial risk management, big data prediction methods based on machine learning are gradually becoming the core means of risk early warning. Such methods integrate enterprise operation data, environmental indicators, policy information and other multi-source heterogeneous data to build a dynamic risk transmission model to identify potential risk nodes in the supply chain. Especially in the field of power grid equipment dispatching and supply chain intersection, real-time monitoring data such as current fluctuations and load changes can reflect the stability of enterprise production, while indicators such as carbon emission policies and financial health constitute risk factors in the environmental and economic dimensions. Existing technologies usually attempt to combine power grid data with enterprise business data to establish a unified supply chain risk assessment model.

[0003] However, the existing methods have two key defects: first, there is a significant time difference between real-time monitoring data and enterprise carbon emissions and financial data (monthly / quarterly period). Traditional data alignment methods can cause feature distortion and cannot capture the dynamic correlation between power grid transient fluctuations and enterprise operating status; second, unstructured information such as policy texts cannot be effectively integrated with quantitative data. Existing models often treat policies as static thresholds or manually adjusted parameters, and cannot quantify the transmission effect of policy changes on supply chain risks. This time conflict and lack of unstructured information processing make the existing model lack accuracy in risk prediction under sudden policy adjustments or power grid abnormal scenarios, and it is difficult to support real-time risk control decisions. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a big data prediction method and device based on machine learning that can automatically integrate high-frequency power grid data and low-frequency business data and accurately quantify policy transmission risks.

[0005] The purpose of the present application is achieved by the following scheme:

[0006] In a first aspect, the present application provides a big data prediction method based on machine learning, comprising the following steps:

[0007] S1: Time-frequency feature extraction is performed on the acquired power grid real-time monitoring data stream, and the sampled power grid real-time monitoring data is converted into frequency domain feature representation through wavelet packet transform to generate a time-frequency matrix;

[0008] S2: Based on the time-frequency matrix, the obtained enterprise carbon emission data, enterprise financial data and power grid topology adjacency matrix from the power grid topology database are fused across scales, different time granularity data are aligned in a unified feature space through tensor direct product operation, and a space-time joint tensor is generated;

[0009] S3: Obtain the policy and regulation text disclosed by the government agency, combine the space-time joint tensor to perform joint feature transformation, decompose the tensor into a low-dimensional factor matrix and convert the policy text into a word vector matrix, splice and map to a unified space based on an activation function, construct a correlation space of carbon emission indicators and financial indicators, and generate a unified risk embedding vector;

[0010] S4: Risk transmission modeling is performed on the unified risk embedding vector, the state of the supply chain node is iteratively updated through a dynamic graph neural network, and the transmission probability distribution is calculated to generate a risk transmission probability;

[0011] S5: Strategy optimization is performed on the risk transmission probability, the difference in penalty weight between overestimation risk and underestimation risk is calculated, the model parameters are adjusted through gradient back propagation, the decision instruction with threshold judgment is output, and the supply chain risk control strategy is generated. The supply chain risk control strategy is used to indicate the adjustment of the purchase order or the raw material reserve.

[0012] In one of the embodiments, the S1 of the big data prediction method based on machine learning provided by the application specifically includes the following steps:

[0013] S11: Select a wavelet base function for the obtained power grid real-time monitoring data stream, select Morlet wavelet as the base function according to the frequency characteristics of the current waveform, and generate an optimized wavelet base;

[0014] S12: Perform wavelet packet decomposition processing on the optimized wavelet base and the sampled power grid real-time monitoring data, calculate the scale coefficients of the fundamental wave and harmonic components in the preset frequency band range, and calculate the discrete wavelet coefficients of the high-frequency noise components, to generate a multi-resolution frequency spectrum feature;

[0015] S13: Perform frequency domain reorganization on the multi-resolution frequency spectrum feature, separate the high-frequency noise components from the fundamental wave components, and reconstruct an energy density matrix to generate a time-frequency matrix.

[0016] In one of the embodiments, the S2 of the big data prediction method based on machine learning provided by the application specifically includes the following steps:

[0017] S21: Perform period segmentation processing on the time-frequency matrix, cut the continuous time-frequency data into equal-length monthly data blocks according to the carbon emission report period in the obtained enterprise carbon emission data, and generate a periodic sub-matrix;

[0018] S22: Feature concatenation is performed on the periodic sub-matrix and the enterprise carbon emission data, the carbon emission intensity value is expanded into a feature plane with the same dimension as the sub-matrix, and a space-time basic tensor is generated;

[0019] S23: Based on the space-time basic tensor, the obtained enterprise financial data and the power grid topology adjacency matrix of the power grid topology structure database are fused in terms of financial features and spatial relationships, the asset-liability ratio quarterly difference is calculated and coded as an additional feature layer, and the power grid topology adjacency matrix is superimposed as a spatial relationship layer, a space-time joint tensor is generated, and the space-time joint tensor is used to indicate the business activity and environmental constraint correlation characteristics between power grid nodes.

[0020] In one of the embodiments, the S3 of the big data prediction method based on machine learning provided by the application specifically comprises the following steps:

[0021] S31: Obtain the policy and regulation text disclosed by the government agency, analyze the legal provisions of the policy and regulation text, extract key constraint conditions such as carbon emission quota and penalty provisions, and generate structured policy features;

[0022] S32: Perform multi-dimensional feature compression processing on the space-time joint tensor, identify periodic change patterns in the time dimension, extract node distribution features in the space dimension, and filter key correlation parameters in the index dimension, reconstruct the data space by retaining principal components with a variance contribution rate exceeding a set threshold in each dimension, and generate a core factor tensor;

[0023] S33: Perform cross-modal fusion processing on the core factor tensor and the structured policy features, map to a unified vector space through a fully connected layer and standardize, and generate a unified risk embedding vector.

[0024] In one of the embodiments, the expression of the unified risk embedding vector of the big data prediction method based on machine learning provided by the application is:

[0025]

[0026] wherein, is the output unified risk embedding vector, is the matrix representation of the core factor tensor after modal expansion, is the structured policy feature matrix, is the fully connected layer weight matrix, is the fully connected layer bias vector, is the rectified linear unit activation function, denotes the layer standardization operation, denotes the matrix longitudinal splicing operation.

[0027] In one embodiment, step S4 of the big data prediction method based on machine learning provided by the present invention specifically includes the following steps:

[0028] S41: Perform dynamic graph initialization on the unified risk embedding vector, construct an initial adjacency matrix containing enterprise nodes and grid nodes based on the grid topology connection relationship, and generate a spatiotemporal constraint graph;

[0029] S42: Dynamically calculate edge weights for the spatiotemporal constraint graph, update the connection strength between enterprise nodes based on the transaction frequency between enterprises and the similarity of risk embedding vectors, and generate an adaptive adjacency matrix;

[0030] S43: Perform graph convolution propagation processing on the adaptive adjacency matrix, and iteratively update the hidden state of enterprise nodes in the supply chain network through gated cyclic units to generate risk transmission probability. The risk transmission probability is used to indicate the likelihood of risk being transmitted from upstream enterprises to downstream enterprises in the supply chain network.

[0031] In one embodiment, S5 of the big data prediction method based on machine learning provided by the present invention specifically includes the following steps:

[0032] S51: Perform asymmetric loss calculation on the risk transmission probability, set different penalty weights based on the difference in the impact of overestimating and underestimating risk on the supply chain, and generate an optimization objective function;

[0033] S52: Perform gradient backpropagation on the objective function, update the prediction model parameters through the optimization algorithm, and generate a robust prediction model;

[0034] S53: Make strategic decisions based on the risk transmission probability output by the robust prediction model. When the transmission probability exceeds the preset risk threshold, identify key risk transmission paths and generate multi-level response instructions, and integrate them to generate supply chain risk control strategies.

[0035] Secondly, the present invention provides a big data prediction device based on machine learning, which is configured with the following modules:

[0036] The power grid time-frequency feature extraction module is used to extract time-frequency features from the acquired real-time power grid monitoring data stream. It converts the sampled real-time power grid monitoring data into a frequency domain feature representation through wavelet packet transform, and generates a time-frequency matrix.

