Search

CN-121998417-A - Artificial intelligence-based ESG risk prediction and assessment method and system

CN121998417ACN 121998417 ACN121998417 ACN 121998417ACN-121998417-A

Abstract

The application relates to the technical field of artificial intelligence, discloses an ESG risk prediction and evaluation method and system based on artificial intelligence, and aims to solve the problems of hysteresis, unilateralness and unexplainability of the existing ESG evaluation. The method comprises the steps of fusing multi-source heterogeneous ESG data to construct a dynamic heterogram, modeling risk propagation and time-varying characteristics by utilizing a dynamic graph neural network and a time sequence attention mechanism, introducing causal reasoning to eliminate confounding deviation, deducing the influence of a quantization factor through the counterfactual, and finally outputting interpretable risk grade and evidence chain. The system comprises a multi-source data access module, a standardized alignment module, a dynamic diagram construction module, a diagram embedding module, a time sequence weighting module, a causal correction module, a counterfactual deduction module and a risk rating module. The application realizes real-time, accurate and interpretable evaluation of the ESG risk of enterprises, and remarkably improves the risk identification and decision support capability.

Inventors

  • Liu Diefang

Assignees

  • 中财碳融(北京)科技有限公司

Dates

Publication Date
20260508
Application Date
20260116

Claims (10)

