Renewable energy green evidence data multi-dimensional classification and monitoring management method

Abstract

The invention discloses a renewable energy source green evidence data multi-dimensional classification and dynamic analysis management method, and relates to the technical field of renewable energy source data management. Green evidence data is collected in real time, basic, policy and transaction attribute labels are extracted, and a label pair set is formed; a feature group list is obtained through double threshold screening, clustering and semantic fusion, feature group attributes are matched with an analysis method mapping table to obtain a corresponding analysis method, feedback scores are calculated through accuracy, decision-making adoption rate and efficiency scores, and an optimization strategy is determined. The problems that in the prior art, the green evidence data classification dimension is single, the dynamic analysis capacity is insufficient, the method matching is lack of systematicness, and a feedback optimization mechanism is lacked are solved, and the data utilization efficiency and the decision accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of renewable energy data management, and in particular to a method for multi-dimensional classification and dynamic analysis management of renewable energy green certificate data. Background Art

[0002] In the renewable energy sector, Green Certificates (GCEs) serve as proof of renewable energy power production and consumption. Effective data management is crucial for energy market transactions. However, existing technologies for GCE data classification and dynamic analysis and management have the following deficiencies: The data classification dimension is single, mostly based on static classification of basic attributes, lacking multi-dimensional coupling analysis of policy attributes and transaction attributes, making it difficult to capture the dynamic changes in the value of green certificates and the interactive impact of policies and markets; Insufficient dynamic analysis capabilities. Traditional methods lack a time window mechanism to perform real-time statistics on co-occurrence patterns of tag pairs. They are unable to adapt to the evolution of data association patterns caused by policy adjustments or changes in market rules. They are prone to mixing in invalid historical data, causing analysis results to lag behind actual business scenarios. The matching strategy between analysis methods and data features lacks systematicity. Existing technologies do not establish a mapping mechanism for analysis methods based on attribute combinations. Instead, they often use a single analysis method, making it difficult to select an adaptation method based on the semantic associations and business needs of different attribute combinations, resulting in low analysis efficiency. Therefore, a multidimensional classification and dynamic analysis management method for renewable energy green certificate data is urgently needed to improve data utilization efficiency and decision-making accuracy. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention provides a multi-dimensional classification and dynamic analysis management method for renewable energy green certificate data, which solves the problems of single dimension classification and analysis of green certificate data and insufficient dynamic management.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: a multi-dimensional classification and dynamic analysis management method for renewable energy green certificate data, including: Step 1: Collect green certificate data in real time, extract basic attribute tags, policy attribute tags, and transaction attribute tags to form a tag pair set; Step 2: Count the co-occurrence times and proportions of the tag pairs within the time window T, filter the high-frequency co-occurrence tag pairs using a double threshold, and obtain a feature group list through clustering and semantic fusion. Step 3: Obtain the attributes of each feature group in the feature group list, and match the analysis method based on the feature group attributes and the analysis method mapping table; Step 4: For each feature group, obtain the analysis method used, its accuracy, decision adoption rate, and efficiency score, calculate the feedback score, and determine the corresponding optimization strategy.

[0005] As a further solution of the present invention, the basic attribute tags specifically include renewable energy type, power generation amount, and generation time. The policy attribute tags specifically include subsidy policy category, certification standard, and regional policy attribute. The transaction attribute tags specifically include transaction region, transaction volume, and transaction subject type.

[0006] As a further solution of the present invention, within window T, calculate the co-occurrence count C of each tag pair and the co-occurrence data proportion P = C / N. Set the co-occurrence count threshold as Cth and the co-occurrence data proportion threshold as Pth. If C > Cth and P > Pth, then filter it into the high-frequency co-occurrence tag pair set, where N is the total number of record pairs.

[0007] As a further solution of the present invention, calculate the co-occurrence data proportion threshold according to the formula Pth = Pmin + (Pmax - Pmin) × exp(-k × Cth), where Pmin is the lower limit of the proportion, Pmax is the upper limit of the proportion, k is the attenuation coefficient, and exp() is the function with base e.

