Ecological credit evaluation model construction method

Abstract

The invention provides a construction method of an ecological credit evaluation model. The construction method comprises the steps of collecting green financial activity data of an evaluation subject; analyzing the green financial activity data to obtain a subject portrait feature library of the evaluation subject, the subject portrait feature library comprising basic information of the evaluation subject, behavior pattern features and preference features of the evaluation subject in the green financial field; distributing a multi-dimensional identity label set for the evaluation subject; constructing an evaluation subject portrait based on the subject portrait feature library of the evaluation subject and the multi-dimensional identity identification set; based on the constructed causal inference model, according to the evaluation subject portrait, mining internal factor characteristics and external factor characteristics of green financial behaviors and ecological credit influencing the evaluation subject; and training the pre-trained neural network model according to the internal factor characteristics and the external factor characteristics until the training is finished, thereby obtaining an ecological credit evaluation model, and remarkably improving scientificity, accuracy and practicability of an ecological credit evaluation result.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent processing, and specifically relates to a method for constructing an ecological credit evaluation model. Background Art

[0002] With the global emphasis on ecological environment protection and the rapid development of the green economy, green finance, as an important tool to promote ecological sustainable development, plays a key role in aspects such as resource allocation and risk prevention. Ecological credit evaluation, as the core link of the green finance system, aims to provide a basis for financial institution decision-making, policy formulation, and industry supervision by scientifically evaluating the credit level of market entities in green finance activities, thereby effectively reducing transaction risks, optimizing resource allocation efficiency, and promoting the healthy development of green industries.

[0003] Currently, traditional credit evaluation methods mainly rely on structured data such as financial indicators and historical transaction records, and conduct credit evaluation by constructing a scoring model with fixed weights or simple machine learning algorithms. Such methods have many limitations: on the one hand, they overly focus on financial data and are difficult to comprehensively capture the behavioral characteristics and ecological impacts of market entities in green finance activities. For example, they lack consideration of non-financial factors such as the implementation effect of enterprise green projects and the environmental risk response ability; on the other hand, the fixed evaluation model lacks adaptability to the dynamic market environment and innovative business scenarios, and cannot timely reflect the impact of external factors such as policy changes and industry fluctuations on ecological credit.

[0004] In recent years, some studies have attempted to apply deep learning technology to the field of credit evaluation and automatically extract data features through neural networks. However, the existing technologies still have obvious defects: at the data processing level, most of them do not fully integrate multi-source heterogeneous data and do not make sufficient use of unstructured data (such as industry news, policy texts, environmental monitoring reports), resulting in incomplete information mining; in terms of model construction, there is a lack of in-depth analysis of the causal relationships between credit influencing factors, and only prediction is carried out through data correlation, making it difficult to accurately reveal the internal mechanism of ecological credit formation, thus affecting the accuracy and reliability of evaluation results; in addition, existing models generally lack targeted processing of key factors such as spatio-temporal characteristics and subject differences in the ecological credit evaluation scenario, and cannot meet the innovative and changeable needs of green finance business.

[0005] Therefore, there is an urgent need for a method for constructing an ecological credit evaluation model that can fully integrate multi-source data, deeply mine causal relationships, and adapt to the characteristics of green finance business, so as to improve the scientificity, accuracy, and practicality of ecological credit evaluation and promote the high-quality development of the green finance industry. Summary of the Invention

[0006] To solve the above technical problems, the present application provides a method for constructing an ecological credit evaluation model to at least solve or alleviate the problems existing in the above prior art.

[0007] To achieve the above object, according to one aspect of the present application, there is provided a method for constructing an ecological credit evaluation model, which includes:

[0008] Collect green financial activity data of the evaluation subject;

[0009] Analyze the green financial activity data to obtain a subject portrait feature library of the evaluation subject, where the subject portrait feature library includes the basic information of the evaluation subject, its behavioral pattern features and preference features in the field of green finance;

[0010] Assign a multi-dimensional identity identifier set to the evaluation subject, where the multi-dimensional identity identifier set includes a composite subject ID of a time stamp, a regional identifier, a subject type identifier and a unique serial number;

[0011] Construct an evaluation subject portrait based on the subject portrait feature library of the evaluation subject and the multi-dimensional identity identifier set;

[0012] Based on the constructed causal inference model, according to the evaluation subject portrait, mine the internal factor features and external factor features that affect the green financial behavior and ecological credit of the evaluation subject;

[0013] Train the pre-trained neural network model according to the internal factor features and external factor features until the training is completed to obtain an ecological credit evaluation model.

[0014] The technical solutions in the present application have at least the following technical advantages:

[0015] ① By collecting the green financial activity data of the evaluation subject and analyzing to obtain a subject portrait feature library including basic information, behavioral pattern features and preference features, it changes the limitation of the traditional method that only relies on structured financial data. Combining with the multi-dimensional identity identifier set to construct an evaluation subject portrait can integrate multi-source heterogeneous data, comprehensively cover various information of the subject in green financial activities, including unstructured behavioral preference data, etc., effectively solve the problem of incomplete information mining in the prior art, and more accurately depict the characteristics of the evaluation subject.

[0016] ② Using the constructed causal inference model, based on the evaluation subject portrait, mine the internal and external factor features that affect ecological credit. Different from the existing method that only relies on data correlation prediction, this method can deeply analyze the internal mechanism of the formation of ecological credit, reveal the influence of each factor on ecological credit from the causal logic level, avoid misjudgment caused by false correlation, and significantly improve the scientificity and accuracy of the ecological credit evaluation result.

[0017] ③ Train the pre-trained neural network model based on the mined internal and external factor features. The dynamic learning ability of the neural network can adapt to the innovative and variable characteristics of green finance business. Compared with the traditional scoring model with fixed weights, this method can learn external factors such as policy changes and industry fluctuations in real time, as well as internal factors such as changes in the behavior patterns of the entity itself, enabling the model to better adapt to different time, space, and business scenarios, solve the problem of poor adaptability of existing models to dynamic environments, and improve the practicality and reliability of ecological credit evaluation. Description of the Drawings

[0018] Figure 1 It is a flowchart of a method for constructing an ecological credit evaluation model according to an embodiment of the present application. Detailed Embodiments

[0019] As Figure 1 shown, a method for constructing an ecological credit evaluation model provided by an embodiment of the present application includes:

[0020] Collect green finance activity data of the evaluation entity;

[0021] Parse the green finance activity data to obtain the entity portrait feature library of the evaluation entity, where the entity portrait feature library includes the basic information of the evaluation entity, its behavior pattern features and preference features in the field of green finance;

[0022] Assign a multi-dimensional identity identifier set to the evaluation entity, where the multi-dimensional identity identifier set includes a composite entity ID of a time stamp, a region identifier, an entity type identifier, and a unique serial number;

[0023] Construct an evaluation entity portrait based on the entity portrait feature library of the evaluation entity and the multi-dimensional identity identifier set;

[0024] Based on the constructed causal inference model, according to the evaluation entity portrait, mine the internal factor features and external factor features that affect the green finance behavior and ecological credit of the evaluation entity;

[0025] Train the pre-trained neural network model according to the internal factor features and external factor features until the training is completed to obtain an ecological credit evaluation model.

[0026] Preferably, in a specific application scenario, it can be implemented by replacing or preferably integrating technologies such as quantum computing, graph theory, and high-order tensor operations, as detailed below:

[0027] 1. Green Finance Activity Data Collection and Preprocessing

[0028] Let the collected green finance activity data form a tensor Where n is the number of evaluation subjects, m is the data feature dimension, k is the length of the time series, and l is the number of data source channels. The tensor decomposition technology is used to denoise and enhance the features of the original data, which is specifically implemented through Higher-Order Singular Value Decomposition (HOSVD): Among them, is the core tensor, and U 1 , U 2 , U 3 , U 4 are the orthogonal basis matrices of the corresponding dimensions respectively. By setting a threshold to screen the elements of the core tensor , the key feature components are retained, and the preprocessed data tensor

[0029] 2. Construction of the subject portrait feature library

[0030] 2.1 Extraction of basic information

[0031] Based on the Graph Convolutional Network (GCN), the basic information is extracted, and the relationship between subjects is constructed as a graph structure where the node corresponds to the evaluation subject, and the edge ε represents the association relationship between subjects. The update formula for the feature vector of the i-th node is: Among them, is the feature vector of node i at the l-th layer, is the set of neighborhood nodes of node i, c ij The normalization coefficient between node i and j, W (l) is the weight matrix of the l-th layer, b (l) is the bias vector, and σ is the activation function (such as LeakyReLU). After multiple layers of iteration, the basic information feature vector b i is output .