[0037] The cross-scale data fusion module is used to perform cross-scale fusion of acquired corporate carbon emission data, corporate financial data and power grid topology adjacency matrix from power grid topology database based on time-frequency matrix. It aligns data of different time granularities in a unified feature space through tensor direct product operation to generate spatiotemporal joint tensor.

[0038] The joint feature transformation embedding module is used to obtain publicly available policy and regulatory texts from government agencies, perform joint feature transformations by combining spatiotemporal joint tensors, decompose the tensors into low-dimensional factor matrices and convert the policy texts into word vector matrices, and then concatenate them and map them to a unified space based on an activation function to construct a correlation space between carbon emission indicators and financial indicators, and generate a unified risk embedding vector.

[0039] The risk transmission modeling module is used to model risk transmission on a unified risk embedding vector. It iteratively updates the state of supply chain nodes and calculates the transmission probability distribution through a dynamic graph neural network to generate risk transmission probability.

[0040] The risk control strategy optimization module is used to optimize the risk transmission probability, calculate the difference in penalty weights between overestimating and underestimating risk, adjust model parameters through gradient backpropagation, output decision instructions with threshold judgment, and generate supply chain risk control strategies. These supply chain risk control strategies are used to indicate adjustments to purchase orders or raw material reserves.

[0041] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned big data prediction methods based on machine learning.

[0042] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the aforementioned machine learning-based big data prediction methods.

[0043] In summary, the big data prediction method based on machine learning provided in this application converts millisecond-level real-time power grid monitoring data into frequency domain features through wavelet packet transform, effectively overcoming the timeliness conflict between high-frequency physical data and low-frequency commercial data, and achieving accurate alignment between power grid transient fluctuation characteristics and enterprise operating status. By utilizing tensor direct product operations to fuse multi-source heterogeneous information such as power grid topology, enterprise carbon emissions, and financial data in a unified feature space, a joint representation that preserves spatiotemporal correlation can be constructed to achieve lossless integration of cross-domain data. Through the collaborative processing of policy and regulatory semantic parsing and tensor decomposition, unstructured policy clauses can be transformed into quantifiable constraint features, and by combining activation function mapping, a dynamic correlation between environmental policies and business activities can be constructed. In terms of space, it achieves accurate quantification of the implicit risk transmission effect; based on the iterative update mechanism of dynamic graph neural network, it can capture the real-time evolution law of supply chain node status and generate risk transmission probability distribution, so as to visualize the propagation path of risk in multi-level supply chain network; adopting the asymmetric loss function to optimize the strategy decision model, it can dynamically adjust the prediction weight according to the difference in the impact of overestimating and underestimating risk, and finally output supply chain control instructions with threshold judgment, which can achieve a closed-loop response from risk monitoring to decision execution. It systematically solves the technical bottlenecks of traditional models in cross-scale data fusion, policy transmission quantification and dynamic risk response, and significantly improves the accuracy of risk prediction and the timeliness of handling in the event of sudden policy adjustment or power grid anomaly.

[0044] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0045] Figure 1 A flowchart illustrating a big data prediction method based on machine learning, provided as an embodiment of this application;

[0046] Figure 2 A schematic diagram illustrating the process of generating a unified risk embedding vector provided in an embodiment of this application;

[0047] Figure 3 This is a schematic diagram of the structure of a big data prediction device based on machine learning, provided as another embodiment of this application. Detailed Implementation

[0048] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0050] In one embodiment, such as Figure 1 As shown, a big data prediction method based on machine learning is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a device including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0051] S1: Extract time-frequency features from the acquired real-time power grid monitoring data stream, and convert the sampled real-time power grid monitoring data into a frequency domain feature representation through wavelet packet transform to generate a time-frequency matrix.

[0052] Specifically, the system acquires real-time monitoring data streams of the power grid. During the acquisition process, smart sensors deployed at key nodes of the power grid continuously capture electrical parameter signals during grid operation. These signals cover various core parameters reflecting the operating status of the power grid. The edge gateway receives the raw signals transmitted by the smart sensors and transmits them to the system's data processing unit in real time according to a preset transmission protocol. Further, the system verifies each data point in the raw data stream based on the reasonable range of electrical parameters during normal power grid operation. If a data point exceeds this reasonable range, it is determined to be abnormal data. The system uses linear interpolation of adjacent data to replace the abnormal data to maintain the continuity of the data stream. The second step is to remove power frequency interference. The system selects filtering technology adapted to the frequency characteristics of power frequency interference to process the data stream after abnormal data removal, filtering out the interference components caused by power frequency influence to obtain a clean data stream.

[0053] Preferably, the system can use wavelet packet transform to extract time-frequency features from the clean data stream, select wavelet basis functions adapted to the non-stationary characteristics of the power grid data, and determine the number of wavelet packet decomposition levels based on the frequency distribution range of transient fluctuations in the power grid to ensure that the frequency bands obtained after decomposition can completely cover the main frequency ranges of transient fluctuations in the power grid. Further, the system employs a sliding window mechanism to perform windowing processing on the clean data stream. The window size setting needs to be combined with the data sampling time interval to ensure that each window can contain sufficient effective data, and the step size setting needs to avoid information redundancy or missing information in the time dimension, so that the data in adjacent windows can maintain correlation without duplication. The system performs wavelet packet decomposition on the data within each window, and calculates the energy value of each frequency band after decomposition. This energy value is directly used as the feature value of the corresponding frequency band. The system arranges the feature values ​​of each frequency band corresponding to all windows sequentially according to the dimensions, with the time window as the row dimension and the frequency band number as the column dimension, to construct a time-frequency matrix. Each element in the time-frequency matrix corresponds to the energy feature of a specific frequency band under a specific time window.

[0054] S2: Based on the time-frequency matrix, cross-scale fusion is performed on the acquired corporate carbon emission data, corporate financial data and the grid topology adjacency matrix from the grid topology database. The data of different time granularities are aligned in a unified feature space through tensor direct product operation to generate a spatiotemporal joint tensor.

[0055] Specifically, the system first performs time-granularity alignment on low-frequency data: For enterprise carbon emission data, after collecting relevant indicators, the system uses a cubic spline interpolation algorithm to convert the data into a time granularity that matches the real-time data of the power grid. Simultaneously, it corrects the interpolation results by combining the enterprise's production schedule to ensure that the corrected data is consistent with the enterprise's actual production rhythm, thereby generating a carbon emission data matrix. For enterprise financial data, after collecting relevant indicators, the system uses linear interpolation combined with an operating cycle correction algorithm, processing based on the enterprise's quarterly financial report data and monthly revenue reports. The correction coefficient is dynamically adjusted according to the industry's average operating cycle, ultimately generating a financial data matrix. Next, the system extracts supply chain-related nodes from the power grid topology database, constructing a power grid topology adjacency matrix using nodes as row and column indices. Matrix elements are quantified and calculated using historical transaction amount percentages, cooperation years, and logistics transportation distances. Elements corresponding to nodes without direct business connections are set to specific values. The system then expands the adjacency matrix into a three-dimensional static tensor. Finally, the system performs tensor direct product fusion operation, sequentially performing direct product of the power grid time-frequency matrix with the carbon emission data matrix and the financial data matrix, and then performing direct product of the resulting tensor with the power grid topology adjacency tensor to generate a spatiotemporal joint tensor, achieving comprehensive fusion of time, node association, time-frequency characteristics, carbon emissions, and financial data.

[0056] S3: Obtain publicly available policy and regulatory texts from government agencies, perform joint feature transformation using spatiotemporal joint tensors, decompose the tensors into low-dimensional factor matrices, and transform the policy texts into word vector matrices. After concatenation, map the resulting matrix to a unified space based on an activation function to construct a correlation space between carbon emission indicators and financial indicators, and generate a unified risk embedding vector.