[0008] As a further solution of the present invention, the specific steps for obtaining the feature group list through clustering and semantic fusion are as follows: Use the K-means++ algorithm to initialize the clustering center for the high-frequency co-occurrence tag pair set; Use the ISODATA algorithm to perform dynamic adjustment on the clusters obtained through clustering. The specific operation is as follows: If the number of samples in a certain cluster < Nmin or the dispersion exceeds the threshold miu, then split it into two sub-clusters; If the distance between the centers of two clusters < Dmerge, then merge the two clusters; Among them, Nmin is the minimum number of samples to avoid isolated small clusters, Dmerge is the merging distance threshold, based on the Euclidean distance of the feature vectors; Use the BERT pre-trained model to encode the tag pairs in each cluster into 384-dimensional semantic vectors, according to the formula Calculate the average cosine similarity between clusters, where m and n are the number of samples in two clusters; If the average similarity between clusters >= Simth and the sum of co-occurrence counts >= Cmerge, trigger the merging between the two clusters, where Simth is the similarity threshold and Cmerge is the co-occurrence count threshold; After merging, recalculate the cluster centroid and update the samples within the cluster; If the number of clusters remains unchanged for 3 consecutive iterations, or reaches the maximum number of iterations, then terminate the iteration; Finally, output the feature group list.

[0009] As a further solution of the present invention, the co-occurring label pairs in the high-frequency co-occurring label pair set are converted into one-hot encoding vectors.

[0010] As a further solution of the present invention, the specific steps of obtaining the feature group attributes are: The occurrence ratio of each attribute in all the label pairs of the feature group is calculated according to the formula Fu(Ai)=(number of label pairs containing attribute Ai / total number of label pairs in the feature group), where Fu(Ai) is the coverage of each attribute, i takes values ​​of 1, 2, and 3, A1 represents basic attributes, A2 represents policy attributes, and A3 represents transaction attributes; If Fu(Ai)>=Futh, then the attribute is retained, otherwise it is removed. The specific value range of Futh is [30%, 50%].

[0011] As a further solution of the present invention, the specific contents of the analysis method mapping table include: If it is the basic attribute, select cluster analysis + principal component analysis PCA; If it is a policy attribute, select Policy Association Network Analysis + Compliance Verification; If it is a transaction attribute, select transaction pattern mining + predictive analysis; If it is basic attributes + policy attributes, select the policy impact assessment model; If it is basic attributes + transaction attributes, select market segmentation and pricing strategy analysis; If it is policy attributes + transaction attributes, select policy compliance risk assessment; If it is basic attributes + policy attributes + transaction attributes, choose a comprehensive value assessment model.

[0012] As a further solution of the present invention, the specific steps of calculating the feedback score are: The three variables of method accuracy AA, decision adoption rate BB, and efficiency score CC are normalized to obtain the corresponding AA', BB', and CC'; According to the formula Calculate the entropy value of each variable, where , xij is the jth variable value of the i-th feature group, n is the sample size, j∈{AA',BB',CC'}; According to the formula Determine the weight of each variable; Map the feedback variable of each feature group to a three-dimensional space point X 反馈 =(AA',BB',CC'), pre-set X 理想 =(1,1,1); Calculate X 反馈 With X 理想 The cosine similarity Sim; According to the formula Grade=(X 反馈 The feedback score is calculated by using the formula (·W)×Sim×100, where · represents the dot product between two vectors and W is a vector consisting of the weights of the three variables.

[0013] As a further solution of the present invention, the specific steps of determining the optimization strategy according to the feedback score are: If 80<=Grade<=100, then this method will be included in the default analysis method of the new feature group; If 60<=Grade<80, check for abnormal independent variables or switch to an alternative method; If Grade < 60, redefine the feature group attributes or start cross-method fusion.