[0032] 2.2 Extraction of behavior pattern features

[0033] The quantum state superposition principle is used to construct a behavior pattern feature extraction model. Define the quantum state vector |ψ i > to represent the behavior pattern of the i-th evaluation subject: Among them, |φ j > is the basic behavior pattern quantum state, and α ij is the quantum state superposition coefficient, satisfying The quantum state is transformed through the quantum gate operation U: |ψ i′ > = U|ψ i > Measure the transformed quantum state to obtain the behavior pattern feature vector a i .

[0034] 2.3 Extraction of preference features

[0035] Preference Feature Extraction Model Based on Attention Mechanism and Reinforcement Learning, Defining the Preference Feature Extraction Function: p i = RL-Attention(d i , R) where R is the reward function matrix, which is continuously optimized through the reinforcement learning algorithm to improve the ability to capture key preference features. The calculation formula of the attention mechanism is: where Q, K, and V are the query, key, and value matrices respectively, and d k is the key vector dimension. The subject portrait feature library F i is still expressed as

[0036] 3. Construction of Multi-Dimensional Identity Identification Set and Subject Portrait

[0037] The multi-dimensional identity identification set I i = [t i , r i , s i , u i T Remains unchanged. When constructing the subject portrait, the tensor fusion technology is introduced to fuse the subject portrait feature library F i with the multi-dimensional identity identification set I i into a tensor

[0038] 4. Causal Inference Model to Mine Internal and External Factor Features

[0039] Based on the dual framework of the Structural Causal Model (SCM) and causal effect estimation, define the structural equation: X = f(U, E) + g(Z) where X is the endogenous variable (green finance behavior and ecological credit indicators), U is the exogenous variable (internal and external factors), E is the error term, and Z is the instrumental variable. Use the causal effect estimation method (such as double machine learning) to calculate the causal effect τ of each factor on ecological credit: where Y(1) and Y(0) are the result variables when the intervention occurs and does not occur respectively. By screening the factors with significant causal effects, the internal factor feature vector and the external factor feature vector

[0040] 5. Training of the Pre-trained Neural Network Model

[0041] Adopt a hybrid deep neural network architecture, combining Transformer and Graph Neural Network (GNN). The input layer receives the internal and external factor feature vectors After being processed by the Transformer encoding layer: Then input it into the graph neural network layer, and the graph convolution update formula is: The output layer adopts an adaptive weighted fusion mechanism to obtain the ecological credit evaluation result​ Among them, ω l is the adaptive weight, which is jointly optimized by backpropagation and meta - learning algorithm. The loss function adopts a composite loss function: L = λ1L MSE + λ2L CE + λ3L TV Among them, L MSE is the mean - square error loss, L CE is the cross - entropy loss, L TV is the total variation regularization term, and λ1, λ2, λ3 are weight coefficients, and the optimal values are determined by the Bayesian optimization algorithm.

[0042] Therefore, the technical advantages of the above - mentioned solution are as follows:

[0043] First, traditional technologies usually only perform simple numerical statistics or linear transformations on structured financial data, transaction records, etc., and it is difficult to process multi - source heterogeneous data. For example, in the face of text policy information and image - based environmental monitoring data in green financial activities, traditional methods often cannot effectively extract key features. This solution introduces the high - order tensor decomposition technology (HOSVD) to process the data tensor of green financial activities which can perform feature extraction and noise reduction simultaneously from multiple dimensions such as the subject, feature, time, data source, etc. By screening and reconstructing the core tensor it can accurately retain the key information related to ecological credit, such as the green financial behavior characteristics of the subject under different time nodes and different data channels, effectively solving the problem of insufficient data utilization in traditional methods and improving the comprehensiveness and depth of data processing.

[0044] Second, traditional behavior pattern and preference feature extraction are mostly based on fixed rules or simple clustering algorithms. For example, using K - means clustering to divide behavior patterns, relying on manually set thresholds and rules to mine preferences, which cannot capture innovative non - linear relationships and dynamic change characteristics. In this solution, the behavior pattern feature extraction uses the principle of superposition of quantum states to construct the quantum state vector |ψ i >, and realizes the non - linear transformation of the behavior pattern through quantum gate operations, which can simulate the diversity and uncertainty of the subject's behavior, breaking through the static and linear description limitations of traditional methods for behavior patterns. The preference feature extraction combines the attention mechanism and reinforcement learning. By dynamically optimizing the reward function matrix R, it can adaptively focus on key preference features, and can capture the dynamic changes of the subject's preferences in green financial activities more accurately than the traditional fixed - rule mining method.

[0045] Third, traditional credit evaluation models mostly make predictions based on data correlation and cannot accurately distinguish causal relationships and spurious associations. For example, accidentally associating market fluctuations with changes in the subject's ecological credit.

[0046] The causal inference model constructed in this solution combines the Structural Causal Model (SCM) with the double machine learning method, and by introducing the instrumental variable Z and calculating the causal effect It can strictly distinguish the causal impacts of various factors on ecological credit. In the green finance scenario, it can accurately judge the true causal effects of factors such as policy and regulation changes and enterprise internal management adjustments on ecological credit, avoid the evaluation biases caused by false correlations in traditional methods, and significantly improve the scientificity and reliability of evaluation results.

[0047] Fourthly, the traditional neural network model has a single structure and often adopts fixed loss functions and optimization algorithms, making it difficult to adapt to the innovation and dynamics of green finance business, and it has poor generalization ability in dealing with ecological credit evaluations of different subjects and scenarios.

[0048] This solution adopts a hybrid architecture of Transformer and Graph Neural Network (GNN). The Transformer encoding layer can effectively capture the long-distance dependence relationships between factors, while the GNN layer can mine the graph structure associations between ecological asset elements. The combination of the two can comprehensively handle the innovative relationships in ecological credit evaluation. At the same time, the composite loss function L = λ1L MSE +λ2L CE +λ3L TV Combined with Bayesian optimization to adjust the weight coefficients, it not only considers the prediction error and classification accuracy, but also improves the model stability through the total variation regularization term. Compared with the traditional single loss function, it can optimize the model more accurately, enabling the model to have stronger adaptability and generalization ability in the innovative and changeable scenarios of green finance.

[0049] Optionally, train the pre-trained neural network model according to the internal factor features and external factor features until the training is completed to obtain an ecological credit evaluation model, including:

[0050] According to the internal and external factor features, construct a network structure with ecological asset elements as associated branches to form an ecological asset relationship map, and assign a unique element association ID to each ecological asset element therein;

[0051] Obtain the green finance business scenario description to construct a scenario simulation module;

[0052] Based on the scenario simulation module, evaluate the scenario interaction score between the evaluation subject and the ecological asset relationship map to form a multi-dimensional scenario association matrix;

[0053] According to the multi-dimensional scenario association matrix, train the pre-trained neural network model until the training is completed to obtain an ecological credit evaluation model.

[0054] Preferably, in a specific application scenario, the above solution is described in a replaceable or preferred manner.

[0055] 1. Construct an ecological asset relationship graph

[0056] Let the set composed of internal and external factor characteristics be where N is the total number of factor characteristics. The ecological asset element set is denoted as M is the number of ecological asset elements. Define the association matrix where the element g ij represents the association strength between the factor characteristic f i and the ecological asset element a j , and its calculation method is: where and are the feature vectors of the factor characteristic f i and the ecological asset element a j respectively, and are obtained through a pre-trained word vector model (such as Word2Vec) or a feature extraction network; cosine(·,·) is the cosine similarity function, which is used to measure the similarity degree of two vectors.

[0057] Based on the association matrix G, construct an ecological asset relationship graph where the node set is The weight w i of the edge (a j , a ij ) in the edge set ε is determined by the following formula: Assign a unique element association ID to each ecological asset element a j , which is generated by using the hash function h(a j ), that is, ID j = h(a j ) to ensure that each element can be uniquely identified during data processing and model training.

