ESG risk prediction and evaluation method and system based on artificial intelligence

Abstract

The invention 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, one-sidedness and uninterpretability of existing ESG evaluation. The method comprises the following steps: fusing multi-source heterogeneous ESG data to construct a dynamic heterogeneous graph; modeling risk propagation and time-varying features by using a dynamic graph neural network and a time sequence attention mechanism; introducing causal reasoning to eliminate hybrid deviation, and deducing the influence of a quantization factor through anti-fact; and finally outputting an interpretable risk level and an evidence chain. The system comprises a multi-source data access module, a standardized alignment module, a dynamic graph construction module, a graph embedding module, a time sequence weighting module, a causal correction module, an anti-fact deduction module and a risk rating module. According to the invention, real-time, accurate and explainable evaluation of the enterprise ESG risk is realized, and the risk identification and decision support capability is significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence prediction, specifically relating to an AI-based ESG risk prediction and assessment method and system. Background Technology

[0002] Against the backdrop of accelerated financial development and a flurry of sustainable regulatory policies, Environmental, Social, and Governance (ESG) has become a core dimension of corporate strategic management. The ESG framework encompasses multi-dimensional issues such as climate risk, supply chain responsibility, data privacy, and corporate governance. Its complexity and dynamism pose unprecedented challenges to companies' information disclosure, risk identification, and value transformation capabilities. Currently, companies generally face problems such as fragmented ESG data sources, inconsistent standards, and difficulties in verification, resulting in high compliance disclosure costs and low credibility. Simultaneously, traditional risk management methods rely on static indicators and manual judgment, making it difficult to capture cross-industry and cross-regional ESG risk evolution trends in real time. Especially in an environment of frequent extreme weather events and escalating geopolitical disturbances, risk warnings are severely delayed. Furthermore, although new sustainable assets such as carbon assets and natural capital are increasingly becoming growth drivers for corporate value, the lack of intelligent tools to support their identification, quantification, and financialization path design has caused ESG practices to remain at the compliance level for a long time, failing to effectively translate into strategic competitive advantages.

[0003] Among these, AI-based ESG intelligent management technology is gradually becoming a key direction for overcoming the aforementioned challenges. This technology aims to construct an intelligent decision-making system covering the entire chain of "information collection—risk warning—value enhancement" through large-scale model-driven data fusion, dynamic risk modeling, and value assessment algorithms. Its core lies in integrating multi-source heterogeneous data (such as corporate annual reports, government platforms, third-party ratings, satellite remote sensing, etc.), utilizing knowledge-enhanced language models to achieve ESG semantic understanding and indicator mapping, and combining time-series prediction and graph neural networks to jointly extrapolate climate-physical risks, supply chain disruption risks, and reputational and public opinion risks, thereby supporting enterprises in transforming from passive compliance to proactive value creation.

[0004] In existing technologies, some platforms attempt to introduce machine learning for ESG scoring or carbon emission estimation, but significant shortcomings remain: First, data collection heavily relies on manual reporting or single interfaces, lacking the ability to fuse unstructured text, images, and IoT device data in a multimodal manner, resulting in insufficient information completeness and real-time performance. Second, risk assessment models often employ static weights or rule engines, failing to dynamically adjust risk factor sensitivity based on industry characteristics, regional policies, or unforeseen events, leading to low early warning accuracy. Third, value enhancement modules are generally lacking, with neither automated accounting mechanisms for new forms of capital such as corporate carbon assets and biodiversity contributions, nor the ability to connect with financial instruments such as green credit and sustainable bonds, making it difficult to form a closed loop of "risk control—asset appreciation—financing empowerment." Especially in the face of differentiated disclosure requirements for listed companies and the rapid iteration of international standards such as TCFD and ISSB, existing systems struggle to achieve flexible adaptation and continuous evolution, necessitating an integrated ESG intelligent platform with a highly credible data foundation, adaptive risk reasoning capabilities, and a scalable value transformation architecture. Summary of the Invention

[0005] This invention provides an AI-based ESG risk prediction and assessment method and system. It constructs a multi-source heterogeneous data fusion architecture to uniformly represent and model structured and unstructured data across environmental, social, and corporate governance dimensions. A dynamic graph neural network is used to perform topological evolution analysis on complex relationships between entities, combined with a temporal attention mechanism to capture the propagation path and intensity decay patterns of risk events over time. Furthermore, a causal reasoning module is introduced to identify the interference effects of potential confounding factors on ESG scores, and robust risk warning signals are generated based on counterfactual deduction. Finally, interpretable enterprise-level ESG risk level assessment results and a report tracing the origins of key driving factors are output.