[0057] Specifically, the system acquires publicly available policy and regulatory texts from government agencies. These official websites cover government departments responsible for carbon emission management, financial supervision, and grid operation, ensuring that the acquired policy texts cover all areas related to supply chain risks. The system retrieves policy and regulatory text data from designated official websites, adhering to the websites' access guidelines to avoid impacting their normal operation. Furthermore, the system employs a word segmentation tool adapted to Chinese semantic structures to segment the plain text sentence by sentence, breaking down continuous text into independent word units. The segmentation process ensures the semantic integrity of each word unit. Based on a pre-defined stop word list, words without actual semantic contribution are removed, retaining only those with practical meaning. The stop word list includes common modifiers and auxiliary words. The system then uses an entity recognition model to extract entities from the segmented and stop-word-removed text. This model is trained based on labeled data of core entities in the policy text, accurately identifying core entities related to carbon emissions, finance, grid operation, and policy supervision, forming a set of policy keywords. The system inputs a set of policy keywords into a word vector model. The word vector model is trained based on the contextual semantic association of keywords in the policy text, and can generate a corresponding word vector for each keyword. The word vectors of all keywords together form a word vector matrix.

[0058] Preferably, the system performs mean pooling on the word vector matrix, calculating the average of all word vectors to obtain a global policy vector that represents the overall semantic features of the policy. Simultaneously, the system performs tensor decomposition on the spatiotemporal joint tensor. The decomposition process aims to minimize the tensor reconstruction error, continuously adjusting the decomposition parameters through iterative calculations until the reconstruction error reaches a preset stable range. This ultimately yields three low-dimensional factor matrices, corresponding to low-dimensional representations of the time dimension, node dimension, and feature dimension, respectively, achieving dimensionality reduction of the high-dimensional tensor data. Further, the system concatenates the low-dimensional factor matrices corresponding to the feature dimensions with the global policy vector along the row dimension. Before concatenation, it ensures that the dimensions of the two matrices match; if dimensional differences exist, they are adjusted through linear transformation to form a fusion matrix. This matrix integrates the feature information of the spatiotemporal joint tensor with the semantic information of the policy text. To eliminate differences in the spatial distribution of different types of features, the system uses an adaptive activation function to perform nonlinear mapping on the fusion matrix, adjusting features from different sources to the same feature space.

[0059] Furthermore, the system calculates the correlation coefficient matrix between carbon emission indicators and financial indicators. The calculation process is based on historical data of carbon emissions and financial indicators, and statistical analysis is used to obtain the degree of correlation between the two types of indicators. This matrix reflects the mutual influence between carbon emission indicators and financial indicators. Using this correlation coefficient matrix as weights, the system constructs a correlation space between carbon emission indicators and financial indicators, which reflects the synergistic change characteristics of the two types of indicators. Finally, the system compresses the feature vector after activation function mapping into a low-dimensional representation through a fully connected layer. The dimensionality setting of the fully connected layer is determined based on the dimensionality requirements of risk features, generating a unified risk embedding vector. The three dimensions of this vector correspond to the time window, grid node, and risk feature, respectively, with each element representing the comprehensive risk feature of a specific node within a specific time window.

[0060] S4: Model risk transmission using a unified risk embedding vector, iteratively update the state of supply chain nodes and calculate the transmission probability distribution through a dynamic graph neural network to generate the risk transmission probability.

[0061] Specifically, the system uses each node in the supply chain as a vertex in a graph. The node features corresponding to each vertex are set as the feature vector of that node within the corresponding time window in a unified risk embedding vector, ensuring that the vertex features reflect the risk status of the node. The system uses the dynamic risk transmission coefficient between nodes as the edge weights of the graph. The initial edge weights are calculated by the system based on the cosine similarity of the unified risk embedding vectors corresponding to two nodes. The higher the similarity, the larger the initial edge weight, indicating a higher probability of risk transmission between nodes. The system updates the edge weights at fixed time intervals, acquiring relevant data on changes in node association strength. This data includes adjustments to transaction amounts between nodes, records of changes in cooperative relationships, and changes in logistics transportation frequency. Based on this data, the system corrects the edge weights so that they reflect the actual association status between nodes in real time. Through these steps, the system constructs a dynamic graph that updates periodically to adapt to changes in the association between supply chain nodes.

[0062] Furthermore, the system can employ an improved graph attention network to iteratively update the supply chain node states. This network introduces a time decay factor to correct the influence of historical node states on the current state, preventing excessive interference from historical states in the current risk assessment. The system calculates the state vector of each node according to a preset formula, which includes parameters such as the node's neighbor set, attention coefficients, weight matrix, and bias term. The neighbor set is determined by the system based on the edge relationships of the dynamic graph. The attention coefficients are obtained by normalizing the original attention scores using the Softmax function. The weight matrix and bias term are parameters that need to be adjusted during model training. The activation function uses a specified type suitable for node state calculation. The system performs node state updates according to a preset number of iterations, determined by the system based on the maximum path length of the supply chain. This ensures that the node state vector converges after multiple iterations, avoiding instability due to insufficient iterations or resource waste due to excessive iterations. After the iterations are complete, the system obtains the converged node state matrix.

[0063] Preferably, the system calculates the risk transmission intensity based on the converged node state vector, and uses the Softmax function to process the risk transmission intensity of each node, transforming the transmission intensity into a probability distribution. The system organizes these probability values ​​into a risk transmission probability matrix, where each element in the matrix represents the probability that the corresponding row node transmits risk to the corresponding column node.

[0064] S5: Optimize the risk transmission probability strategy, calculate the difference in penalty weights between overestimating and underestimating risk, adjust the model parameters through gradient backpropagation, output decision instructions with threshold judgment, and generate supply chain risk control strategies. Supply chain risk control strategies are used to indicate adjustments to purchase orders or raw material reserves.

[0065] Specifically, the system retrieves data from the historical risk event database and uses this data to calculate two types of risk losses: overestimation and underestimation of risk. The first type is supply chain disruption losses caused by risks that actually occurred but were not predicted. These losses include order default compensation, equipment idle costs during production stoppages, labor costs, and customer churn losses. The second type is excessive inventory costs caused by risks that did not occur but were misjudged as occurring. These costs include warehousing costs from raw material stockpiling, interest costs from capital occupation, and losses from expired or depreciated raw materials.

[0066] Furthermore, the system calculates the proportion of each of the two types of losses in the total loss based on the statistical results, and then sets corresponding penalty weights according to the proportions. During the setting process, the system ensures that the penalty weights for overestimating and underestimating risks differ. For risk types with a higher loss proportion, the system sets a higher penalty weight to highlight the priority of prevention and control for such risks, avoiding insufficient prevention and control of key risks due to equal weights. Preferably, the system optimizes the risk transmission probability strategy and constructs an asymmetric loss function. This function incorporates the calculated penalty weights and risk thresholds. The risk threshold is determined by the system using a preset statistical method, which is based on a comprehensive calculation of indicators such as prediction accuracy, false negative rate, and false positive rate of historical risk events. The system can use a specified type of optimizer for gradient backpropagation. During the propagation process, the system adjusts various weight matrices in the model, including the factor matrix generated during tensor decomposition, the weight matrix in the graph attention network, and the parameters corresponding to the activation function. The system performs parameter adjustments through multiple iterations, calculating the loss function value after each iteration until the loss function value reaches the preset minimum threshold or the number of iterations reaches the upper limit. Through this process, the optimized risk transmission probability matrix and the optimal risk threshold are obtained.