[0058] 2. Construct a scenario simulation module

[0059] Denote the set of green finance business scenario description texts as S is the number of scenario description texts. Use a pre-trained language model (such as BERT) to encode each text t s to obtain the feature vector where D is the dimension of the feature vector.

[0060] When constructing the scenario simulation module, introduce a variational autoencoder (VAE) model. The encoder q φ (z|h s ) maps the text feature h s to the latent space K is the dimension of the latent space, and its calculation process is: z s = μs +σ s ⊙∈where, is the random noise of standard normal distribution, ⊙ represents the element multiplication. Decoder p θ (h s |z) Reconstruct text features based on latent variables z, and train the VAE model by minimizing reconstruction loss and KL divergence loss to achieve effective modeling and simulation of green finance business scenarios.

[0061] 3. Form a multi-dimensional scenario association matrix

[0062] For the evaluation subject u, under scenario s, the ecological asset element a j The interaction score score(u,s,j) is calculated using reinforcement learning method. Define the policy network π θ (a|s,u) represents the probability of selecting ecological asset element a under the conditions of scenario s and subject u. By interacting with the ecological asset relationship graph, rewards are obtained according to the reward function R(u,s,a). The reward function can be set according to the green finance business goals, such as ecological benefit improvement and risk reduction. The deep Q-network (DQN) algorithm is used to optimize the policy network, and the Q value function Q θ (s,u,a) represents the expected cumulative reward for selecting action a under scenario s and subject u. Its update formula is: Q θ (s,u,a)←Q θ (s,u,a)+α[r+γmax a′ Q θ′ (s′,u,a′)-Q θ (s,u,a)], where α is the learning rate, γ is the discount factor, r is the immediate reward, (s′,a′) is the next scenario and action, and θ′ is the target network parameter. Finally, the multi-dimensional scenario association matrix The element m usj =score(u,s,j), where U is the number of evaluation subjects.

[0063] 4. Pre-trained neural network model training

[0064] Assume that the pre-trained neural network model is The input is a multi-dimensional context association matrix M, which is processed by a multi-layer neural network layer. The loss function L is defined as: L = λ1L regression +λ2L classification +λ3L entropy Among them, L regression is the regression loss, which is used to measure the difference between the predicted ecological credit score and the actual score, and is calculated using the mean square error (MSE): is the ecological credit score predicted by the model, y us Rating for authenticity.classification is the classification loss, which is used to measure the difference between the predicted ecological credit rating and the true rating, and is calculated using cross-entropy loss: C is the number of categories of the ecological credit rating, and are the probability distributions of the true and predicted ecological credit ratings respectively. L entropy is the entropy regularization term of the policy network, which is used to increase the diversity of the policy, and is calculated as: λ1, λ2, λ3 are weight coefficients, and the optimal values are determined by the cross-validation method. The gradient of the loss function with respect to the model parameters ω is calculated by the backpropagation algorithm The model parameters are updated using an adaptive learning rate optimization algorithm (such as Adam): where β is the learning rate, and the training is iterated until the loss function converges to obtain the final ecological credit evaluation model.

[0065] Therefore, the technical benefits of the above solution are as follows:

[0066] First, traditional technologies usually determine the relationship between ecological asset elements by manually setting rules or simple weighting methods. For example, the fixed weights of financial indicators and ecological credit are set based on experience, which is difficult to adapt to the innovative and changing green finance scenarios and lacks quantitative analysis of unstructured factors.

[0067] In this solution, the association strength g between factor features and ecological asset elements is calculated through the association matrix G ij , and using cosine similarity combined with a pre-trained word vector model or feature extraction network, the internal and external factor features and ecological asset elements are transformed into vector forms for quantitative calculation, which can effectively process the information contained in unstructured data such as text and images. The edge weight w ij is calculated by comprehensively considering the influence of all factor features, making the ecological asset relationship graph more accurately reflect the actual association. For example, it can dynamically capture the impact of policy changes on different ecological asset elements. Compared with traditional methods, it greatly improves the accuracy and dynamic adaptability of relationship modeling.

[0068] Second, traditional scenario simulations rely on expert experience or simple rule matching, which are difficult to simulate innovative green finance business scenarios and cannot handle semantic information and uncertainties in scenario descriptions. For example, classifying business scenarios only according to a fixed template cannot reflect the dynamic changes and individual differences of the scenarios.

[0069] This solution introduces a variational autoencoder (VAE), uses the pre-trained language model BERT to encode the text of the green finance business scenario, maps the text features to the latent space through the encoder, and then reconstructs the features by the decoder. It is trained by minimizing the reconstruction loss and the KL divergence loss to achieve probabilistic modeling of the scenario features, and can generate diverse and real-distribution-compliant scenario representations. For example, different business scenario simulations can be generated according to policy adjustments, providing richer and more realistic inputs for ecological credit evaluation, and solving the problems of single scenario modeling and lack of flexibility in traditional methods.

[0070] Thirdly, the calculation of traditional interaction scores is mostly based on fixed scoring rules, such as simply scoring according to the transaction amount and frequency, which cannot adapt to the innovative behaviors of different subjects in diverse scenarios, and it is difficult to dynamically adjust the evaluation strategy, resulting in the lack of accuracy and real-time nature of the evaluation results.

[0071] This solution adopts a reinforcement learning method to optimize the calculation of interaction scores through the policy network π θ (a|s,u) and the deep Q-network (DQN). The policy network dynamically generates the probability of selecting ecological asset elements according to the scenario and the subject. The DQN learns the optimal strategy by interacting with the ecological asset relationship graph and dynamically adjusts its behavior according to the reward function R(u,s,a). For example, when the market fluctuates, the model can quickly adjust the evaluation strategy for different ecological asset elements, enabling the multi-dimensional scenario correlation matrix to reflect the real interaction situation of the subject in the innovative scenario in real time. Compared with the traditional fixed strategy, it significantly improves the dynamics and accuracy of the evaluation.

[0072] Fourthly, traditional neural network training often uses a single loss function (such as mean squared error), only focusing on the difference between the predicted value and the true value, ignoring the generalization ability and policy diversity of the model, which is prone to overfitting and performs poorly when facing new scenarios.

[0073] This solution adopts a composite loss function L = λ1L regression +λ2L classification +λ3L entropy , and simultaneously considers the regression loss, the classification loss, and the entropy regularization term of the policy network. The regression loss and the classification loss ensure the accuracy of the prediction, and the entropy regularization term increases the policy diversity to prevent the model from falling into local optima. The weight coefficients λ1, λ2, λ3 are determined through cross-validation, and the model parameters are updated by combining the adaptive learning rate optimization algorithm, enabling the model to ensure both the prediction accuracy and good generalization ability in the green finance innovative scenario, overcoming the limitations of traditional training methods.

[0074] Optionally, according to the multi-dimensional scenario correlation matrix, the pre-trained neural network model is trained until the training ends to obtain an ecological credit evaluation model, including:

[0075] Obtain the real-time performance evaluation value of the evaluation subject in green finance activities and the actual contribution evaluation value of ecological asset elements to green finance business;

[0076] Based on the real-time performance evaluation value and the actual contribution evaluation value, calculate the dynamic weight for ecological asset elements to calibrate the multi-dimensional scenario correlation matrix;

[0077] According to the calibrated multi-dimensional scenario correlation matrix, train the pre-trained neural network model until the training is completed to obtain the ecological credit evaluation model.

[0078] Preferably, in a specific application scenario, the above solution is described in a replaceable or preferred manner.

[0079] 1. Obtain the real-time performance evaluation value and the actual contribution evaluation value

[0080] Let the set of evaluation subjects be where U is the number of evaluation subjects; the set of ecological asset elements is M is the number of ecological asset elements.