[0014] The present invention provides a method for multi-dimensional classification and dynamic analysis and management of renewable energy green certificate data, which has the following beneficial effects compared with the existing technology: (1) This invention uses multi-dimensional attribute labels for classification and dynamic clustering to achieve full-dimensional characterization of green certificate data, solve the problem of single dimension in traditional classification, accurately capture the interactive relationship between policies and markets, and improve the level of data structuring; (2) The present invention improves the analysis efficiency and accuracy by constructing an intelligent mapping mechanism between feature groups and analysis methods and matching strategies based on attribute groups; (3) The present invention introduces the entropy weight-cosine similarity feedback model and the hierarchical optimization strategy, and forms a closed-loop iteration of the analysis method through dynamic verification of the three dimensions of accuracy, adoption rate, and efficiency, so as to continuously adapt to the dynamic changes of the green certificate market. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of the steps of the present invention. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] like Figure 1 The present invention provides a method for multi-dimensional classification and dynamic analysis management of renewable energy green certificate data, including: Step 1: Collect green certificate data in real time, extract its basic attribute labels, policy attribute labels, and transaction attribute labels, and combine the obtained labels in pairs to form a label pair set corresponding to each green certificate; The basic attribute tags specifically include renewable energy type, power generation, and generation time; Renewable energy type. Energy type is the core physical attribute of green certificates, which directly affects their market value, policy subsidies and trading demand; Power generation is the direct carrier of the value of green certificates and reflects energy production capacity. Differences in scale lead to different market behaviors. Market participants can be divided into different types through clustering. For example, large-scale power generation companies are more inclined to long-term agreement transactions. It can also be combined with generation time to identify peak and trough periods of power generation, such as high photovoltaic power generation in summer, supporting supply and demand balance analysis; Generation time: The timeliness of the green certificate directly affects its market value; Recently generated green certificates are more popular. They can capture seasonal power generation characteristics from a temporal perspective. For example, wind power generation is high in winter. Combined with trading data, they can predict price fluctuation cycles. The mapping of policy changes on the time axis can be used to assess the lagged impact of policies on power generation behavior. Policy attribute tags specifically include subsidy policy categories, certification standards, and regional policy attributes; Subsidy policy category: Subsidies are a key driving force for the development of renewable energy and directly affect the production cost and market price of green certificates; Subsidies at different levels form differentiated markets. Clustering can reveal subsidy-trading patterns, such as the tendency of national subsidy projects to trade across regions. Furthermore, the time series correlation between subsidy policy changes and power generation behavior supports policy effectiveness evaluation. Certification standards determine the credibility and circulation of green certificates and are a key threshold for international market transactions. Internationally certified green certificates frequently co-occur with cross-border transactions and carbon footprint declarations, and clustering can be used to identify high-end market participants. Regional policy attributes: different regions have significant differences in their positioning and policies on green certificates. For example, provinces piloting the carbon market have a higher demand for green certificates. Combining regional policy attributes with trading areas, we can discover policy-driven cross-regional flow patterns, such as the flow of green certificates from non-pilot provinces to pilot provinces; Transaction attribute tags specifically include transaction area, transaction volume, and transaction entity type; Trading area: The trading area reflects the geographical distribution of supply and demand of green certificates and is a key dimension for analyzing market liquidity; The correlation between inter-regional price differences and transaction flows supports cross-regional arbitrage analysis and is also combined with energy types to identify the match between regional resource endowments and market demand; Trading volume, which directly reflects the market demand and liquidity of green certificates; Combining transaction volume with transaction entity type, such as large-scale transactions by power generation groups versus scattered transactions by small and medium-sized enterprises, can help identify behavioral patterns of market participants. From a temporal perspective, fluctuations in transaction volume over time series are associated with policy changes, paving the way for further prediction of market turning points. The types of trading entities: different trading entities have different motivations and behavior patterns, which affect the market price formation mechanism; Power generation companies are more concerned about subsidy policies and long-term agreements, while power purchasing companies are more concerned about green certificate prices and environmental benefits. Market segments can be divided through clustering; Combining entity type with trading area can identify high-end market demand.