[0006] According to one aspect of the present invention, an artificial intelligence-based ESG risk prediction and assessment method is provided, comprising: The original ESG-related data of the target enterprise is obtained, and the original ESG-related data is processed to convert it into a unified vector representation data. Based on the corresponding enterprise entity identifier, a cross-modal data index is established for the vector representation data. A dynamic heterogeneous graph structure is constructed based on the vector representation data. A dynamic graph neural network model is used to perform multiple rounds of message passing and aggregation operations on the dynamic heterogeneous graph structure to generate the embedding vector of each node in the current time window. Time series modeling of the embedded vectors is performed to calculate the time weight distribution of the contribution of each time point to the overall ESG risk of the enterprise. Based on the embedding vector and time weight distribution, the comprehensive ESG risk score of the target enterprise is calculated, and the risk level is divided according to the preset threshold range. The main evidence chain leading to the rating result and its propagation path in the graph structure are obtained.

[0007] An AI-based ESG risk prediction and assessment system includes: The multi-source data access module is used to obtain the original ESG-related data of the target enterprise from multiple external data sources, process the original ESG-related data to convert it into unified vector representation data, and establish a cross-modal data index for the vector representation data based on the corresponding enterprise entity identifier. The data construction module is used to construct a dynamic heterogeneous graph structure based on the vector representation data, and to perform multi-round message passing and aggregation operations on the dynamic heterogeneous graph structure using a dynamic graph neural network model to generate the embedding vector of each node in the current time window. The time series modeling module is used to perform time series modeling on the embedded vectors and calculate the time weight distribution of the contribution of each time point to the overall ESG risk of the enterprise. The risk rating and interpretation generation module is used to calculate the comprehensive ESG risk score of the target enterprise based on the embedding vector and time weight distribution, divide the risk level according to the preset threshold range, and obtain the main evidence chain leading to the rating result and its propagation path in the graph structure.

[0008] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0009] This invention overcomes the limitations of traditional ESG assessment methods that rely solely on static financial statements and questionnaires by constructing a dynamic heterogeneous graph structure that integrates multi-source heterogeneous data. It achieves real-time perception and correlation analysis of actual corporate operations. By introducing dynamic graph neural networks and temporal attention mechanisms, it effectively captures the nonlinear propagation patterns and temporal evolution characteristics of ESG risks in complex business networks. Combining causal reasoning and counterfactual deduction techniques, it significantly improves the accuracy of risk attribution and the reliability of decision-making recommendations, avoiding the common error of misjudging correlation as causation. The final interpretable report not only provides risk level determination but also reveals the specific paths and key nodes of risk transmission, offering valuable technical tools for corporate managers, investment institutions, and regulatory authorities. This addresses the three core problems of existing ESG assessment systems: lag, bias, and lack of interpretability. Attached Figure Description

[0010] Figure 1 This is a flowchart of the ESG risk prediction and assessment method based on artificial intelligence proposed in this invention; Figure 2This is a network framework diagram of the AI-based ESG risk prediction and assessment method proposed in this invention. Figure 3 This is a schematic diagram of the core principle framework of the dynamic heterogeneous graph neural network and temporal attention fusion mechanism in this invention; Figure 4 This is a logical flowchart of the standardization, cleaning, semantic alignment, and unified vector representation construction of multi-source heterogeneous data in this invention. Figure 5 This is a schematic diagram of the logical framework and deviation correction process of the causal reasoning and counterfactual deduction module in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between enterprise-level ESG risk level determination and the generation of interpretable evidence chains in this invention. Figure 7 This is a framework diagram of the AI-based ESG risk prediction and assessment system proposed in this invention. Figure 8 This is a schematic diagram of the overall technical solution architecture of the AI-based ESG risk prediction and assessment system proposed in this invention. Detailed Implementation

[0011] Example 1

[0012] This invention provides an artificial intelligence-based ESG risk prediction and assessment method and system. Its core lies in using multi-source heterogeneous data fusion, dynamic graph neural network modeling, temporal attention mechanisms, causal reasoning, and counterfactual deduction to achieve accurate prediction, quantitative assessment, and explainable attribution of corporate environmental, social, and governance (ESG) risks. The following will describe in detail the specific implementation methods of this approach.