[0081] 1.1 Calculation of real-time performance evaluation value

[0082] Construct a real-time performance evaluation index system N is the number of indicators. For the evaluation subject u i , its real-time data under the indicator I j is denoted as x ij . Use principal component analysis (PCA) combined with the entropy weight method to determine the weight w of each indicator j . First, perform dimensionality reduction on the original data matrix X = [x ij U×N through PCA to obtain the principal component matrix PC and the eigenvalue vector λ. The formula for calculating the indicator weight by the entropy weight method is: where, e j is the information entropy of the indicator I j . Then the indicator weight The real-time performance evaluation value R i of the evaluation subject u i is:

[0083] 1.2 Calculation of actual contribution evaluation value

[0084] For the ecological asset element a k , its actual contribution evaluation value C k to the green finance business is calculated by using data envelopment analysis (DEA) combined with grey relational analysis. Let the input vector of the decision-making unit (ecological asset element) be I k = [i​k1 , i k2 , …, i kP , the output vector is O k = [o k1 , o k2 , …, o kQ , P is the number of input indicators, and Q is the number of output indicators. The efficiency value θ of the ecological asset factors is calculated through the DEA model k : where λ = [λ1, λ2, …, λ M T is the weight vector

[0085] Then, the grey relational analysis is used to calculate the correlation degree γ between the ecological asset factors and the green finance business objectives k , and the final actual contribution evaluation value C k = αθ k + (1 - α)γ k , where α is the weight adjustment coefficient

[0086] 2. Calculate the dynamic weights and calibrate the multi-dimensional scenario correlation matrix

[0087] Let the multi-dimensional scenario correlation matrix be M = [m ijs U×M×S , where S is the number of scenarios, and m ijs represents the interaction score of the evaluation subject u i with the ecological asset factor a j under the scenario s. A dynamic weight calculation model is constructed, and the adaptive weight neural network (AWNN) is used to calculate the dynamic weights w = [w1, w2, …, w M T . The input layer receives the real-time performance evaluation value vector R = [R1, R2, …, R U T and the actual contribution evaluation value vector C = [C1, C2, …, C M T , the activation function of the hidden layer is σ(·) (such as LeakyReLU), and the output layer outputs the dynamic weights. The forward propagation formula of the network is: h l = σ(W l h l-1 + b l ), where h l is the output vector of the l-th layer, W l is the weight matrix of the l-th layer, b l is the bias vector, and h0 = [R T , C T T . The calibrated multi-dimensional scenario correlation matrix is calculated as follows:​​​​​​

[0088] 3. Training the pre-trained neural network model based on the calibration matrix

[0089] Let the pre-trained neural network model be The input is the calibrated multi-dimensional scenario correlation matrix M. The improved hybrid loss function L is used for training: L = λ1L MSE + λ2L Focal + λ3L Contrastive , where L MSE is the mean squared error loss, which is used to measure the difference between the predicted value and the true value: is the model predicted value, and y ijs is the true ecological credit value. L Focal is the focal loss, which is used to solve the problem of sample imbalance: where is the predicted probability, and γ is the adjustment parameter. L Contrastive is the contrastive learning loss, which enhances the model feature representation ability: where z ijs is the feature vector output by the model, is the positive sample feature vector, z iks is the negative sample feature vector, sim(·,·) is the similarity function, and τ is the temperature parameter.

[0090] Calculate the gradient through the backpropagation algorithm Use the AdamW optimization algorithm to update the model parameters ω: where β is the learning rate, and λ wd is the weight decay coefficient. Continuously iterate the training until the loss function converges to obtain the final ecological credit evaluation model.

[0091] Therefore, the above scheme is compared with the traditional technology, and it focuses on the dynamic weight calibration and model optimization in the training of the ecological credit evaluation model. When compared with the traditional technology, it has the following advantages in terms of data evaluation, weight calculation, model training, etc.:

[0092] 1. When evaluating the real-time performance of the evaluation subject and the contribution of ecological asset elements, traditional ecological credit evaluation mostly uses a single indicator or a simple weighted average method. For example, only evaluating the ecological credit of an enterprise based on the investment amount of its green projects ignores important factors such as the implementation efficiency and environmental benefits of the projects; when evaluating the contribution of ecological asset elements, there is a lack of systematic analysis of input-output, making it difficult to accurately measure its actual value. This method is highly subjective and has a single indicator, and it cannot comprehensively reflect the true situation of the evaluation subject and ecological asset elements.

[0093] In this application, the real-time performance evaluation value is calculated by combining principal component analysis (PCA) with the entropy weight method. PCA can reduce the dimension of multi-dimensional raw data, remove data redundancy, and extract key information; the entropy weight method objectively determines the index weights based on the variation degree of the data itself, avoiding the interference of human factors. For example, in the green finance scenario, multiple indicators such as the enterprise's capital flow, project progress, and environmental impact can be integrated to accurately evaluate its real-time performance. The actual contribution evaluation value is calculated by combining data envelopment analysis (DEA) and grey relational analysis. DEA is used to evaluate the input-output efficiency of ecological asset elements, and grey relational analysis measures the degree of association between elements and the green finance business objectives. The combination of the two can comprehensively and accurately evaluate the actual contribution of ecological asset elements.

[0094] 2. Traditional weight determination methods are often fixed and difficult to dynamically adjust according to the actual situation once set. For example, in a credit evaluation model, the weights of various evaluation indicators remain unchanged for a long time and cannot adapt to the impacts brought by changes in the green finance market environment, policy adjustments, and the development of the evaluation subject itself. This static weight setting makes the evaluation results lack flexibility and timeliness.

[0095] This solution uses an adaptive weight neural network (AWNN) to calculate the dynamic weights of ecological asset elements. Taking the real-time performance evaluation value vector and the actual contribution evaluation value vector as inputs, AWNN can automatically capture the dynamic changes in the real-time performance of the evaluation subject and the actual contribution of ecological asset elements through the learning and training of a multi-layer neural network, and then adjust the weights of each element. When the green finance business scenario changes, such as a sudden surge in the market demand for a certain ecological asset, AWNN can timely increase the weight of this asset element, making the multi-dimensional scenario correlation matrix more in line with the actual situation and enhancing the adaptability and accuracy of the evaluation model.

[0096] 3. Traditional neural network model training usually uses a single loss function, such as the mean square error loss, which only focuses on the difference between the predicted value and the true value, making it difficult to solve the problem of sample imbalance and unable to effectively enhance the feature representation ability of the model. In ecological credit evaluation, there are cases where the number of samples of certain ecological credit grades is extremely small. The single loss function will lead to poor learning effects of the model on a small number of samples and reduce the overall evaluation accuracy.

[0097] This application uses an improved hybrid loss function, including mean square error loss (L MSE ), focal loss (L Focal ), and contrastive learning loss (L Contrastive ). L MSE ensures the basic fitting of the predicted value and the true value; L Focal reduces the weight of easy-to-classify samples and focuses on difficult-to-classify samples by adjusting the parameter γ, effectively solving the problem of sample imbalance; L ContrastiveBy comparing the positive and negative sample feature vectors, the model's ability to distinguish different ecological credit features is enhanced, and the feature representation ability is improved. The synergistic effect of multiple loss functions enables the model to be more comprehensively optimized during training, improving the accuracy of ecological credit evaluation and the generalization ability of the model.

[0098] Optionally, according to the calibrated multi-dimensional scenario correlation matrix, the pre-trained neural network model is trained until the training is completed to obtain an ecological credit evaluation model, including:

[0099] Quantitatively encode the risks and benefits brought by ecological asset elements according to the calibrated multi-dimensional scenario correlation matrix to obtain a risk coding vector and a benefit coding vector;

[0100] Based on the risk coding vector and the benefit coding vector, input them into the pre-trained neural network model for training until the training is completed to obtain an ecological credit evaluation model.

[0101] Preferably, in a specific application scenario, the above solution is described in a replaceable or preferred manner.

[0102] 1. Quantification and coding of risks and benefits of ecological asset elements

[0103] Let the calibrated multi-dimensional scenario correlation matrix be where U is the number of evaluation subjects, M is the number of ecological asset elements, S is the number of scenarios, represents the evaluation subject u i under the scenario s and the ecological asset element a j calibrated interaction score.

[0104] 1.1 Risk quantification and coding

[0105] Construct a risk assessment function Risk(a j ) to quantify the risk of the ecological asset element a j , adopting a hybrid model based on fuzzy comprehensive evaluation and Bayesian network. First, determine the risk evaluation index set N is the number of indicators, and the corresponding weight vector w r = [w r1 , w r2 , …, w rN T , and the weights are determined by combining the analytic hierarchy process (AHP) with the entropy weight method. For each evaluation index R k , establish a fuzzy membership function for mapping the original index value x to the membership degree in the interval [0, 1]. The fuzzy membership vector of the ecological asset element a j under the index R k is ​where \(i = 1, 2, \ldots, U\) and \(s = 1, 2, \ldots, S\). Calculate the preliminary risk value Risk through fuzzy transformation pre (a j ): Then, construct a Bayesian network where the nodes include risk assessment indicators and the risk status of ecological asset elements, and the edge \(\varepsilon\) represents the probabilistic dependence relationship between nodes. Learn the conditional probability table CPT of the Bayesian network from historical data. Use the Bayesian inference formula to correct the preliminary risk value: where \(X\) is the observed value vector of risk assessment indicators. Normalize and encode the risk values of all ecological asset elements to obtain the risk coding vector \(R = [R_1, R_2, \ldots, R M T , and the coding function is R all is the set of risk values of all ecological asset elements.