[0018] Step 2: Count the co-occurrence times and proportions of the green certificate tag pairs within the time window T, and filter out high-frequency co-occurrence tag pairs using a double threshold. Policy adjustments or changes in market rules can lead to dynamic evolution of label co-occurrence patterns. Statistics within the time window T can prevent historical data from interfering with current association analysis. For example, in Q2 2024, a province adjusted its photovoltaic subsidies from provincial to national levels. T = co-occurrence statistics for the past 30 days can accurately capture the high-frequency co-occurrence of "photovoltaic-national subsidies", while long-term windows will be mixed with invalid co-occurrence data of "photovoltaic-provincial subsidies" in Q1. Green certificate data may fluctuate significantly due to the centralized entry at the end of the month. The co-occurrence ratio can be normalized by the total amount within the time window to avoid the problem of "the surge in data volume leads to an artificially low ratio of low-correlation label pairs, and the sudden decrease in data volume leads to an artificially high ratio of accidental co-occurrences"; For example, the difference in the proportion of a certain label pair in windows T1 (N=1000, 50 occurrences, 5% of the total) and T2 (N=5000, 200 occurrences, 4% of the total) can truly reflect the change in its association strength with the total amount of data; Calculate the co-occurrence count C and co-occurrence data ratio P = C / N for each tag pair in window T, where N is the total number of record pairs. Set the co-occurrence count threshold to Cth and the co-occurrence data ratio threshold to Pth. If C>Cth and P>Pth, filter it out to obtain a set of high-frequency co-occurrence tag pairs. The co-occurrence data ratio threshold is calculated according to the formula Pth=Pmin+(Pmax-Pmin)×exp(-k×Cth), where Pmin is the lower limit of the ratio to prevent the threshold from tending to 0, Pmax is the upper limit of the ratio, and k is the attenuation coefficient, which can be calibrated by historical data. If Cth is in the low frequency range, such as Cth < 50, the initial slope of the exponential function is large, and the proportion threshold decreases rapidly with the increase of the frequency, ensuring that a high proportion must be met at low frequencies to filter out accidental co-occurrences; For example, when the total number of records N = 300, the co-occurrence of 20 times needs to account for 6.7% to pass the screening, avoiding false positives caused by a small N; If Cth is in the high-frequency range, such as Cth > 100, the exponential function approaches the horizontal, and the proportion threshold decreases slowly, allowing a low proportion to pass at high frequencies; For example, when N = 30000, the co-occurrence of 200 times accounts for 0.6%, but the absolute frequency is sufficient to indicate a strong correlation; Convert the co-occurring label pairs in the high-frequency co-occurrence label pair set into one-hot encoded vectors. 9 dimensions correspond to 3 types of labels. For example, "wind power - Hebei" corresponds to [1, 0, 0, 0, 0, 0, 1, 0, 0]; One-hot encoding retains the information of the large label categories, that is, whether the basic, policy, and trading attributes appear, facilitating the calculation of the category differences of label pairs; Use the K-means++ algorithm to initialize the clustering centers, and preferentially select label pairs with a larger distance as the initial centroids to avoid local optima caused by random selection; Use the ISODATA algorithm to dynamically adjust the clusters obtained by clustering. The specific operations are as follows: If the number of samples in a cluster < Nmin or the dispersion exceeds the threshold miu, it is split into two sub-clusters; If the distance between the centers of two clusters < Dmerge, the two clusters are merged; Among them, Nmin is the minimum number of samples to avoid isolated small clusters, and Dmerge is the merging distance threshold, based on the Euclidean distance of feature vectors; Obtain the initial clustering result, and each cluster contains a set of semantically related label pairs; The green certificate data is dynamic. The number of clusters in traditional K-means clustering is fixed. ISODATA can automatically adjust the number of clusters. Its splitting mechanism avoids noise clusters (isolated clusters with few samples), and the merging mechanism eliminates redundant clusters (similar clusters with small differences); Use the BERT pre-trained model to encode the label pairs in each cluster into 384-dimensional semantic vectors, according to the formula Calculate the average cosine similarity between clusters. Among them, m and n are the number of samples in two clusters; Statistical co-occurrence can only capture surface associations, and semantic similarity can mine implicit relationships. For example, "carbon trading mechanism" and "international certification" do not directly co-occur, but their semantics are both related to "carbon market"; The BERT model is pre-trained on a large number of policy texts and energy reports, and can accurately capture semantic associations in the field of green certificates; Too low a dimension (e.g., 50 dimensions) will lose semantic details, while too high a dimension (e.g., 768 dimensions) retains more information but increases computational complexity. 384 dimensions have been proven in practice to achieve a good balance between semantic preservation (capturing the domain association and contextual semantics of label pairs) and computational efficiency (reducing vector operation overhead). When the average similarity between clusters is greater than or equal to Simth and the sum of co-occurrence times is greater than or equal to Cmerge, the merging of the two clusters is triggered; Avoid over-abstraction caused by purely semantic merging, such as merging unrelated but semantically vaguely similar label pairs. It is necessary to combine statistical thresholds (i.e., co-occurrence counts) to ensure that the association is supported by actual data. After merging, recalculate the cluster centroid, take the mean of the semantic vector, and update the samples within the cluster; If the number of clusters remains unchanged for three consecutive iterations or the maximum number of iterations is reached, the iteration is terminated; The final output is a list of feature groups, each of which contains semantically related label pairs.