[0013] like Figure 1-6 As shown, the specific steps of this invention are as follows: S1. Obtain the original ESG-related data of the target enterprise, process the original ESG-related data to convert it into unified vector representation data, and establish a cross-modal data index for the vector representation data based on the corresponding enterprise entity identifier. In step S1, raw ESG-related data of the target company needs to be obtained from multiple external data sources. This raw ESG-related data includes regulatory penalty records, supply chain public opinion texts, carbon emission monitoring indicators, employee satisfaction survey results, board composition information, and industry benchmark parameters. Specifically, records of pollutant emission exceeding standards issued by the ecological and environmental departments and labor inspection penalty decisions published by the human resources and social security departments are obtained through public database interfaces; breach of contract litigation documents and contract performance evaluations between upstream and downstream companies are captured through a business credit platform; the rate of vegetation cover change and the urban heat island effect index around the factory are analyzed through satellite remote sensing images; public sentiment regarding corporate social responsibility activities is collected through social media crawling programs; and data on the proportion of female board members, the number of times independent directors have performed their duties, and the compensation structure of senior executives are extracted through the company's annual report analysis module. All data exists in structured or unstructured forms, covering multiple modalities such as text, numerical values, images, and time-series signals, and has different update frequencies and reliability levels.

[0014] The raw ESG-related data is processed to transform it into a unified vector representation. Specifically, standardization cleaning and semantic alignment are performed on the raw ESG-related data to convert data of different formats into a unified vector representation. A cross-modal data index is then established based on preset enterprise entity identifiers. For numerical data, the Z-score normalization method is used to map it to a zero-mean unit variance space, as shown in the following formula:

[0015] Where x is the original value, μ is the mean of the feature within the historical observation window, and σ is its standard deviation.

[0016] For categorical data, one-hot encoding is used in conjunction with a domain knowledge ontology for semantic extension. For example, “environmental violation” in “administrative penalty type” is mapped to the “environmental dimension-pollution emission-excess emission” path node in the ontology.

[0017] For text-based data, a pre-trained language model is used to generate context-aware sentence vectors, which are then mapped to a predefined ESG topic tagging system using cosine similarity matching. This system includes three main categories: environment, society, and corporate governance, along with over thirty fine-grained subcategories. All processed data is appended with a unified timestamp field and a unique enterprise identifier, forming structured data tuples with spatiotemporal coordinates, facilitating subsequent cross-modal association and graph structure construction.

[0018] S2. Construct a dynamic heterogeneous graph structure based on the vector representation data, and use a dynamic graph neural network model to perform multiple rounds of message passing and aggregation operations on the dynamic heterogeneous graph structure to generate the embedding vector of each node in the current time window. A dynamic heterogeneous graph structure is constructed based on the vector representation. In this structure, nodes represent entities such as enterprises, suppliers, regulatory agencies, or natural events, while edges represent relationships such as equity control, contract performance, administrative penalties, or public opinion connections. Edge weights are determined by a combination of historical interaction frequency and a time-decrease function. For example, the time-decrease function is defined as an exponential decay function with a half-life parameter set to 180 days to reflect the natural decay of the impact of ESG negative events over time. The initial edge weight is determined by the product of the event severity score and the closeness of the relationship between the entities, and is subsequently updated daily according to the time-decrease function. Specifically, if an enterprise is fined heavily for environmental violations on day t, an administrative penalty edge is established between it and the environmental regulatory department. The initial weight is the product of the event severity score (e.g., the fine amount after logarithmic transformation and normalization) and the historical interaction frequency between the enterprise and the department; thereafter, the daily weight is updated according to the time-decrease function. Update, in which The date of the incident. These are the initial weights.