[0067] Furthermore, the system compares each element in the optimized risk transmission probability matrix with the optimal risk threshold. If the element value is higher than the optimal risk threshold, the system determines the risk transmission path between the corresponding nodes to be a high-risk transmission path; if the element value is lower than the optimal risk threshold, the system determines the corresponding path to be a low-risk transmission path. Based on the determination results, the system generates structured decision instructions, which include information such as node ID, risk level, transmission probability value, and warning level. The instruction format is set to a standard format that can be interfacing with the enterprise's ERP system and supply chain management system, ensuring that the instructions can be recognized and executed by the enterprise's existing systems. For high-risk transmission paths, the system generates purchase order adjustment strategies and raw material reserve increase strategies. The purchase order adjustment strategy specifies the types of materials that need to be reduced from high-risk nodes and the adjustment range, while the raw material reserve increase strategy specifies the types and quantities of raw materials that need to be increased in reserves. For low-risk transmission paths, the system generates a strategy to maintain the original purchase plan, while determining that raw material reserves will be maintained at normal levels. The normal reserve quantity is calculated by the system using a safety stock model, which considers factors such as average daily consumption, purchase lead time, and demand fluctuation coefficient.

[0068] In summary, the big data prediction method based on machine learning provided in this application converts millisecond-level real-time power grid monitoring data into frequency domain features through wavelet packet transform, effectively overcoming the timeliness conflict between high-frequency physical data and low-frequency commercial data, and achieving accurate alignment between power grid transient fluctuation characteristics and enterprise operating status. By utilizing tensor direct product operations to fuse multi-source heterogeneous information such as power grid topology, enterprise carbon emissions, and financial data in a unified feature space, a joint representation that preserves spatiotemporal correlation can be constructed to achieve lossless integration of cross-domain data. Through the collaborative processing of policy and regulatory semantic parsing and tensor decomposition, unstructured policy clauses can be transformed into quantifiable constraint features, and by combining activation function mapping, a dynamic correlation between environmental policies and business activities can be constructed. In terms of space, it achieves accurate quantification of the implicit risk transmission effect; based on the iterative update mechanism of dynamic graph neural network, it can capture the real-time evolution law of supply chain node status and generate risk transmission probability distribution, so as to visualize the propagation path of risk in multi-level supply chain network; adopting the asymmetric loss function to optimize the strategy decision model, it can dynamically adjust the prediction weight according to the difference in the impact of overestimating and underestimating risk, and finally output supply chain control instructions with threshold judgment, which can achieve a closed-loop response from risk monitoring to decision execution. It systematically solves the technical bottlenecks of traditional models in cross-scale data fusion, policy transmission quantification and dynamic risk response, and significantly improves the accuracy of risk prediction and the timeliness of handling in the event of sudden policy adjustment or power grid anomaly.

[0069] In one embodiment, S1 of the big data prediction method based on machine learning provided by the present invention specifically includes the following steps:

[0070] S11: Select wavelet basis functions for the acquired real-time power grid monitoring data stream. Based on the frequency characteristics of the current waveform, select Morlet wavelet as the basis function to generate an optimized wavelet basis.

[0071] Specifically, the system extracts current waveform-related data from the real-time power grid monitoring data stream, and obtains the frequency distribution range, oscillation pattern, and correlation characteristics between the time and frequency domains of the current waveform through a frequency analysis process. Based on these characteristics, the system clarifies the adaptation requirements that the wavelet basis functions must meet, including the ability to track non-stationary current signals, the ability to distinguish different frequency components, and stability performance during the calculation process.

[0072] Furthermore, the system retrieves a pre-defined wavelet basis function library, which contains various wavelet basis functions commonly used in signal processing. The system extracts the characteristic data of each wavelet basis function in the library, covering time-domain support range, frequency-domain bandwidth, oscillation count, and decay rate. The system compares and analyzes the frequency characteristics of the current waveform with the characteristic data of each wavelet basis function, comparing dimensions including time-domain feature matching, frequency-domain resolution fit, and the ability to capture key components in the current signal. Through comparison, the system determines that the Morlet wavelet has the best fit with the frequency characteristics of the current waveform. The time-domain oscillation pattern of the Morlet wavelet is consistent with the periodic characteristics of the current waveform, and its frequency-domain bandwidth can accurately cover the frequency range of the fundamental wave, harmonics, and high-frequency noise in the current signal. Based on the determined Morlet wavelet, the system adjusts the parameters to ensure that the center frequency of the wavelet basis function matches the main frequency components of the current waveform. This adjustment process generates an optimized wavelet basis.

[0073] S12: Perform wavelet packet decomposition on the optimized wavelet basis and sampled real-time power grid monitoring data, calculate the scaling coefficients of the fundamental and harmonic components within the preset frequency band, and calculate the discrete wavelet coefficients of the high-frequency noise components to generate multi-resolution spectral features.

[0074] Specifically, the system decomposes the sampled real-time power grid monitoring data using an optimized wavelet basis. The system determines the wavelet packet decomposition level based on the frequency distribution of signal components in the real-time power grid monitoring data, ensuring that the decomposition level covers the frequency ranges corresponding to the fundamental wave, harmonics, and high-frequency noise. This avoids insufficient levels leading to undecomposed frequency components or excessive computational resource consumption due to too many levels. The system performs wavelet packet decomposition according to the determined level. During the decomposition process, the system performs multi-scale partitioning of the real-time power grid monitoring data by optimizing the wavelet basis. After each level of decomposition, the system divides the signal into multiple frequency band components, corresponding to the fundamental wave component, harmonic components, and high-frequency noise components, respectively.

[0075] Furthermore, the system calculates scaling coefficients for the fundamental and harmonic components. The calculation process follows the scaling space mapping rules of wavelet packet decomposition, which convert the time-domain signals of the fundamental and harmonics into scaling coefficients characterizing the low-frequency energy distribution. These scaling coefficients reflect key information such as the amplitude and phase of the fundamental and harmonics. Simultaneously, the system performs discrete wavelet coefficient calculations for the high-frequency noise components, following the wavelet space mapping rules of wavelet packet decomposition. These rules convert the transient characteristics of high-frequency noise into discrete wavelet coefficients, which can distinguish high-frequency noise from useful signal components. After completing the coefficient calculations, the system organizes the scaling coefficients and discrete wavelet coefficients according to the decomposition level and frequency order to form multi-resolution spectral features.

[0076] S13: The multi-resolution spectral features are reconstructed in the frequency domain. After separating the high-frequency noise components from the fundamental components, the energy density matrix is ​​reconstructed to generate the time-frequency matrix.

[0077] Specifically, the system defines frequency limits based on the current waveform frequency characteristic dataset. The definition is based on the highest frequency values ​​of the fundamental and harmonic components in the historical current waveforms. These highest frequency values ​​are used as frequency limits to ensure that the fundamental component frequency range includes both fundamental and harmonic components, while the high-frequency noise component frequency range exceeds this range. The system then classifies the frequency bands in the multi-resolution spectral features according to these frequency limits, marking those frequency bands within the high-frequency noise component range as noise data and those within the fundamental component range as valid data, thus completing the separation.

[0078] After separation, the system determines the matrix dimensions. The row dimension corresponds to the time window, and the time window division criteria are consistent with the time window used during decomposition to ensure the continuity of the time dimension. The column dimension corresponds to the separated effective frequency components, and the number of effective frequency components is determined by the number of frequency bands corresponding to the fundamental component. The system calculates the energy value of each effective frequency component under each time window, calculates the scaling factor corresponding to that frequency component using a sum of squares operation, and the calculation range covers all scaling factor data within the time window. The sum of squares operation follows numerical calculation rules. The system fills the matrix with the energy values ​​according to the correspondence between the time window and the effective frequency components, forming a preliminary energy density matrix. The system performs an integrity check on the preliminary matrix to check for missing matrix elements and whether the energy values ​​are within the reasonable range determined by historical energy data. If missing or abnormal elements are found, the system recalculates the corresponding energy values. After the integrity check passes, the system performs standardization processing on the energy density matrix by dividing each element value of the matrix by the maximum value of all elements to ensure that the element values ​​are within a uniform numerical range. After the standardization processing is completed, the system generates a time-frequency matrix, which reflects the energy distribution characteristics of the real-time power grid monitoring data in the time and frequency dimensions.