[0019] Step 3: After obtaining the feature group list, match each feature group attribute with the analysis method mapping table to obtain the corresponding analysis method; The specific method for obtaining feature group attributes is: The occurrence ratio of each attribute in all the label pairs of the feature group is calculated according to the formula Fu(Ai)=(number of label pairs containing attribute Ai / total number of label pairs in the feature group), where Fu(Ai) is the coverage of each attribute, i takes values ​​of 1, 2, and 3, A1 represents basic attributes, A2 represents policy attributes, and A3 represents transaction attributes; Only attributes with coverage greater than or equal to Futh are retained. The threshold can be adjusted according to the business scenario, with a specific value range of [30%, 50%], to filter out occasional secondary attributes; Enforce that attribute coverage is greater than or equal to the threshold to ensure that the extracted features are the mainstream business attributes of the feature group. For example, a feature group based on "basic + transaction" will not contain "policy" attributes due to individual policy label pairs. Only one coverage calculation is added, and the time complexity is O(n), where n is the number of label pairs, which is suitable for large-scale data processing; For example, a feature group E contains 8 label pairs, of which 5 are {foundation + transaction} (such as "wind power - Guangdong" and "photovoltaic - Jiangsu"), 2 are {foundation + policy} (such as "wind power - national subsidy" and "photovoltaic - provincial subsidy"), and 1 is {policy + transaction} (such as "national subsidy - inter-provincial transaction"). The coverage of the basic attributes is calculated to be 100%, the coverage of the transaction attributes is 75%, the complexity of the policy attributes is 37.5%, and the coverage threshold Futh is set to 40%. The policy attributes are filtered out, and the final attribute of feature group E is {foundation, transaction}. The specific contents of the analysis method mapping table are as follows: Basic attributes, all label pairs belong to the category of basic attributes (such as "renewable energy type", "power generation", "equipment parameters", etc.). The data usually has high dimensions, strong correlation, and the inherent structure needs to be explored. The processing method selected is: cluster analysis + principal component analysis PCA; Basic attribute data often have implicit semantic or logical associations. For example, "wind power" and "solar energy" belong to the same energy type. Cluster analysis (such as K-means and DBSCAN) can automatically group similar label pairs based on the similarity of feature vectors without pre-defining categories, which meets the needs of exploratory analysis of basic attributes. After being encoded by BERT, basic attribute labels may contain high-dimensional features, and there may be strong correlations between different attributes. PCA maps high-dimensional features to low-dimensional principal components through linear transformation, preserving the main variation information of the data while eliminating redundancy. Policy attributes: There are complex relationships between policies, such as references, hierarchies, and conflicts. At the same time, policy updates may render old policies invalid. The solution is to analyze the policy association network and verify compliance. Policy attribute data naturally has graph structure characteristics. Policies can be represented as nodes, and reference or conflict relationships as edges. Network analysis can quantify the strength of the association between policies. For example, in a corporate compliance scenario, "new energy subsidy policy" may be associated with "renewable energy quota system" and "environmental protection acceptance standards." Through network analysis, the policy dependency path can be sorted out; When multiple policies run in parallel, conflicts may arise. Compliance verification can identify conflicts and provide priority recommendations. Transaction attributes: Transaction data changes in real time with time, market rules, and supply and demand relationships. Similar transactions often show regular patterns, and future trends can be predicted based on historical patterns. The processing method selected is: transaction pattern mining + predictive analysis; There are a large number of recurring patterns in transaction data. Pattern mining can quantify these regularities and provide a basis for prediction. Trading data (such as price and volume) has strong time series characteristics, and forecasting models (such as ARIMA and LSTM) can capture trends, seasonality, and cyclicality. By predicting trading outcomes under different market conditions, such as the impact of rising carbon prices on green certificate trading, it assists in formulating trading strategies and helps companies avoid risks when extreme situations are predicted; Basic attributes + policy attributes. The policy effect depends on the adaptability of the basic attributes. The same policy may produce significantly different results under different basic conditions. The method selected for processing is: policy impact assessment model; By comparing the differences between scenarios with and without policies, we can strip away the interference of other factors and evaluate the impact of policies on basic attributes separately. Policy impacts often involve multiple dimensions, including economic, social, and environmental. Models must combine basic attributes to construct a multidimensional assessment framework. The policy impact is constructed by taking the basic attributes as constraints. The specific