[0019] The dynamic graph neural network model is used to perform multiple rounds of message passing and aggregation operations on the dynamic heterogeneous graph structure to generate the embedding vector of each node within the current time window.

[0020] The embedding vector encodes the node's comprehensive state characteristics across three dimensions: environment, society, and corporate governance. The dynamic graph neural network employs a relationship-aware message passing mechanism, configuring independent weight matrices for different types of relationship edges. In each aggregation round, the information received by a node from its neighbors undergoes a linear transformation corresponding to the relationship type, followed by a weighted summation based on attention coefficients. These attention coefficients are obtained by normalizing the inner product of the embedding vectors of the source and target nodes using the softmax function, as shown in the following formula:

[0021] Where hi and hj are the current embedding vectors of nodes i and j, respectively, and Wr is the learnable weight matrix corresponding to relation type r. Let be the set of neighbors of node i under relation r. After L rounds of message passing, the final embedding vector of each node is the concatenation or weighted average of the outputs of each round, with a dimension of 768, corresponding to the 256-dimensional feature subspaces of the three dimensions of environment, society, and corporate governance.

[0022] S3. Perform time series modeling on the embedded vector and calculate the time weight distribution of the contribution of each time point to the overall ESG risk of the enterprise. Time series modeling is applied to the embedded vectors, and a gated recurrent unit network (GRN) is used to extract long-term dependency patterns. A multi-head self-attention mechanism is then used to calculate the weight distribution of each time point's contribution to the overall ESG risk of the enterprise. The hidden state dimension of the GRN is set to 256 dimensions to capture historical behavioral patterns spanning more than two years. The input sequence is the daily node embedding vectors for 360 consecutive days. After processing by the GRN, the hidden state sequence for the corresponding time step is output. Subsequently, this hidden state sequence is input into the multi-head self-attention mechanism, which contains eight parallel heads, each with a 32-dimensional query, key, and value projection matrix. The attention weight calculation formula is as follows:

[0023] Where Q, K, and V are the query, key, and value matrices, respectively, and dk=32 is the scaling factor. The outputs of each header are concatenated, normalized by the layer, and connected by the residuals before being fed into the fully connected layer to generate a time weight vector of length 360. This vector, after being normalized by softmax, represents the relative importance of each time point to the current ESG risk assessment.

[0024] Preferably, a causal discovery sub-model can also be constructed, which identifies the set of key covariates affecting ESG scores based on the constrained gene-fruit search algorithm, and uses a backdoor adjustment formula to eliminate selection bias in the observation data.

[0025] First, a preliminary causal skeleton graph is constructed based on the PC algorithm, which gradually removes irrelevant edges through conditional independence testing; Then, orientation rules (such as V-structure identification and acyclic constraints) are used to determine the direction of the edges, forming a complete causal graph. Next, all backdoor paths pointing to ESG scoring nodes are identified, and the conditional probability distributions of each confounding variable are estimated.

[0026] Finally, the do operator is applied to estimate the expected intervention value, obtaining the unbiased causal effect value. The backdoor adjustment formula is as follows:

[0027] Where Y is the ESG score, X is the ESG factor to be evaluated, and Z is the set of confounding variables that satisfy the backdoor criterion.

[0028] Furthermore, a counterfactual intervention simulation is performed. While keeping other conditions constant, the values ​​of specific ESG factors are modified, and the aforementioned graph neural network and time-series modeling process is rerun to quantify the impact of a single factor change on the final risk rating. During the intervention, the counterfactual inference module only allows modification of the observed values ​​of a single ESG factor, while other variables remain in their original observed state. The graph structure reconstruction process after intervention reuses the original time-decrease function and edge weight initialization logic to ensure the simulation results are realistically feasible. For example, increasing the proportion of female members on a company's board of directors from 20% to 40%, reconstructing the graph structure, and again using a dynamic graph neural network model to perform multiple rounds of message passing and aggregation operations on the dynamic heterogeneous graph structure, and applying time-series modeling operations to the embedded vectors; comparing the difference in the comprehensive ESG risk score before and after the intervention, this difference represents the counterfactual impact of the factor.

[0029] S4. Based on the embedding vector and time weight distribution, calculate the comprehensive ESG risk score of the target enterprise, divide the risk level according to the preset threshold range, and obtain the main evidence chain leading to the rating result and its propagation path in the graph structure.