[0079] In one embodiment, S2 of the big data prediction method based on machine learning provided by the present invention specifically includes the following steps:

[0080] S21: Perform periodic segmentation on the time-frequency matrix. According to the carbon emission reporting period in the acquired enterprise carbon emission data, the continuous time-frequency data is divided into monthly data blocks of equal length to generate a periodic submatrix.

[0081] Specifically, the system retrieves enterprise carbon emission data and extracts carbon emission reporting cycle information from it. This cycle information represents the fixed time interval for enterprises to submit carbon emission reports, and the system uses this cycle as the time basis for segmenting the time-frequency matrix. Further, the system segments the continuous time-frequency data in the time-frequency matrix. Before segmentation, the system first determines the time dimension range of the time-frequency matrix to ensure that this range fully covers the time span corresponding to the enterprise's carbon emission reporting cycle. The system divides the continuous time-frequency data into multiple data blocks according to the duration of the carbon emission reporting cycle. During the segmentation process, the system ensures that the duration of each data block is equal, and that the time interval corresponding to each data block corresponds one-to-one with the time interval of the enterprise's carbon emission report, avoiding uneven data block durations or misaligned time intervals.

[0082] Preferably, after segmentation, the system performs integrity verification on each data block. The verification includes checking for missing data and ensuring that the time and frequency dimensions of the data block fully retain the original characteristics of the time-frequency matrix. If missing data is found, the system uses interpolation algorithms to supplement the missing data, ensuring the integrity of the data block. After completing the segmentation and verification of all data blocks, the system organizes these data blocks into periodic sub-matrices, with each periodic sub-matrix corresponding to the time-frequency data within a company's carbon emission reporting period.

[0083] S22: Perform feature stitching on the periodic submatrix and corporate carbon emission data, expand the carbon emission intensity values ​​into a feature plane with the same dimension as the submatrix, and generate a spatiotemporal fundamental tensor.

[0084] Specifically, the system extracts carbon emission intensity values ​​from enterprise carbon emission data. During extraction, the system ensures that the carbon emission intensity values ​​are consistent with the time interval of the periodic submatrix, meaning each periodic submatrix corresponds to a set of carbon emission intensity values ​​for the same period. Further, the system performs dimensional expansion on the carbon emission intensity values, aiming to make the dimensions of the carbon emission intensity values ​​the same as the dimensions of the periodic submatrix. The system determines the dimensional composition of the periodic submatrix, including time, frequency, and feature dimensions. Based on these dimensional parameters, a single carbon emission intensity value is expanded into a multi-dimensional feature plane. During the expansion process, the system follows dimensional mapping rules to ensure that each dimension of the feature plane matches the corresponding dimension of the periodic submatrix, avoiding dimensional incompatibility issues.

[0085] After expanding the dimensionality of the carbon emission intensity values, the system performs a concatenation operation between the generated feature plane and the corresponding periodic submatrix. The concatenation direction is the feature dimension direction, meaning the feature plane is added as a new feature layer and superimposed on the feature dimension of the periodic submatrix. After concatenation, the system performs a dimensionality consistency check on the result to confirm that the concatenated matrix has no conflicts in the time, frequency, and feature dimensions. If the check passes, the system integrates the concatenation result into a spatiotemporal fundamental tensor, which simultaneously contains both time-frequency data features and carbon emission data features.

[0086] S23: Based on the spatiotemporal fundamental tensor, the financial features and spatial relationships of the acquired enterprise financial data and the power grid topology adjacency matrix of the power grid topology database are fused. The quarterly difference of the asset-liability ratio is calculated and encoded as an additional feature layer. At the same time, the power grid topology adjacency matrix is ​​superimposed as a spatial relationship layer to generate a spatiotemporal joint tensor. The spatiotemporal joint tensor is used to indicate the correlation features between business activities and environmental constraints between power grid nodes.

[0087] Specifically, the system extracts the debt-to-asset ratio from corporate financial data, calculates the quarterly difference of the debt-to-asset ratio based on the debt-to-asset ratio values ​​of two adjacent quarters, and reflects the quarterly trend of the debt-to-asset ratio through difference calculation. Subsequently, the system encodes the calculated quarterly difference of the debt-to-asset ratio. The encoding goal is to convert the difference result into an additional feature layer conforming to the tensor feature layer format. During the encoding process, the system uses a feature standardization method to ensure that the numerical range of the additional feature layer is consistent with the feature numerical range of the spatiotemporal underlying tensor, avoiding the impact of differences in numerical range on the fusion effect.

[0088] After generating the additional feature layer, the system processes the power grid topology adjacency matrix in the power grid topology database. First, the system confirms the association between power grid nodes and supply chain nodes corresponding to the power grid topology adjacency matrix, ensuring that the adjacency matrix accurately reflects the connection status between relevant power grid nodes in the supply chain. The system then superimposes the power grid topology adjacency matrix into a spatial relationship layer. During the superposition process, the system adjusts the dimensions of the spatial relationship layer according to the spatial dimension parameters of the spatiotemporal fundamental tensor, ensuring that the node dimensions of the spatial relationship layer match the power grid node dimensions in the spatiotemporal fundamental tensor. The system then superimposes the additional feature layer and the spatial relationship layer onto the spatiotemporal fundamental tensor, respectively. The additional feature layer is superimposed onto the feature dimension, and the spatial relationship layer is superimposed onto the spatial dimension. After superposition, the system performs a fusion verification, confirming that the fused tensor completely retains the key features of each data source in the time, spatial, frequency, and feature dimensions, without any feature loss or conflict.

[0089] In one embodiment, S3 of the big data prediction method based on machine learning provided by the present invention specifically includes the following steps:

[0090] S31: Obtain publicly available policy and regulatory texts from government agencies, analyze the legal clauses of these texts, extract key constraints such as carbon emission limits and penalty clauses, and generate structured policy features.

[0091] Specifically, after obtaining publicly available policy and regulatory texts from government agencies, the system initiates a legal clause parsing process. The system first performs text preprocessing on the policy and regulatory texts, sequentially completing word segmentation, stop word removal, and legal entity identification. The identified objects include key constraints related to supply chain risks, such as "carbon emission limits," "penalty clauses," and "industry regulatory requirements." The system then quantifies the identified key constraints. For constraints like "carbon emission limits," the system converts them into numerical forms based on the explicit limits in the text; for constraints like "penalty clauses," the system converts them into corresponding quantitative indicators based on descriptions such as penalty amounts and frequency. The system then organizes all quantified key constraints into a structured policy feature matrix, constructed according to the following formula:

[0092]

[0093] in, Represents the structured policy feature matrix. This represents the quantized value of the i-th key constraint, and k represents the total number of key constraints. After generating the matrix, the system performs a consistency check on the values ​​within the matrix to ensure that the quantized values ​​of different types of constraints are comparable in the same dimension. After the check passes, structured policy features are generated.

[0094] S32: Perform multidimensional feature compression processing on the spatiotemporal joint tensor, identify periodic change patterns in the time dimension, extract node distribution features in the spatial dimension, screen key correlation parameters in the indicator dimension, reconstruct the data space by retaining principal components whose variance contribution rate in each dimension exceeds the set threshold, and generate core factor tensors.

[0095] Specifically, the system first analyzes the spatiotemporal joint tensor in the time dimension, identifying periodic change patterns in the data through time series analysis to determine the main fluctuation cycles of the data in the time dimension. In the spatial dimension, the system extracts the distribution characteristics of power grid nodes in the tensor, clarifying the correlation density and distribution patterns between nodes. In the indicator dimension, the system screens key parameters with high correlation to supply chain risks and removes redundant parameters with weak correlation. The system then uses principal component analysis to compress the multidimensional features. First, it calculates the covariance matrix of each dimension of features, then solves for the eigenvalues ​​and eigenvectors of the covariance matrix, and calculates the variance contribution rate of each principal component using the following formula:

[0096]

[0097] in, Indicates the first The variance contribution rate of each principal component Indicates the first 1 eigenvalue, This represents the total number of eigenvalues. The system retains principal components whose variance contribution rate exceeds a set threshold and reconstructs the data space using the following formula:

[0098]

[0099] in, Represents the core factor tensor. Indicates the first The vectors corresponding to the principal components. Indicates the first The loading matrix of each principal component, This represents the set variance contribution rate threshold, which generates the core factor tensor after reconstruction.