[0106] 1.2 Income Quantification Coding

[0107] Construct an income evaluation function Return(a j ), adopting a method combining random forest regression and Markov chain Monte Carlo (MCMC) simulation. Use the random forest regression model RF to train the historical income data \(Y j = [y j , y j1 , \ldots, y j2 , \ldots, y jT of the ecological asset element \(a j = [x j1 , x j2 , \ldots, x jT , where \(T\) is the number of time periods, and the corresponding feature vector \(X j = [x j1 , x j2 , \ldots, x jT . The feature vector includes factors related to the income of ecological asset elements, such as market demand and policy support. The random forest regression prediction function is: where \(B\) is the number of decision trees in the random forest, and \(f b (x jt ) is the prediction value of the \(b\)-th decision tree for \(x jt . Conduct uncertainty analysis on the prediction results of random forest regression through MCMC simulation to estimate the probability distribution of income. Let the Markov chain state transition probability matrix be \(P\), the initial state probability vector be \(\pi_0\), and after \(L\) iterations, obtain the probability distribution vector \(\pi L . The income value is the expected value of the probability distribution: Normalize and encode the income values of all ecological asset elements to obtain the income coding vector \(V = [V_1, V_2, \ldots, V M T , and the coding function is V all is the set of return values of all ecological asset elements.

[0108] 2. Training of the pre-trained neural network model based on the encoded vector

[0109] Let the pre-trained neural network model be The input layer receives the risk encoded vector R and the return encoded vector V, and concatenates them into the input vector I = [R T , V T . T . A network structure that combines the deep residual network (ResNet) with the attention mechanism (Attention) is adopted. In the residual block, let the input of the l-th layer be h l , and the output be h l+1 . The residual connection formula is: h l+1 = h l + F(h l , W l ) where F(·) is the residual function and W l is the weight matrix of the l-th layer. The calculation process of the attention mechanism is: Q = W q I, K = W k I, V = W v I where Q, K, and V are the query, key, and value matrices respectively, and W q , W k , W v are the weight matrices, and d k is the key vector dimension. The output layer adopts the loss function L of multi-task learning, which includes the ecological credit score regression loss L regression and the credit rating classification loss L classification : L = λ1L regression + λ2L classification where is the predicted ecological credit score, and y i is the true score; C is the number of credit rating categories, and y ic and are the true and predicted credit rating probability distributions respectively. λ1 and λ2 are the weight coefficients, which are determined by cross-validation. Calculate the gradient through the backpropagation algorithm and update the model parameters ω using the AdaGrad optimization algorithm: where η is the learning rate, T is the number of training epochs, and ∈ is a very small positive number to prevent the denominator from being zero. Continuously iterate the training until the loss function converges to obtain the final ecological credit evaluation model.

[0110] Therefore, in the scenario of ecological credit evaluation, traditional technologies have many limitations in risk-return quantification and model training. This application integrates cutting-edge technologies such as fuzzy comprehensive evaluation, Bayesian network, and random forest, achieving breakthroughs from multiple dimensions and having the following significant advantages:

[0111] 1. The risk quantification of traditional ecological credit evaluation mostly relies on fixed thresholds or simple linear weighted models. For example, the risk level is divided only based on a few indicators such as the number of corporate defaults and the asset-liability ratio, lacking dynamic assessment of innovative risk factors and handling of uncertainties. This method cannot adapt to non-linear risks such as policy changes and environmental fluctuations in green finance business, easily leading to misjudgment of risks.

[0112] This application adopts a hybrid model of fuzzy comprehensive evaluation and Bayesian network. Fuzzy comprehensive evaluation maps multi-dimensional risk evaluation indicators (such as policy compliance, degree of environmental impact, etc.) to fuzzy membership degrees by establishing a fuzzy membership function, which can effectively handle the fuzziness and uncertainty of evaluation indicators. For example, when evaluating the risk faced by an enterprise due to changes in environmental protection policies, fuzzy factors such as the implementation intensity of the policy and the enterprise's adaptability can be transformed into quantified membership degrees through a fuzzy function. The Bayesian network learns the probabilistic dependence relationship between nodes based on historical data and corrects the preliminary risk value through Bayesian inference, enabling dynamic update of the risk assessment result. For example, when a new environmental policy appears, the Bayesian network can quickly adjust the risk value of ecological asset elements according to the correlation probability between the policy and risk indicators, depicting risks more accurately and dynamically compared with traditional methods.

[0113] 2. Traditional income assessment often uses deterministic models, such as simple linear regression or empirical formulas, only considering a few key variables and ignoring the impact of uncertain factors such as market fluctuations and emergencies, and unable to accurately predict the income situation of ecological asset elements in an innovative environment.

[0114] This application uses a method combining random forest regression and Markov chain Monte Carlo (MCMC) simulation. Random forest regression can handle high-dimensional and non-linear data by integrating multiple decision trees, effectively capturing the innovative relationships between ecological asset income and multiple factors such as market demand and policy support. For example, when predicting the income of a new energy project, factors such as market electricity price fluctuations and government subsidy changes can be comprehensively considered. MCMC simulation then conducts uncertainty analysis on the prediction results of random forest, estimating the probability distribution of income by constructing a Markov chain and calculating the expected value. This makes the evaluation result not only a definite value but also contains uncertainty information, being able to more comprehensively reflect the true situation of ecological asset income and having a significant improvement in accuracy and reliability compared with traditional methods.

[0115] III. Traditional neural network models have a simple structure and often use a single loss function, such as mean square error, which only focuses on the difference between the predicted value and the true value. When dealing with multi-objective tasks in ecological credit evaluation (such as score prediction and grade classification), it is difficult to balance the importance of different tasks, and the model generalization ability is weak.

[0116] This application uses a deep residual network (ResNet) combined with an attention mechanism and multi-task learning. ResNet solves the problem of gradient disappearance in the training of deep networks through residual connections, enabling the construction of deeper network structures and the extraction of more innovative features; the attention mechanism can make the model focus on key risk and return coding features, improving the feature representation ability. The loss function of multi-task learning simultaneously includes regression loss and classification loss, and determines the weight coefficients through cross-validation, which can effectively balance the ecological credit score prediction and credit grade classification tasks. For example, during the training process, λ1 and λ2 can be adjusted according to actual needs to make the model achieve better performance on different tasks. Compared with traditional models, this combination has stronger adaptability and higher prediction accuracy in innovative ecological credit evaluation tasks.

[0117] Optionally, according to the calibrated multi-dimensional scenario association matrix, quantify and encode the risks and returns brought by ecological asset elements to obtain a risk coding vector and a return coding vector, including:

[0118] Perform structured and unstructured data extraction processing on the calibrated multi-dimensional scenario association matrix, separate the structured data in the matrix, and collect the unstructured data in green financial activities to obtain a data set to be fused;

[0119] Extract features from the text data in the unstructured data to obtain a text feature vector;

[0120] Fuse the features of the structured data and the text feature vector to obtain enhanced scenario association data;

[0121] Perform data integration processing on the enhanced scenario association data and the timestamp and regional identification information in the multi-dimensional identity identification set of the evaluation subject, associate the timestamp and regional identification information with the enhanced scenario association data to obtain spatio-temporal association data;

[0122] According to the spatio-temporal association data, mine the influence probability of factors on risks and returns to screen out core factors;

[0123] Perform quantum computing principle mapping processing on the screened core factors, map the core factors into a qubit form that simultaneously represents multiple risk and return states to obtain a quantumized factor representation;

[0124] Perform quantum entanglement phenomenon simulation processing on the quantization factor representation, establish an innovative non-linear correlation between factors based on the correlation relationship between factors, and construct a quantum-inspired risk-return quantization model;

[0125] Perform quantum solution on the quantum-inspired risk-return quantization model to calculate the quantization values of the risks and returns brought to the ecological asset elements;

[0126] Input the quantization values into the constructed risk coding generator and return coding generator to generate the initial risk coding vector and return coding vector;

[0127] Perform discriminator judgment processing on the generated risk coding vector and return coding vector to obtain the risk coding vector and return coding vector.