[0030] Preferably, the target company's comprehensive ESG risk score and risk level can be calculated based on the embedding vector, time weight distribution, and counterfactual impact magnitude.

[0031] Based on the embedding vector, time weight distribution, and counterfactual impact magnitude, the comprehensive ESG risk score of the target enterprise is calculated, and risk levels are classified according to a preset threshold range. Simultaneously, the main chain of evidence leading to the rating result and its propagation path in the graph structure are output. The comprehensive ESG risk score is obtained by weighted summation of three parts: the basic embedding score, the time-weighted score, and the causal correction score.

[0032] A four-level risk classification standard is preset: a comprehensive ESG risk score below 30 is considered low risk, between 30 and 60 is medium risk, between 60 and 80 is high risk, and above 80 is extremely high risk. The main evidence chain consists of the three propagation paths with the highest activation intensity in a graph neural network. Each path is labeled with the starting event type, intermediate transmission nodes, and the ending corporate entity. For example, a typical evidence chain is "environmental penalty event → damage to the parent company's reputation → increased supply chain financing costs → target company's tight cash flow." This path is represented in the graph structure as a high-weight directed path from the regulatory agency node through the parent company node to the target company node. Its activation intensity is determined by the product of the attention coefficients of each edge and the changes in node embedding.

[0033] Obtain raw ESG-related data from multiple external data sources for the target company; The original ESG-related data is subjected to standardization cleaning and semantic alignment processing. A dynamic heterogeneous graph structure is constructed based on the aforementioned vector representation. A dynamic graph neural network model is used to perform multi-round message passing and aggregation operations on the dynamic heterogeneous graph structure. Apply time series modeling operations to the embedded vector; Construct a causal discovery sub-model to identify the set of key covariates that influence ESG scores; Perform counterfactual intervention simulation; Based on the embedding vector, time weight distribution, and counterfactual impact magnitude, the comprehensive ESG risk score of the target enterprise is calculated, the risk level is classified, and the main evidence chain and its propagation path are output.

[0034] like Figure 7 , 8 As shown, the present invention also provides an ESG risk prediction and assessment system based on artificial intelligence, including: a multi-source data access module 1, a data construction module 2, a time series modeling module 3, and a risk rating and interpretation generation module 4; The multi-source data access module 1 is used to obtain the original ESG-related data of the target enterprise from multiple external data sources, process the original ESG-related data to convert it into unified vector representation data, and establish a cross-modal data index for the vector representation data based on the corresponding enterprise entity identifier. Data construction module 2 is used to construct a dynamic heterogeneous graph structure based on the vector representation data, and to perform multi-round message passing and aggregation operations on the dynamic heterogeneous graph structure using a dynamic graph neural network model to generate the embedding vector of each node in the current time window. Time series modeling module 3 is used to perform time series modeling on the embedded vector and calculate the time weight distribution of the contribution of each time point to the overall ESG risk of the enterprise. The risk rating and interpretation generation module 4 is used to calculate the comprehensive ESG risk score of the target enterprise based on the embedding vector and time weight distribution, divide the risk level according to the preset threshold range, and obtain the main evidence chain that leads to the rating result and its propagation path in the graph structure.

[0035] The multi-source data access module collects multimodal ESG data in real time from government, commercial platforms, social media, and publicly available corporate documents via API interfaces, web crawlers, satellite data decoders, and document parsing engines. Deployed on a distributed computing cluster, it cleans, normalizes, and vectorizes the raw data stream in real time, outputting structured data tuples with spatiotemporal labels. The data construction module 2 maintains an incrementally updated graph database, dynamically adding and deleting nodes and edges daily based on newly arrived data, and refreshing edge weights according to a time decay function. The graph neural network embedding module adopts a distributed graph computing framework, supporting parallel message passing and embedding generation for large-scale heterogeneous graphs. The temporal modeling module 3 is integrated into a deep learning inference engine, performing temporal modeling on node embedding sequences and outputting time weight distributions. Furthermore, an attention weighting module 5 and a subsequent causal inference and bias correction module run in a statistical computing environment, performing causal graph construction, backdoor path identification, and do operator estimation. Preferably, a counterfactual inference module 6 is also included, reusing the aforementioned graph and sequence modeling components to perform intervention simulations in an isolated environment. The Risk Rating and Interpretation Generation Module 4 integrates all intermediate results to generate a structured risk report, which includes risk level, score details, key driving factors, and a visualized chain of evidence.