[0100] S33: Perform cross-modal fusion processing on core factor tensors and structured policy features, map them to a unified vector space through a fully connected layer and standardize them to generate a unified risk embedding vector.

[0101] Specifically, after obtaining the core factor tensor and structured policy features, the system performs modal expansion on the core factor tensor, converting the multidimensional tensor into a two-dimensional matrix form, denoted as... Simultaneously confirm the structured policy feature matrix. The dimension ensures its consistency with Maintain consistency across sample dimensions. The system uses a matrix vertical concatenation operation to... and By merging, we obtain the joint characteristic matrix. The symbol "|" represents the vertical concatenation operation of matrices. The system inputs the joint feature matrix into the fully connected layer and generates a unified risk embedding vector using the following formula:

[0102]

[0103] in, The output is a unified risk embedding vector. This is the weight matrix of the fully connected layer, used to perform linear transformations on the joint feature matrix. This is the bias vector for the fully connected layer, used to adjust the offset of the linear transformation result. To modify the activation function of the linear unit, a nonlinear transformation is introduced to enhance the feature representation capability. The representation layer normalization operation is used to normalize the output of the activation function and avoid gradient anomalies.

[0104] In one embodiment, step S4 of the big data prediction method based on machine learning provided by the present invention specifically includes the following steps:

[0105] S41: Perform dynamic graph initialization on the unified risk embedding vector, construct an initial adjacency matrix containing enterprise nodes and grid nodes based on the grid topology connection relationship, and generate a spatiotemporal constraint graph.

[0106] Specifically, the system retrieves power grid topology connection data and identifies two types of core nodes from the data: power grid nodes (such as substations and transmission line nodes) and enterprise nodes (such as raw material suppliers, manufacturers, and downstream distributors), integrating them to form a node set. This set contains all nodes involved in the supply chain and power grid connections, where N is the total number of nodes. The system constructs an initial adjacency matrix based on the node set and the power grid topology connections. Elements within the matrix This represents the initial connection state between node i and node j. When node i and node j have a direct topological or business association, the element at the corresponding position in the matrix is ​​1; when node i and node j have no direct association, the element at the corresponding position is 0. Specifically, whether there is an association between power grid nodes is determined by whether their topological lines are connected; whether there is an association between enterprise nodes and power grid nodes is determined by their power supply or power consumption relationship; and whether there is an association between enterprise nodes is temporarily determined by their basic business cooperation relationship. The system then combines the time dimension features in the unified risk embedding vector with the initial adjacency matrix, and through timestamp matching, ensures that each node connection corresponds to the risk features of a specific time period, ultimately generating a spatiotemporal constraint graph. This graph contains the node set, initial connection relationships, and spatiotemporal constraint information.

[0107] S42: Dynamically calculate edge weights for the spatiotemporal constraint graph, update the connection strength between enterprise nodes based on the transaction frequency between enterprises and the similarity of risk embedding vectors, and generate an adaptive adjacency matrix.

[0108] Specifically, the system first retrieves historical transaction data between enterprises, extracts the transaction frequency data of enterprise node pairs (\( v_i, v_j \)), and normalizes the transaction frequency using the following formula:

[0109]

[0110] in, This represents the original transaction frequency between enterprise nodes i and j. This represents the maximum transaction frequency for all enterprise node pairs. The normalized transaction frequency ranges from 0 to 1. The system also calculates the risk embedding vector similarity of enterprise node pairs using the cosine similarity formula:

[0111]

[0112] in, , Let i and j be the unified risk embedding vectors for enterprise nodes i and j, respectively. Let L2 norm represent the vector. The system updates edge weights using a weighted fusion formula:

[0113]

[0114] in, Weighting coefficients are assigned to balance the impact of transaction frequency and risk similarity in the initial adjacency matrix. Replace the non-zero elements in with the corresponding Generate an adaptive adjacency matrix This matrix can reflect the dynamic changes in the connection strength between enterprise nodes in real time.

[0115] S43: Perform graph convolution propagation processing on the adaptive adjacency matrix, and iteratively update the hidden state of enterprise nodes in the supply chain network through gated cyclic units to generate risk transmission probability. The risk transmission probability is used to indicate the likelihood of risk being transmitted from upstream enterprises to downstream enterprises in the supply chain network.

[0116] Specifically, the system adapts the adjacency matrix Perform a normalization operation to eliminate the impact of node degree differences using the following formula:

[0117]

[0118] in, For degree matrix, , representing the sum of edge weights connecting node i. The system inputs the normalized matrix into the graph convolutional layer, combines it with the unified risk embedding vector as the initial node features, and iteratively updates the hidden state of the enterprise nodes through a gated recurrent unit (GRU). The iteration formula is:

[0119]

[0120] in, Let be the hidden state of enterprise node i in the t-th iteration. This is the hidden state from the previous iteration. Here is the weight matrix of the GRU. Here is the bias vector of the GRU. This represents message passing from neighboring nodes to the current node. After the system completes a preset number of iterations, it calculates the risk propagation probability based on the final hidden state, using the following formula:

[0121]

[0122] in, Let be the probability that risk is transmitted from enterprise node i to node j. , For the hidden state of the final iteration, The probability mapping weight matrix, This is the probability mapping bias vector. This is the sigmoid activation function, which maps the output to a range between 0 and 1. (System compilation of all) Generate a risk transmission probability set to indicate the likelihood of risk being transmitted from upstream to downstream enterprises in the supply chain network.

[0123] In one embodiment, S5 of the big data prediction method based on machine learning provided by the present invention specifically includes the following steps:

[0124] S51: Perform asymmetric loss calculation on the risk transmission probability, set different penalty weights based on the difference in the impact of overestimating and underestimating risk on the supply chain, and generate an optimized objective function.

[0125] Specifically, the system retrieves historical risk event records from the supply chain and separates them into two categories: underestimated risk events (where the predicted probability of the risk actually occurred is lower than the actual probability) and overestimated risk events (where the predicted probability of the risk did not occur is higher than the actual probability). The system then calculates the impact of these two types of events on the supply chain, including order default costs and production stoppage losses due to underestimated risks, and raw material backlog costs and capital tied up losses due to overestimated risks. Based on the statistical results, corresponding penalty weights are assigned, with the penalty weight for underestimated risks denoted as […]. Overestimating the risk penalty weight is recorded as ,and This reflects the higher priority of risk prevention and control, which is intended to mitigate underestimation of risk. The system then constructs an optimization objective function that minimizes asymmetric loss, expressed as:

[0126]

[0127] in, This indicates the total number of risk transmission paths. This is the true risk label for the i-th path, where 1 indicates that the risk has occurred and 0 indicates that the risk has not occurred. Predict the probability of risk transmission for the i-th path. This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. For losses that underestimate risk, Losses due to overestimating risk.

[0128] S52: Perform gradient backpropagation on the objective function to update the prediction model parameters through the optimization algorithm and generate a robust prediction model.

[0129] Specifically, the system selects a suitable optimization algorithm that supports adaptive learning rate adjustment and momentum optimization to balance parameter update speed and stability. The system calculates the partial derivative of the objective function with respect to all trainable parameters in the prediction model, obtaining the gradient values ​​for each parameter. The gradient direction indicates the direction of parameter adjustment (a positive gradient indicates a decrease in parameters, and a negative gradient indicates an increase in parameters). The system performs parameter updates based on the gradient values, using the following formula:

[0130]

[0131] in, The updated model parameters include graph convolutional layer weights, gated recurrent unit biases, probability mapping matrices, etc. The parameters before the update. For adaptive learning rate, This is the gradient term driving the equation. The system repeatedly executes the "gradient calculation-parameter update" process. After each iteration, it calculates the current value of the objective function. When the difference in function values ​​across multiple consecutive iterations is less than a preset fluctuation threshold, the model is considered to have converged. After convergence, the system saves the final parameter configuration and generates a robust prediction model. This model can stably output the risk transmission probability under different supply chain scenarios.