[0128] Preferably, in a specific application scenario, the above solution is described in a replaceable or preferred manner.

[0129] 1. Structured and unstructured data extraction and processing

[0130] Let the calibrated multi-dimensional scenario correlation matrix be where n is the number of evaluation subjects, m is the feature dimension, and s is the number of scenarios. Using tensor decomposition technology, decompose it into a core tensor and multiple factor matrices That is where, × i represents the tensor product operation along the i-th dimension, and r1, r2, and r3 are the ranks of the corresponding dimensions. The structured data tensor can be obtained by retaining the dimension slice related to the structured features in the core tensor , and the information related to the unstructured data is extracted from the original matrix unstr through the factor matrix and a specific mapping function f to form an unstructured data set

[0131] 2. Unstructured text data feature extraction

[0132] For the text data in the unstructured data set , encode it using a pre-trained deep language model (such as BERT-Large). Let the text sequence be w1, w2,..., w T , and after encoding by the model, obtain the hidden state tensor where d h is the hidden layer dimension. Calculate the feature vector through the multi-head attention mechanism:

[0133] where, Wq , W k , W v is the projection matrix, d k is the dimension of the key vector, h is the number of attention heads, and finally the text feature tensor is obtained d f is the dimension of the fused feature.

[0134] 3. Feature Fusion

[0135] Convert the structured data tensor to matrix S mat and then fuse it with the text feature tensor . Introduce a gating mechanism and define the gating vector where σ is the Sigmoid function, denotes matrix concatenation, W g and b g are learnable parameters. The fused enhanced scenario correlation data ε is: where ⊙ denotes element-wise multiplication.

[0136] 4. Data Integration and Processing

[0137] The timestamp information of the evaluation subject and the region identification information are both represented in tensor form. Through tensor concatenation operation, integrate them with the enhanced scenario correlation data ε into the spatio-temporal correlation data tensor and perform high-order singular value decomposition (HOSVD) on it: where, is the core tensor, and V1, V2, V3 are factor matrices.

[0138] 5. Core Factor Screening

[0139] Based on the spatio-temporal correlation data tensor construct a dynamic Bayesian network where is the node set, ε is the edge set, is the time slice set. Define the conditional probability distribution P(X i |Pa(X i ), t), where is the node variable, and Pa(X i ) is its parent node set. Through the particle filter algorithm for inference, calculate the dynamic influence probabilities of each factor on the risk R and the return G and where is the observation information at time t. According to the information gain criterion IG(X i ) = H(R) + H(G) - H(R, G|Xi ) Screen the core factors, where H(·) is the Shannon entropy.

[0140] 6. Quantum computing principle mapping processing

[0141] For the set of screened core factors Map it to a quantum state. Let the number of core factors be c, and construct a quantum register where the state of each qubit |ψ i > is parameterized by the core factor value x i : where β and γ are adjustment parameters, and atan2 is the four-quadrant arctangent function.

[0142] 7. Quantum entanglement phenomenon simulation processing

[0143] Based on the correlation matrix between core factors Use controlled quantum gates to construct quantum entanglement. For the elements λ ij ≠0 in the matrix, apply a controlled phase gate (CPHASE) between the i-th and j-th qubits: Through a series of quantum gate operations, obtain the quantum state |Φ> of the quantum heuristic risk-return quantification model.

[0144] 8. Quantum solution

[0145] Adopt the quantum Monte Carlo method to solve the quantum state |Φ>. Define the Hamiltonian whose eigenstate |∈ n > and eigenvalue ∈ n satisfy Obtain a sample set through quantum sampling Calculate the quantified values of risk and return: where, and are the Hamiltonian components related to risk and return respectively.

[0146] 9. Encoding generation

[0147] Construct a deep generative adversarial network (GAN) as the encoding generator. The generator G receives the risk and return quantified values R q and G q as well as random noise and outputs the initial risk encoding vector r and return encoding vector g: r, g = G([R q ; G q ; z]). The discriminator D discriminates the generated encoding vectors and outputs the probability p = D([r; g]), and optimizes the parameters of the generator and discriminator through adversarial training.

[0148] [[ID=

[0149] Introduce a reinforcement learning mechanism to optimize the discriminator. Define the reward function R(r,g), and give rewards according to the difference between the discrimination result and the real data distribution. Use the Deep Q-Network (DQN) to learn the optimal policy π * , and iteratively update the discriminator parameter θ D : where γ is the discount factor and α is the learning rate. Through continuous training, the final reliable risk coding vector and return coding vector are obtained.

[0150] Therefore, the above scheme integrates cutting-edge technologies such as tensor analysis, quantum computing, and deep learning. Compared with traditional ecological asset risk-return assessment technologies, it has the following significant advantages in data processing, feature extraction, quantitative modeling, etc.

[0151] 1. Traditional methods usually process data as a two-dimensional matrix, which can only handle structured numerical data and is difficult to effectively integrate unstructured data such as text and images. For example, in ecological asset risk assessment, traditional methods only rely on structured data such as financial statements to calculate risk indicators, ignoring the risk information contained in unstructured texts such as policy documents and market news, resulting in a single assessment dimension.

[0152] This application uses tensor decomposition technology to process high-dimensional matrices to separate the characteristics of structured and unstructured data. At the same time, use a pre-trained language model (such as BERT-Large) combined with a multi-head attention mechanism to process unstructured text and convert the text sequence into a feature tensor This method can not only process multi-source heterogeneous data, but also retain the multi-dimensional correlation information of the data through tensor operations. For example, when analyzing the risk of regional ecological assets, structured transaction data and unstructured policy texts at different times and different entities can be fused simultaneously to comprehensively capture risk signals.

[0153] 2. Traditional feature fusion mostly uses simple splicing or weighted summation, without considering the dynamic correlation and importance difference between features. For example, in credit assessment, directly adding features such as enterprise asset size and transaction frequency cannot reflect the weight changes of different features in different scenarios.

[0154] This application introduces a gating mechanism to dynamically adjust the fusion ratio of structured data S mat and text features through the Sigmoid function. In the green finance scenario, when the policy changes, the gating vector will automatically enhance the weight of text features related to the policy and weaken the influence of other features, so that the fused enhanced scenario-correlated data ε can more accurately reflect the key factors of the current ecological asset risk-return.

[0155] 3. Traditional methods rely on correlation analysis or fixed rules to screen factors and cannot handle dynamic causal relationships. For example, only by determining risk indicators through historical data correlation, when the market environment changes, the original indicators become invalid.

[0156] This application constructs a dynamic Bayesian network and uses the particle filter algorithm to calculate the dynamic influence probability and screens core factors through the information gain criterion IG(X i )). In ecological asset assessment, this method can update the influence probability of factors on risk and return in real time. For example, when a sudden environmental event occurs in a certain area, the dynamic Bayesian network can quickly identify the impact of event-related factors on the risk of ecological assets, which is more timely and causally interpretable than traditional methods.

[0157] Traditional quantitative models are based on classical probability or linear regression and are difficult to handle high-dimensional and non-linear innovative systems. For example, when using a linear model to predict the return of ecological assets, it is impossible to accurately describe the coupling effect of market fluctuations and policy interventions.

[0158] This application maps the core factors to the quantum state |ψ>, constructs quantum entanglement through controlled phase gates, and then uses the quantum Monte Carlo method to solve the quantitative values of risk and return. At the same time, the generative adversarial network (GAN) and the deep Q-network (DQN) are combined for encoding generation and discriminator optimization. This interdisciplinary integration enables the model to represent the superposition states of multiple risk and return states at the same time. For example, when evaluating new energy projects, the quantum-inspired model can simulate the risk and return distributions under the simultaneous changes of multiple factors such as policy subsidies and technological breakthroughs, and can more realistically reflect the uncertainty and innovation of ecological assets compared with traditional models.

[0159] Optionally, the risk coding vector and the return coding vector are input into a pre-trained neural network model for training until the training is completed to obtain an ecological credit evaluation model.