[0036] The aforementioned methods and system work together to achieve an end-to-end automated process from raw data collection to interpretable risk assessment. The entire process is updated daily to ensure that the assessment results reflect the company's latest ESG status. Internal modules exchange data via message queues and shared memory, guaranteeing high throughput and low latency. All model parameters are regularly updated through an online learning mechanism to adapt to the long-term evolution of ESG policies, market environments, and corporate behavior patterns. The final output of risk levels and evidence chains not only serves investment decisions and compliance reviews but also acts as a diagnostic tool for corporate ESG improvement, guiding targeted optimization of weaknesses.

[0037] The basic principles, main features, and advantages of the present invention have been described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based ESG risk prediction and assessment method, characterized in that, include: The original ESG-related data of the target enterprise is obtained, and the original ESG-related data is processed to convert it into a unified vector representation data. Based on the corresponding enterprise entity identifier, a cross-modal data index is established for the vector representation data. A dynamic heterogeneous graph structure is constructed based on the vector representation data. A dynamic graph neural network model is used to perform multiple rounds of message passing and aggregation operations on the dynamic heterogeneous graph structure to generate the embedding vector of each node in the current time window. Time series modeling of the embedded vectors is performed to calculate the time weight distribution of the contribution of each time point to the overall ESG risk of the enterprise. Based on the embedding vector and time weight distribution, the comprehensive ESG risk score of the target enterprise is calculated, and the risk level is divided according to the preset threshold range. The main evidence chain leading to the rating result and its propagation path in the graph structure are obtained.

2. The ESG risk prediction and assessment method based on artificial intelligence according to claim 1, characterized in that, The processing of the original ESG-related data to convert it into a unified vector representation specifically includes: The original ESG-related data includes regulatory penalty records, supply chain public opinion texts, carbon emission monitoring indicators, employee satisfaction survey results, board composition information, and industry benchmark parameters. For numerical data in the original ESG-related data, the Z-score normalization method is used to map them to a zero-mean unit variance space. For categorical data in the original ESG-related data, semantic extension is performed using one-hot encoding combined with a domain knowledge ontology. For text-based data in the original ESG-related data, a pre-trained language model is called to generate context-aware sentence vectors, and these vectors are mapped to a predefined ESG topic label system through cosine similarity matching. All processed data are appended with a unified timestamp field and a unique enterprise identifier, forming a vector representation of data with spatiotemporal coordinates.

3. The AI-based ESG risk prediction and assessment method according to claim 2, characterized in that, Constructing a dynamic heterogeneous graph structure based on the aforementioned vector representation includes: The vector representation data includes enterprise entities, supplier entities, regulatory entities, and natural event entities as graph nodes; edges are established based on the actual business or event relationships between entities, and the edge types include equity control, contract performance, administrative penalties, or public opinion correlation. The initial weight of the edge is calculated as the product of the event severity score and the closeness of the relationship between the subjects, and is updated daily according to the time decay function in the form of exponential decay, wherein the half-life parameter of the time decay function is set to 180 days.

4. The AI-based ESG risk prediction and assessment method according to claim 3, characterized in that, The dynamic heterogeneous graph structure is subjected to multi-round message passing and aggregation operations using a dynamic graph neural network model, specifically including: Configure an independent learnable weight matrix for each type of relation edge; In each round of aggregation, the embedding vectors of neighboring nodes are linearly transformed according to the correspondence type and then weighted and summed according to the attention coefficients. The attention coefficients are obtained by normalizing the inner product of the embedding vectors of the source node and the target node using the softmax function. After multiple rounds of message passing, the outputs of each round are concatenated or weighted averaged to generate the embedding vector of each node within the current time window.