[0132] S53: Make strategic decisions based on the risk transmission probability output by the robust prediction model. When the transmission probability exceeds the preset risk threshold, identify key risk transmission paths and generate multi-level response instructions, and integrate them to generate supply chain risk control strategies.

[0133] Specifically, the system first sets a preset risk threshold, which is determined based on the supply chain's risk tolerance and must cover the division criteria of "acceptable risk range" and "intervention-required risk range." The system compares the probability of each risk transmission path with the preset threshold: when the probability exceeds the threshold, it is determined as a high-risk path; when the probability is below the threshold, it is determined as a low-risk path. The system performs critical path identification on high-risk paths, and uses a ranking algorithm to select the several paths with the highest probability values. These paths are the critical risk transmission channels in the supply chain (the channels through which risk is most easily transmitted from upstream to downstream).

[0134] Furthermore, the system generates multi-level response instructions based on the risk level of the critical path: for the highest-risk path, a Level 1 response instruction is generated, including emergency adjustments to purchase orders and increasing the reserve of core raw materials; for medium-risk paths, a Level 2 response instruction is generated, including delaying the delivery of non-core orders and verifying the risk status of suppliers; and for low-risk paths, a Level 3 response instruction is generated, including increasing the frequency of risk monitoring and maintaining regular procurement reserves. The system integrates all response instructions, supplementing execution details such as the proportion of purchase order adjustments and the raw material reserve cycle, to form a complete supply chain risk control strategy. This strategy clearly indicates the operational direction of each link in the supply chain and can be directly used to guide the purchasing department to adjust orders and the warehousing department to adjust raw material reserves, achieving synergy between risk control and supply chain operations.

[0135] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0136] Based on the same inventive concept, this application also provides a machine learning-based big data prediction apparatus for implementing the machine learning-based big data prediction method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more machine learning-based big data prediction apparatus embodiments provided below can be found in the limitations of the machine learning-based big data prediction method described above, and will not be repeated here.

[0137] Preferably, such as Figure 3 As shown, the present invention provides a big data prediction device 600 based on machine learning, which is configured with the following modules:

[0138] The power grid time-frequency feature extraction module 610 is used to extract time-frequency features from the acquired real-time power grid monitoring data stream. It converts the sampled real-time power grid monitoring data into a frequency domain feature representation through wavelet packet transform, and generates a time-frequency matrix.

[0139] The cross-scale data fusion module 620 is used to perform cross-scale fusion of acquired corporate carbon emission data, corporate financial data and power grid topology adjacency matrix from power grid topology database based on time-frequency matrix. It aligns data of different time granularities in a unified feature space through tensor direct product operation to generate spatiotemporal joint tensor.

[0140] The joint feature transformation embedding module 630 is used to obtain publicly available policy and regulatory texts from government agencies, perform joint feature transformations by combining spatiotemporal joint tensors, decompose the tensors into low-dimensional factor matrices and convert the policy texts into word vector matrices, and after concatenation, map them to a unified space based on an activation function to construct a correlation space between carbon emission indicators and financial indicators, and generate a unified risk embedding vector.

[0141] The risk transmission modeling module 640 is used to model risk transmission on a unified risk embedding vector. It iteratively updates the state of supply chain nodes and calculates the transmission probability distribution through a dynamic graph neural network to generate risk transmission probability.

[0142] The risk control strategy optimization module 650 is used to optimize the risk transmission probability, calculate the difference in penalty weights between overestimating and underestimating risk, adjust model parameters through gradient backpropagation, output decision instructions with threshold judgment, and generate supply chain risk control strategies. The supply chain risk control strategies are used to indicate adjustments to purchase orders or raw material reserves.

[0143] Preferably, the power grid time-frequency feature extraction module 610 provided in this application is configured with the following units:

[0144] The wavelet basis function selection unit is used to select wavelet basis functions for the acquired real-time power grid monitoring data stream. Based on the frequency characteristics of the current waveform, the Morlet wavelet is selected as the basis function to generate an optimized wavelet basis.

[0145] The wavelet packet decomposition processing unit is used to perform wavelet packet decomposition processing on the optimized wavelet basis and sampled real-time power grid monitoring data. It calculates the scaling coefficients of the fundamental and harmonic components and the discrete wavelet coefficients of the high-frequency noise components within a preset frequency band, generating multi-resolution spectral features.

[0146] The frequency domain feature reconstruction unit is used to reconstruct the frequency domain of multi-resolution spectral features, and reconstruct the energy density matrix after separating the high-frequency noise component from the fundamental component to generate the time-frequency matrix.

[0147] Preferably, the cross-scale data fusion module 620 provided in this application is configured with the following units:

[0148] The time-frequency matrix periodic segmentation unit is used to perform periodic segmentation processing on the time-frequency matrix. According to the carbon emission reporting period in the acquired enterprise carbon emission data, the continuous time-frequency data is divided into equal-length monthly data blocks to generate a periodic submatrix.

[0149] The carbon emission feature stitching unit is used to perform feature stitching processing on periodic submatrices and corporate carbon emission data, expanding the carbon emission intensity values ​​into a feature plane with the same dimension as the submatrices, and generating a spatiotemporal fundamental tensor.

[0150] The financial-spatial feature fusion unit is used to fuse financial features and spatial relationships between the acquired corporate financial data and the power grid topology adjacency matrix of the power grid topology database based on the spatiotemporal fundamental tensor. It calculates the quarterly difference of the debt-to-equity ratio and encodes it as an additional feature layer. At the same time, it superimposes the power grid topology adjacency matrix as a spatial relationship layer to generate a spatiotemporal joint tensor. The spatiotemporal joint tensor is used to indicate the correlation features between business activities and environmental constraints between power grid nodes.

[0151] Preferably, the joint feature transformation embedding module 630 provided in this application is configured with the following units:

[0152] The policy feature structure extraction unit is used to obtain publicly available policy and regulatory texts from government agencies, analyze the legal clauses of the policy and regulatory texts, extract key constraints such as carbon emission limits and penalty clauses, and generate structured policy features.

[0153] The multidimensional feature compression and reconstruction unit is used to perform multidimensional feature compression processing on the spatiotemporal joint tensor. It identifies periodic change patterns in the time dimension, extracts node distribution features in the spatial dimension, and filters key correlation parameters in the indicator dimension. It reconstructs the data space by retaining principal components whose variance contribution rate in each dimension exceeds a set threshold, and generates core factor tensors.

[0154] The cross-modal feature fusion unit is used to perform cross-modal fusion processing on core factor tensors and structured policy features. It maps to a unified vector space through a fully connected layer and standardizes the data to generate a unified risk embedding vector.

[0155] Preferably, the risk transmission modeling module 640 provided in this application is configured with the following units:

[0156] The dynamic graph initialization building unit is used to perform dynamic graph initialization on the unified risk embedding vector. It constructs an initial adjacency matrix containing enterprise nodes and grid nodes based on the grid topology connection relationship and generates a spatiotemporal constraint graph.

[0157] The edge weight dynamic update unit is used to dynamically calculate the edge weights of the spatiotemporal constraint graph, update the connection strength between enterprise nodes based on the transaction frequency between enterprises and the similarity of risk embedding vectors, and generate an adaptive adjacency matrix.

[0158] The graph convolutional risk propagation unit is used to perform graph convolutional propagation processing on the adaptive adjacency matrix. It iteratively updates the hidden state of enterprise nodes in the supply chain network through the gated loop unit to generate risk propagation probability. The risk propagation probability is used to indicate the likelihood of risk being transmitted from upstream enterprises to downstream enterprises in the supply chain network.

[0159] Preferably, the risk control strategy optimization module 650 provided in this application is configured with the following units:

[0160] The asymmetric loss calculation unit is used to perform asymmetric loss calculation on the risk transmission probability. It sets different penalty weights based on the difference in the impact of overestimating and underestimating risk on the supply chain and generates an optimization objective function.