  1. 1. An artificial intelligence-based ESG risk prediction and assessment method, comprising: Acquiring original ESG related data of a target enterprise, processing the original ESG related data to convert the original ESG related data into unified vector representation data, and establishing a cross-modal data index for the vector representation data according to a corresponding enterprise entity identifier; Constructing a dynamic abnormal pattern structure based on the vector representation data, and performing multi-round message transfer and aggregation operation on the dynamic abnormal pattern structure by utilizing a dynamic pattern neural network model to generate an embedded vector of each node in a current time window; Performing time sequence modeling on the embedded vector, and calculating time weight distribution of contribution degree of each time node to the enterprise whole ESG risk; And calculating comprehensive ESG risk scores of target enterprises according to the embedded vectors and the time weight distribution, and dividing risk grades according to a preset threshold interval to obtain a main evidence chain causing the grading result and a propagation path of the main evidence chain in a graph structure.
  2. 2. The artificial intelligence based ESG risk prediction and assessment method of claim 1, wherein the processing the original ESG related data to convert the original ESG related data into unified vector representation data specifically includes: the original ESG related data comprises supervision punishment records, supply chain public opinion texts, carbon emission monitoring indexes, staff satisfaction survey results, board composition information and industry benchmark comparison parameters; For the numerical data in the original ESG related data, a Z fraction normalization method is adopted to map the numerical data to a zero mean unit variance space; for category data in the original ESG related data, semantic expansion is carried out by combining single thermal coding with a domain knowledge ontology base; For text data in the original ESG related data, invoking a pre-training language model to generate a context-aware sentence vector, and mapping the context-aware sentence vector into a predefined ESG topic label system through cosine similarity matching; And adding a unified time stamp field and an enterprise unique identification code to all the processed data to form vector representation data with space-time coordinates.
  3. 3. The artificial intelligence based ESG risk prediction and assessment method of claim 2, wherein constructing a dynamic heterogeneous map structure based on the vector representation includes: Establishing edges according to actual business or event relations among entities, wherein the types of the edges comprise stock right control, contract performance, administrative punishment or public opinion association; And calculating the initial weight of the edge as the product of the event severity score and the relation compactness between the main bodies, and updating the aging decay function in an exponential decay form day by day, wherein the half-life parameter of the aging decay function is set to be one hundred eighty days.
  4. 4. The artificial intelligence based ESG risk prediction and assessment method of claim 3, wherein the dynamic iso-composition structure is subjected to multiple rounds of message passing and aggregation operations using a dynamic graph neural network model, and specifically comprising: Configuring independent learnable weight matrixes for the edges of each relation type; In each round of aggregation process, carrying out weighted summation on embedded vectors of neighbor nodes according to attention coefficients after linear transformation of corresponding relation types, wherein the attention coefficients are obtained by normalizing embedded vector inner products of source nodes and target nodes through a softmax function; after multiple rounds of message passing, the output of each round is spliced or weighted averaged to generate an embedded vector of each node in the current time window.
  5. 5. The artificial intelligence based ESG risk prediction and assessment method of claim 4, wherein applying time series modeling to the embedded vector specifically includes: Setting a gating circulation unit network, wherein the hidden state dimension of the gating circulation unit network is set to be two hundred fifty-six dimensions, and the daily node embedded vector is set as an input value; inputting a hidden state sequence output by a gating circulation unit network into a multi-head self-attention mechanism comprising eight parallel heads, wherein the dimensions of a query, a key and a value projection matrix of each head are thirty-two dimensions; And after the outputs of the heads are spliced, the outputs are connected with residual errors through layer normalization, and then a time weight vector with the length of three hundred sixty is generated through a full-connection layer, wherein the time weight vector represents weight distribution of the contribution degree of each time node to the whole ESG risk of the enterprise after being normalized by softmax.
  6. 6. The artificial intelligence based ESG risk prediction and assessment method according to claim 1, wherein the modeling of the time series of the embedded vectors, after calculating the time weight distribution of the contribution degree of each time node to the whole ESG risk of the enterprise, further includes: constructing a causal discovery sub-model, identifying a key covariate set influencing ESG scores based on a constraint gene fruit search algorithm, and eliminating selective deviation in observed data by utilizing back door adjustment; the method specifically comprises the following steps: constructing a causal skeleton diagram through condition independence test based on a PC algorithm; the causal skeleton diagram is oriented by utilizing V structure identification and an acyclic constraint rule to form a complete causal diagram; Identifying all back door paths pointing to ESG scoring nodes and estimating the conditional probability distribution of each confounding variable; and calculating an intervention expectation by using the do operator and a back door adjustment formula to obtain an unbiased causal effect value.
  7. 7. The artificial intelligence based ESG risk prediction and assessment method of claim 1 or 6, further comprising: Performing inverse fact intervention on the ESG risk contribution degree, modifying the ESG factor value and re-passing through the graph neural network and the time sequence modeling flow under the premise of keeping other conditions unchanged so as to quantify single factor variation and modify the influence amplitude value of the final risk rating; And calculating the comprehensive ESG risk score of the target enterprise according to the embedded vector, the time weight distribution and the inverse fact influence amplitude.
  8. 8. The artificial intelligence based ESG risk prediction and assessment method of claim 7, wherein the performing a counterfacts intervention on the ESG risk contribution includes: modifying the observation value of a single ESG factor, and keeping the original observation state of the rest variables; reconstructing a dynamic heterogeneous graph structure based on the modified data, and multiplexing an original aging attenuation function and an edge weight initialization logic; And re-running the dynamic graph neural network and the time sequence modeling flow, and calculating the comprehensive ESG risk score difference before and after the intervention as the counter fact influence amplitude value of the factor.
  9. 9. The artificial intelligence based ESG risk prediction and assessment method of claim 7, wherein calculating a comprehensive ESG risk score for the target enterprise based on the embedded vector, the time weight distribution and the counter fact influence magnitude specifically includes: weighting and summing the basic embedding score, the time sequence weighting score and the causal correction score to obtain a comprehensive ESG risk score; When the comprehensive ESG risk score is lower than thirty, the risk is judged to be low, the risk is medium between thirty and sixty, the risk is high between sixty and eighty, and the risk is extremely high above eighty; And taking three propagation paths with highest activation intensity in the graph neural network as main evidence chains, and marking an initial event type, an intermediate conducting node and an end point enterprise entity by each path.
  10. 10. An artificial intelligence based ESG risk prediction and assessment system, comprising: The multi-source data access module is used for acquiring original ESG related data of a target enterprise from a plurality of external data sources, processing the original ESG related data, converting the original ESG related data into unified vector representation data, and establishing a cross-modal data index for the vector representation data according to a corresponding enterprise entity identifier; The data construction module is used for constructing a dynamic abnormal pattern structure based on the vector representation data, and carrying out multi-round message transmission and aggregation operation on the dynamic abnormal pattern structure by utilizing a dynamic graph neural network model to generate an embedded vector of each node in a current time window; The time sequence modeling module is used for performing time sequence modeling on the embedded vector and calculating time weight distribution of the contribution degree of each time node to the whole ESG risk of the enterprise; And the risk rating and interpretation generating module is used for calculating the comprehensive ESG risk score of the target enterprise according to the embedded vector and the time weight distribution, dividing the risk rating according to a preset threshold interval, and acquiring a main evidence chain causing the rating result and a propagation path of the main evidence chain in the graph structure.