5. The AI-based ESG risk prediction and assessment method according to claim 4, characterized in that, Applying time series modeling to the embedded vector specifically includes: A gated cyclic unit network is set up, with the hidden state dimension of the gated cyclic unit network set to 256 dimensions, and the daily node embedding vector is set as the input value. The hidden state sequence output by the gated recurrent unit network is input into a multi-head self-attention mechanism with eight parallel heads, where the query, key, and value projection matrices of each head are all 32-dimensional. After splicing the outputs of each header, the layers are normalized and connected to the residuals. Then, a time weight vector of length 360 is generated through a fully connected layer. After being normalized by softmax, the time weight vector represents the weight distribution of the contribution of each time point to the overall ESG risk of the enterprise.

6. The ESG risk prediction and assessment method based on artificial intelligence according to claim 1, characterized in that, The step of performing time series modeling on the embedded vector and calculating the time weight distribution of the contribution of each time point to the overall ESG risk of the enterprise also includes: A causal discovery sub-model was constructed, and a set of key covariates affecting ESG scores was identified based on the constrained gene-effect search algorithm. Backdoor adjustment was used to eliminate selection bias in the observation data. Specifically, it includes: A causal skeleton graph is constructed based on the PC algorithm through conditional independence testing; V-structure recognition and acyclic constraint rules are used to orient the causal skeleton graph to form a complete causal graph. Identify all backdoor paths pointing to ESG scoring nodes and estimate the conditional probability distributions of each confounding variable; The expected intervention value is calculated by applying the do operator and the backdoor adjustment formula, thus obtaining the unbiased causal effect value.

7. The AI-based ESG risk prediction and assessment method according to claim 1 or 6, characterized in that, Also includes: Counterfactual intervention is carried out on the contribution of ESG risks. Under the premise of keeping other conditions unchanged, the values ​​of ESG factors are modified and the graph neural network and time series modeling process are re-processed to quantify the changes of individual factors and modify the impact value of the final risk rating. The overall ESG risk score of the target enterprise is calculated based on the embedding vector, time weight distribution, and counterfactual impact magnitude.

8. The ESG risk prediction and assessment method based on artificial intelligence according to claim 7, characterized in that, The aforementioned counterfactual intervention in the contribution of ESG risks specifically includes: Modify the observed value of a single ESG factor, while keeping the other variables in their original observed state; Reconstruct the dynamic heterogeneous graph structure based on the modified data, and reuse the original time-decrease function and edge weight initialization logic. Rerun the dynamic graph neural network and time series modeling process to calculate the difference in the comprehensive ESG risk score before and after the intervention, which is used as the counterfactual impact magnitude of this factor.

9. The AI-based ESG risk prediction and assessment method according to claim 7, characterized in that, Based on the embedding vector, time weight distribution, and counterfactual impact magnitude, the comprehensive ESG risk score of the target enterprise is calculated, specifically including: The basic embedding score, the time-series weighted score, and the causal correction score are weighted and summed to obtain the comprehensive ESG risk score. When the comprehensive ESG risk score is below 30, it is considered low risk; between 30 and 60, it is considered medium risk; between 60 and 80, it is considered high risk; and above 80, it is considered very high risk. The three propagation paths with the highest activation intensity in the graph neural network are used as the main evidence chains, and each path is labeled with the starting event type, intermediate transmission nodes and the ending enterprise entity.

10. An ESG risk prediction and assessment system based on artificial intelligence, characterized in that, include: The multi-source data access module is used to obtain the original ESG-related data of the target enterprise from multiple external data sources, process the original ESG-related data to convert it into unified vector representation data, and establish a cross-modal data index for the vector representation data based on the corresponding enterprise entity identifier. The data construction module is used to construct a dynamic heterogeneous graph structure based on the vector representation data, and to perform multi-round message passing and aggregation operations on the dynamic heterogeneous graph structure using a dynamic graph neural network model to generate the embedding vector of each node in the current time window. The time series modeling module is used to perform time series modeling on the embedded vectors and calculate the time weight distribution of the contribution of each time point to the overall ESG risk of the enterprise. The risk rating and interpretation generation module is used to calculate the comprehensive ESG risk score of the target enterprise based on the embedding vector and time weight distribution, divide the risk level according to the preset threshold range, and obtain the main evidence chain leading to the rating result and its propagation path in the graph structure.

Images