[0161] The gradient backpropagation optimization unit is used to perform gradient backpropagation on the objective function, update the prediction model parameters through the optimization algorithm, and generate a robust prediction model.

[0162] The risk probability strategy decision unit is used to make strategy decisions based on the risk transmission probability output by the robust prediction model. When the transmission probability exceeds the preset risk threshold, it identifies key risk transmission paths and generates multi-level response instructions, which are then integrated to generate supply chain risk control strategies.

[0163] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described machine learning-based big data prediction method.

[0164] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described machine learning-based big data prediction method.

[0165] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0166] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0167] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A big data prediction method based on machine learning, characterized in that, Includes the following steps: S1: Extract time-frequency features from the acquired real-time power grid monitoring data stream, and convert the sampled real-time power grid monitoring data into a frequency domain feature representation through wavelet packet transform to generate a time-frequency matrix; S2: Based on the time-frequency matrix, cross-scale fusion is performed on the acquired corporate carbon emission data, corporate financial data, and the power grid topology adjacency matrix from the power grid topology database. The data at different time granularities are aligned in a unified feature space through tensor direct product operation to generate a spatiotemporal joint tensor. S3: Obtain publicly available policy and regulatory texts from government agencies, perform joint feature transformations using the spatiotemporal joint tensor, decompose the tensor into a low-dimensional factor matrix, and convert the policy text into a word vector matrix. After concatenation, map the vectors to a unified space based on an activation function to construct a correlation space between carbon emission indicators and financial indicators, and generate a unified risk embedding vector. S4: Perform risk transmission modeling on the unified risk embedding vector, iteratively update the supply chain node status through a dynamic graph neural network and calculate the transmission probability distribution to generate the risk transmission probability; S5: Optimize the risk transmission probability, calculate the difference in penalty weights between overestimating and underestimating risk, adjust the model parameters through gradient backpropagation, output decision instructions with threshold judgment, and generate a supply chain risk control strategy. The supply chain risk control strategy is used to indicate the adjustment of purchase orders or the reserve of raw materials.

2. The method according to claim 1, characterized in that, S1 includes: S11: Select wavelet basis functions for the acquired real-time power grid monitoring data stream. Select Morlet wavelet as the basis function based on the frequency characteristics of the current waveform to generate an optimized wavelet basis. S12: Perform wavelet packet decomposition on the optimized wavelet basis and sampled real-time power grid monitoring data, calculate the scaling coefficients of the fundamental and harmonic components within a preset frequency band, and calculate the discrete wavelet coefficients of the high-frequency noise components to generate multi-resolution spectral features. S13: The multi-resolution spectral features are reconstructed in the frequency domain, and the energy density matrix is ​​reconstructed after separating the high-frequency noise component from the fundamental component to generate the time-frequency matrix.

3. The method according to claim 1, characterized in that, S2 includes: S21: Perform periodic segmentation processing on the time-frequency matrix, and divide the continuous time-frequency data into equal-length monthly data blocks according to the carbon emission reporting period in the acquired enterprise carbon emission data to generate a periodic sub-matrix; S22: Perform feature concatenation processing on the periodic submatrix and the enterprise carbon emission data, expand the carbon emission intensity value into a feature plane with the same dimension as the submatrix, and generate a spatiotemporal fundamental tensor; S23: Based on the spatiotemporal fundamental tensor, the financial features and spatial relationships of the acquired corporate financial data and the power grid topology adjacency matrix of the power grid topology database are fused. The quarterly difference of the asset-liability ratio is calculated and encoded as an additional feature layer. At the same time, the power grid topology adjacency matrix is ​​superimposed as a spatial relationship layer to generate a spatiotemporal joint tensor. The spatiotemporal joint tensor is used to indicate the correlation features between business activities and environmental constraints between power grid nodes.

4. The method according to claim 1, characterized in that, S3 includes: S31: Obtain publicly available policy and regulatory texts from government agencies, analyze the legal clauses of the policy and regulatory texts, extract key constraints such as carbon emission limits and penalty clauses, and generate structured policy features; S32: Perform multidimensional feature compression processing on the spatiotemporal joint tensor, identify periodic change patterns in the time dimension, extract node distribution features in the spatial dimension, screen key correlation parameters in the index dimension, reconstruct the data space by retaining principal components whose variance contribution rate in each dimension exceeds a set threshold, and generate core factor tensors. S33: Perform cross-modal fusion processing on the core factor tensor and the structured policy features, map them to a unified vector space through a fully connected layer and standardize them to generate a unified risk embedding vector.

5. The method according to claim 4, characterized in that, The expression for the unified risk embedding vector is: in, The output is a unified risk embedding vector. This is the matrix representation of the core factor tensor after modal expansion. For the structured policy feature matrix, This is the weight matrix of the fully connected layer. For the bias vector of the fully connected layer, To modify the activation function of the linear unit, Presentation layer standardized operations, This indicates a vertical matrix concatenation operation.

6. The method according to claim 1, characterized in that, S4 includes: S41: Perform dynamic graph initialization on the unified risk embedding vector, construct an initial adjacency matrix containing enterprise nodes and power grid nodes based on the power grid topology connection relationship, and generate a spatiotemporal constraint graph; S42: Dynamically calculate the edge weights of the spatiotemporal constraint graph, update the connection strength between enterprise nodes based on the transaction frequency between enterprises and the similarity of risk embedding vectors, and generate an adaptive adjacency matrix; S43: Perform graph convolution propagation processing on the adaptive adjacency matrix, and iteratively update the hidden state of enterprise nodes in the supply chain network through gated loop units to generate risk transmission probability. The risk transmission probability is used to indicate the likelihood of risk being transmitted from upstream enterprises to downstream enterprises in the supply chain network.

7. The method according to any one of claims 1-6, characterized in that, S5 includes: S51: Perform asymmetric loss calculation on the risk transmission probability, set different penalty weights according to the difference in the impact of overestimating and underestimating risk on the supply chain, and generate an optimization objective function; S52: Perform gradient backpropagation on the optimization objective function, update the prediction model parameters through the optimization algorithm, and generate a robust prediction model; S53: Make strategy decisions on the risk transmission probability output by the robust prediction model. When the transmission probability exceeds the preset risk threshold, identify key risk transmission paths and generate multi-level response instructions, and integrate them to generate supply chain risk control strategies.

8. A big data prediction device based on machine learning, characterized in that, The device includes: The power grid time-frequency feature extraction module is used to extract time-frequency features from the acquired real-time power grid monitoring data stream. It converts the sampled real-time power grid monitoring data into a frequency domain feature representation through wavelet packet transform, and generates a time-frequency matrix. The cross-scale data fusion module is used to perform cross-scale fusion of acquired corporate carbon emission data, corporate financial data and power grid topology adjacency matrix from the power grid topology database based on the time-frequency matrix. It aligns data of different time granularities in a unified feature space through tensor direct product operation to generate a spatiotemporal joint tensor. The joint feature transformation embedding module is used to obtain publicly available policy and regulatory texts from government agencies, perform joint feature transformations in conjunction with the spatiotemporal joint tensor, decompose the tensor into a low-dimensional factor matrix and convert the policy text into a word vector matrix, and after concatenation, map it to a unified space based on an activation function to construct a correlation space between carbon emission indicators and financial indicators, and generate a unified risk embedding vector. The risk transmission modeling module is used to model the risk transmission of the unified risk embedding vector, and to generate the risk transmission probability by iteratively updating the state of the supply chain nodes and calculating the transmission probability distribution through a dynamic graph neural network. The risk control strategy optimization module is used to optimize the risk transmission probability, calculate the difference in penalty weights between overestimating and underestimating risk, adjust model parameters through gradient backpropagation, output decision instructions with threshold judgment, and generate supply chain risk control strategies. The supply chain risk control strategies are used to indicate adjustments to purchase orders or raw material reserves.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images