Description

Artificial intelligence-based ESG risk prediction and assessment method and system Technical Field The invention belongs to the field of artificial intelligence prediction, and particularly relates to an ESG risk prediction and assessment method and system based on artificial intelligence. Background In the context of accelerated financial development and intensive discharge of sustainable regulatory policies, environmental, social and governance (ESG) have become the core dimension of enterprise strategic management. The ESG system covers multidimensional issues such as climate risk, supply chain responsibility, data privacy, corporate governance and the like, and the complexity and dynamics of the ESG system provide unprecedented challenges for information disclosure, risk identification and value conversion capability of enterprises. At present, enterprises generally face the problems of scattered ESG data sources, different standards, difficult verification and the like, so that the compliance disclosure cost is high, the reliability is low, meanwhile, the traditional risk management method relies on static indexes and manual research and judgment, and is difficult to capture ESG risk evolution trend across industries and regions in real time, and particularly, risk early warning is seriously lagged under the environments of frequent extreme climate events and aggravated geopolitical disturbance. In addition, although new sustainable assets such as carbon assets, natural capital, etc. are increasingly becoming points of increase in enterprise value, the lack of intelligent tools to support their identification, quantification and financial path design has led to ESG practice staying on a compliance level for a long period of time and failing to translate into a strategic competitive advantage. Among them, the ESG intelligent management technology based on artificial intelligence is gradually becoming a key direction for breaking the above dilemma. The technology aims to construct an intelligent decision system covering a full chain of information acquisition, risk early warning and value improvement through a data fusion, dynamic risk modeling and value evaluation algorithm driven by a large model. The method is characterized by integrating multi-source heterogeneous data (such as annual report of enterprises, government affair platforms, third party ratings, satellite remote sensing and the like), realizing ESG semantic understanding and index mapping by utilizing a knowledge enhancement language model, and carrying out combined deduction on climate physical risk, supply chain interruption risk and reputation public opinion risk by combining time sequence prediction and a graphic neural network, thereby supporting the enterprises to create transformation from passive compliance to active value. In the prior art, a part of platforms try to introduce machine learning to perform ESG scoring or carbon emission estimation, but have the obvious defects that firstly, data acquisition is highly dependent on manual reporting or a single interface, and the multi-mode fusion capability of unstructured text, images and equipment data of the Internet of things is lacking, so that the information integrity and the instantaneity are insufficient, secondly, a risk assessment model adopts a static weight or a rule engine, the sensitivity of risk factors cannot be dynamically adjusted according to industry characteristics, regional policies or emergency events, the early warning accuracy is low, thirdly, a value lifting module is commonly lacking, and the system has no automatic accounting mechanism for novel capital such as contribution to enterprise carbon assets and biodiversity, and lacks the capability of interfacing with financial tools such as green credit and sustainable bonds, so that a closed loop of 'risk control-asset value-added-financing enabling' is difficult to form. Particularly, in the face of the differentiated disclosure requirement of the marketing company and the rapid iteration of international standards such as TCFD, ISSB, the conventional system is difficult to realize flexible adaptation and continuous evolution, and an integrated ESG intelligent platform with a high-public-confidence data base, self-adaptive risk reasoning capability and an extensible value conversion architecture is needed. Disclosure of Invention The invention provides an ESG risk prediction and assessment method and system based on artificial intelligence, which are characterized in that a multisource heterogeneous data fusion architecture is constructed to perform unified characterization modeling on structured and unstructured data in environmental, social and corporate governance dimensions, a dynamic graph neural network is adopted to perform topological evolution analysis on complex association relations among entities, a time sequence attention mechanism is combined to capture the propagation path and intensity attenu