[0160] Perform self-attention mechanism fusion processing on the risk coding vector and the return coding vector to generate an enhanced feature coding vector set;

[0161] Perform stratified sampling processing on the enhanced feature coding vector set, and divide the enhanced feature coding vector set into different layers according to the risk level and return potential data of ecological asset elements to obtain a risk-return coding vector layer;

[0162] Input the risk-return coding vector layer into the embedded dynamic routing module in the pre-trained neural network model to extract the key features of the risk and return of ecological asset elements;

[0163] Input the key risk-return characteristics of ecological asset elements into the graph neural network module of the pre-trained neural network model to mine the propagation law of different levels of risk-return coding vector layers in the relationship graph, and generate the semantic features among ecological credit evaluation factors;

[0164] Input the semantic features among ecological credit evaluation factors into the transformer module of the pre-trained neural network model to capture the distance-dependent features among ecological credit evaluation factors;

[0165] Fuse the semantic features and distance-dependent features among ecological credit evaluation factors into a feature evaluation feature vector, and input it into the output layer of the pre-trained neural network model to generate an ecological credit evaluation prediction result and compare it with the true ecological credit label to calculate the loss value;

[0166] Optimize the model parameters of the pre-trained neural network model based on the loss value through backpropagation until the loss value reaches the preset model convergence condition, then end the training to obtain an ecological credit evaluation model.

[0167] Preferably, in a specific application scenario, the above scheme is described in a replaceable or preferred manner.

[0168] 1. Self-attention mechanism fusion processing

[0169] Let the risk coding vector be The return coding vector is Concatenate the two into an initial coding vector Introduce the multi-head self-attention mechanism. For the h-th attention head, calculate the query matrix Q h The key matrix K h And the value matrix V h : Among them, Is a learnable weight matrix, d k Is the key vector dimension, d v Is the value vector dimension.

[0170] Calculate the attention score matrix A h : Get the output z of the h-th attention head h : z h = A h V h Concatenate the outputs of all attention heads and perform a linear transformation to obtain the feature-enhanced coding vector z: z = W o [z1; z2; …; z H Among them, Is the output weight matrix, H is the number of attention heads, d outis the output dimension. Repeat the above operations on the encoding vectors of multiple samples to obtain the enhanced feature encoding vector set. N is the number of samples.

[0171] 2. Stratified sampling processing

[0172] Let the risk level data of ecological asset elements be and the revenue potential data be Normalize them as follows: Define the stratification index where ω1 and ω2 are weight coefficients determined by cross-validation.

[0173] According to the stratification index I, divide the enhanced feature encoding vector set into L levels, denoted as to form the risk-reward encoding vector layer. The specific division method is as follows: Set L - 1 thresholds {τ1, τ2, …, τ L-1}, when τ l-1 <I i ≤τ l , divide z i into the l-th level (where τ0 = -∞, τ L = +∞).

[0174] 3. Feature extraction by embedding the dynamic routing module

[0175] For each risk-reward encoding vector layer Convert it into a tensor form N l is the number of samples in the l-th level. The embedded dynamic routing module consists of multiple dynamic routing layers. At the t-th dynamic routing layer, for each encoding vector in the l-th level calculate its coupling coefficient b with the capsules in the next layer lij : where, is a learnable weight matrix, d c is the capsule dimension, is the output vector of the j-th capsule in the t-th layer. By iteratively updating the coupling coefficient, obtain the final coupling coefficient and then calculate the input vector of the j-th capsule in the t-th layer After passing through the activation function (such as the squash function), obtain the capsule output vector After passing through multiple dynamic routing layers, extract the key risk-reward features of ecological asset elements C is the total number of capsules.

[0176] 4. Mining the Propagation Law of the Graph Neural Network Module

[0177] Represent the ecological asset relationship graph as a graph structure where the nodes correspond to ecological asset elements, and the edge ε represents the association relationship between elements. Use the key risk-return features of ecological asset elements as node features to construct a node feature matrix

[0178] The graph neural network module uses a graph convolutional network (GCN). At the l-th layer, the feature update formula for node i is: where is the feature vector of node i at the l-th layer, is the set of neighboring nodes of node i, c ij is the normalization coefficient between nodes i and j, is the weight matrix at the l-th layer, is the bias vector, and σ is the activation function (such as ReLU). After passing through multiple layers of graph convolutional networks, the final feature vectors of each node are obtained, the propagation law of different levels of risk-return coding vector layers in the relationship graph is mined, and the semantic features between ecological credit evaluation factors are generated d s is the dimension of the semantic features.

[0179] 5. Capturing Distance Dependence of the Transformer Module

[0180] Input the semantic features S between ecological credit evaluation factors into the Transformer module. In the encoder layer of the Transformer, for each semantic feature vector s i ∈ S, calculate its position encoding p i , and obtain the input vector x with position information i = s i + p i .

[0181] After passing through the multi-head attention mechanism and the feed-forward neural network layer, calculate the attention score matrix A: where Q = W q X, K = W k X, V = W v X are the query matrix, key matrix, and value matrix respectively, obtain the attention output Z: Z = AV. After passing through the feed-forward neural network layer and the layer normalization operation, capture the distance dependence features between ecological credit evaluation factors d d is the dimension of the distance dependence features.

[0182] 6. Feature Fusion and Model Training ​

[0183] Fuse the semantic feature S and the distance-dependent feature D among the ecological credit evaluation factors to obtain the feature evaluation feature vector F: F = W f [S T ; D T T Among them, is the fusion weight matrix, and d f is the dimension of the fused feature. Input the feature evaluation feature vector F into the output layer of the pre-trained neural network model to output the ecological credit evaluation prediction result C label is the number of ecological credit label categories. Define the loss function L, which adopts the weighted sum of cross-entropy loss and mean square error loss: L = λ1L CE + λ2L MSE Among them, is the true ecological credit label, and λ1 and λ2 are weight coefficients. Based on the loss value L, calculate the gradient through the backpropagation algorithm Adopt the AdamW optimization algorithm to update the parameters θ of the pre-trained neural network model: Among them, β is the learning rate, and λ wd is the weight decay coefficient. Continuously iterate and train until the loss value reaches the preset model convergence condition (such as L < ∈, where ∈ is the convergence threshold) to obtain the final ecological credit evaluation model.

[0184] Therefore, the above solution integrates cutting-edge technologies such as multi-head self-attention, dynamic routing, and graph neural networks, and has the following technical advantages compared with traditional technologies in terms of feature processing, relationship modeling, model optimization, etc.:

[0185] 1. Traditional feature fusion mostly adopts the linear combination method with fixed weights. For example, in ecological credit evaluation, simply adding risk indicators and return indicators according to empirical weights cannot adapt to the dynamic changes of feature importance in different samples and different scenarios. This method ignores the innovative interaction relationship between features and leads to low information utilization efficiency.

[0186] This application introduces the multi-head self-attention mechanism. By calculating the query matrix Q h , the key matrix K h , and the value matrix V h , dynamically generate the attention score matrix A h . In ecological asset risk assessment, this mechanism can automatically capture the correlation strength between the risk coding vector r and the return coding vector g in different dimensions. For example, when the market fluctuates violently, the model will enhance the weight of risk-related features and weaken the interference of return features. The multi-head design can also extract feature information from multiple angles, and can fuse risk and return features more comprehensively and flexibly than traditional methods, improving the accuracy of feature representation.​

[0187] 2. Traditional hierarchical methods rely on manual experience to divide data levels. For example, ecological assets are simply divided into high, medium, and low risk levels based on fixed risk thresholds, which cannot reflect the dynamic coupling relationship between return potential and risk levels and are difficult to adapt to changes in data distribution.

[0188] This application is based on risk level data R level and return potential data G level , and calculates the hierarchical index I through normalization and weighted calculation to achieve data-driven dynamic stratification. In the green finance scenario, when the return potential of a certain ecological asset is improved due to policy support, this method can adjust its belonging level in real time, making the risk-return coding vector layer more in line with the actual situation. Cross-validation determines the weight coefficients ω1 and ω2 to further optimize the stratification strategy. Compared with traditional methods, it can more accurately depict the risk-return characteristic distribution of ecological assets and provide a better data structure for subsequent analysis.

[0189] 3. Traditional neural networks extract features through fully connected layers or simple convolutional layers, lacking effective modeling of the feature hierarchical structure and spatial relationships. For example, when dealing with the innovative features of ecological assets, it is difficult to distinguish the importance and interaction of different factors.

[0190] In this application, the embedded dynamic routing module uses the principle of capsule networks to achieve dynamic information transfer between the coding vector and the capsule through iterative update of the coupling coefficient . In the extraction of key risk-return features of ecological assets, this module can capture the hierarchical structure of features. For example, it stratifies the macro policy impact and the micro enterprise operation features to avoid information aliasing. The squash function ensures that the direction and length of the capsule output vector have practical significance. Compared with traditional methods, it can more effectively extract key features with structured information and enhance the model's expression ability for innovative ecological asset features.

[0191] 4. Traditional machine learning models mostly assume that data samples are independent of each other and cannot effectively handle the innovative correlation relationships between ecological assets. For example, when evaluating regional ecological credit, the cooperation and competition relationships between enterprises are ignored, which have an impact on credit.

[0192] In this application, the graph neural network module is based on the ecological asset relationship graph , and updates the node features through the graph convolutional network (GCN). In ecological credit evaluation, the feature update formula of node i takes into account the neighboring nodes The information can uncover the propagation rules of different ecological assets in the relationship graph. For example, when a certain enterprise defaults on its credit, through graph convolution operations, the model can automatically spread the impact of this information on the credit of its associated enterprises. Compared with traditional methods, it can more realistically simulate the interactions between ecological assets, improving the accuracy and interpretability of credit evaluations.

[0193] 5. Traditional sequence modeling methods (such as recurrent neural networks) are difficult to capture long-distance dependencies and lack effective modeling of the synergistic effects of different types of features during feature fusion, resulting in insufficient comprehensive analysis capabilities for innovative factors in ecological credit evaluations.

[0194] In this application, the Transformer module can capture long-distance dependence features among ecological credit evaluation factors through the multi-head attention mechanism and positional encoding. In the risk assessment of ecological assets, it can analyze the impact of ecological policies at different times and in different regions on the credit of current assets. Combining the fusion strategy of semantic feature S and distance dependence feature D, through the weighting matrix W f realizes multi-feature collaboration. Compared with traditional methods, it can more comprehensively consider various factors and their interrelationships in ecological credit evaluations, optimize the model prediction performance, and enhance the reliability and stability of ecological credit evaluations.

[0195] The above description is only a preferred embodiment of this application and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by the specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for constructing an ecological credit evaluation model, characterized in that, Including: Collecting green finance activity data of the evaluation subject; Analyzing the green finance activity data to obtain a subject portrait feature library of the evaluation subject, where the subject portrait feature library includes the basic information of the evaluation subject, its behavioral pattern features and preference features in the field of green finance; Assigning a multi-dimensional identity identifier set to the evaluation subject, where the multi-dimensional identity identifier set includes a composite subject ID of a time stamp, a regional identifier, a subject type identifier, and a unique serial number; Constructing an evaluation subject portrait based on the subject portrait feature library of the evaluation subject and the multi-dimensional identity identifier set; Based on the constructed causal inference model, mining the internal factor features and external factor features that affect the green finance behavior and ecological credit of the evaluation subject according to the evaluation subject portrait; Training a pre-trained neural network model according to the internal factor features and external factor features until the training is completed to obtain an ecological credit evaluation model.

2. The construction method of an ecological credit evaluation model according to claim 1, characterized in that Training a pre-trained neural network model according to the internal factor features and external factor features until the training is completed to obtain an ecological credit evaluation model, including: Constructing a network structure with ecological asset elements as associated branches according to the internal and external factor features to form an ecological asset relationship map, and assigning a unique element association ID to each ecological asset element therein; Obtaining a description of the green finance business scenario to construct a scenario simulation module; Based on the scenario simulation module, evaluating the scenario interaction score between the evaluation subject and the ecological asset relationship map to form a multi-dimensional scenario association matrix; Training a pre-trained neural network model according to the multi-dimensional scenario association matrix until the training is completed to obtain an ecological credit evaluation model.

3. The construction method of an ecological credit evaluation model according to claim 2, characterized in that, Training a pre-trained neural network model according to the multi-dimensional scenario association matrix until the training is completed to obtain an ecological credit evaluation model, including: Obtaining the real-time performance evaluation value of the evaluation subject in green finance activities and the actual contribution evaluation value of ecological asset elements to green finance business; Calculating dynamic weights for ecological asset elements based on the real-time performance evaluation value and the actual contribution evaluation value to calibrate the multi-dimensional scenario association matrix; Training the pre-trained neural network model according to the calibrated multi-dimensional scenario association matrix until the training is completed to obtain an ecological credit evaluation model.

4. The construction method of an ecological credit evaluation model according to claim 3, characterized in that Training the pre-trained neural network model according to the calibrated multi-dimensional scenario association matrix until the training is completed to obtain an ecological credit evaluation model, including: Quantitatively coding the risks and benefits brought by ecological asset elements according to the calibrated multi-dimensional scenario association matrix to obtain a risk coding vector and a benefit coding vector; Training by inputting the risk coding vector and the benefit coding vector into the pre-trained neural network model until the training is completed to obtain an ecological credit evaluation model.

5. The construction method of an ecological credit evaluation model according to claim 4, characterized in that, Quantitatively coding the risks and benefits brought by ecological asset elements according to the calibrated multi-dimensional scenario association matrix to obtain a risk coding vector and a benefit coding vector, including: Perform structured and unstructured data extraction processing on the calibrated multi-dimensional scenario correlation matrix, separate the structured data in the matrix, and collect the unstructured data in green finance activities to obtain the data set to be fused; Extract features from the text data in the unstructured data to obtain text feature vectors; Perform feature fusion on the structured data and text feature vectors to obtain enhanced scenario correlation data; Perform data integration processing on the enhanced scenario correlation data and the timestamp and regional identification information in the multi-dimensional identity identification set of the evaluation subject, associate the timestamp and regional identification information with the enhanced scenario correlation data to obtain spatio-temporal correlation data; According to the spatio-temporal correlation data, mine the influence probability of factors on risks and returns to screen out core factors; Perform quantum computing principle mapping processing on the screened core factors, map the core factors into the form of qubits that simultaneously represent multiple risk and return states to obtain the quantum factor representation; Perform quantum entanglement phenomenon simulation processing on the quantum factor representation, establish an innovative non-linear correlation between factors based on the correlation relationship between factors, and construct a quantum-inspired risk-return quantification model; Perform quantum solution on the quantum-inspired risk-return quantification model to calculate the quantified values of risks and returns brought to ecological asset elements; Input the quantified values into the constructed risk coding generator and return coding generator to generate initial risk coding vectors and return coding vectors; Perform discriminator judgment processing on the generated risk coding vectors and return coding vectors to obtain risk coding vectors and return coding vectors; 6. The construction method of an ecological credit evaluation model according to claim 5, characterized in that, Based on the risk coding vectors and return coding vectors, input them into a pre-trained neural network model for training until the training is completed to obtain an ecological credit evaluation model; Perform self-attention mechanism fusion processing on the risk coding vectors and return coding vectors to generate an enhanced feature coding vector set; Perform stratified sampling processing on the enhanced feature coding vector set, and divide the enhanced feature coding vector set into different levels according to the risk level and return potential data of ecological asset elements to obtain the risk-return coding vector layer; Input the risk-return coding vector layer into the embedded dynamic routing module in the pre-trained neural network model to extract the key risk-return features of ecological asset elements; Input the key risk-return features of ecological asset elements into the graph neural network module in the pre-trained neural network model, mine the propagation law of different levels of risk-return coding vector layers in the relationship graph, and generate the semantic features between ecological credit evaluation factors; Input the semantic features between ecological credit evaluation factors into the transformer module in the pre-trained neural network model to capture the distance-dependent features between ecological credit evaluation factors; Fuse the semantic features and distance-dependent features between ecological credit evaluation factors into feature evaluation feature vectors, and input them into the output layer in the pre-trained neural network model to generate an ecological credit evaluation prediction result and compare it with the true ecological credit label to calculate the loss value; Based on the loss value, optimize the model parameters of the pre-trained neural network model through backpropagation until the loss value reaches the preset model convergence condition, then end the training to obtain the ecological credit evaluation model.

Images