formula can be expressed as policy impact = f(policy intensity, basic attributes, random disturbance). The policy tool is parameterized by the model and the state changes after the policy implementation are predicted by combining the basic attributes. Basic attributes + transaction attributes: Users with similar basic attributes may present different values ​​due to differences in transaction behavior. The method for selecting solutions is: market segmentation and pricing strategy analysis; By intersecting basic attributes with transaction attributes, we can accurately target different demand groups and avoid one-size-fits-all pricing; Transaction attributes such as regional supply and demand, transaction entity type, and policy milestones determine the fluctuation range of the actual transaction price. The benchmark price is determined based on these basic attributes, and a differentiated pricing system is designed based on transaction attributes. Policy attributes + transaction attributes. Policies are the rules of transactions. Transactions must be conducted within the policy framework. Any transaction behavior that deviates from the policy may trigger compliance risks. The method of handling is: policy compliance risk assessment. If green certificate transactions deviate from the policy framework, it may lead to risks such as invalid transactions, damaged credit, and economic losses. Through quantitative assessment models, abstract policies can be transformed into executable transaction compliance indicators; For example, scoring items such as compliance of green certificate sources, compliance of trading platforms, and compliance of write-off processes can be set. Each item corresponds to a specific policy clause. When the policy is updated, the system automatically marks expired green certificates in existing transactions, prompting companies to deal with them in a timely manner. Basic attributes + policy attributes + transaction attributes. The value of green certificates is not the product of a single attribute, but the dynamic coupling result of basic attributes, policy attributes, and transaction attributes. Basic attributes are the foundation of value, policy attributes determine the upper limit of value, and transaction attributes reflect the path to value realization. The method of selecting processing is: comprehensive value assessment model; Single attribute evaluation cannot capture the linkage effect of value transmission. The interactive impact of different attributes can be quantified by constructing a linkage formula of basic attribute × policy coefficient × transaction factor. Green certificate policies and trading markets are highly time-sensitive. Single-dimensional evaluation can easily lead to delayed decision-making. By embedding a policy early warning module and a real-time interface for trading data, the evaluation parameters are automatically adjusted and the value assessment results are updated synchronously when policies are updated or market fluctuations exceed thresholds. Single attribute assessment may mask potential risks, such as policy compliance risk and transaction liquidity risk. A comprehensive model can identify risks in advance. It is necessary to embed "policy compliance verification" and "transaction entity credit scoring" modules into the value assessment system. Step 4: Obtain the accuracy, decision adoption rate, and efficiency score of the analysis method used for each feature group, calculate the feedback score, and determine the corresponding optimization strategy based on the feedback score range. The method accuracy rate AA reflects the degree of consistency between the analysis results and the actual business scenario, avoiding decision-making errors due to technical deviations, and is the basic parameter for measuring the effectiveness of the analysis method. Decision adoption rate (BB) is the ratio of analysis results adopted by business systems or manual decision-makers, which directly reflects the adaptability of the analysis method to real business needs. The efficiency score CC, which measures the execution efficiency of the analysis method and reversely correlates resource consumption, is calculated using the formula (1-actual time consumed / benchmark time consumed) × 100. This ensures the operability of the analysis method in scenarios such as high-frequency trading and real-time analysis, and meets engineering practice requirements. The specific steps for calculating feedback scores are: Normalize the method accuracy AA, decision adoption rate BB, and efficiency score CC. AA and BB are already between [0, 1] and therefore do not need to be changed. However, CC needs to be normalized using the formula CC / 100 to obtain the final AA', BB', and CC'. According to the formula Calculate the entropy value of each variable, where , xij is the jth variable value of the i-th feature group, n is the sample size, j∈{AA',BB',CC'}; According to the formula Determine the weight of each variable. The smaller the entropy value, the higher the variable discrimination and the greater the weight. Map the feedback variable of each feature group to a three-dimensional space point X 反馈 =(AA',BB',CC'), pre-set X 理想 =(1,1,1), all variables are in full score state; To measure how close the feedback is to the ideal state, calculate X 反馈 With X 理想 The cosine similarity Sim; According to the formula Grade=(X 反馈 The feedback score is calculated by ·W)×Sim×100, where · represents the dot product between two vectors and W is a vector consisting of the weights of the three variables; The integration of entropy weight dynamic weight and spatial geometric distance can not only reflect the difference in variable importance, but also avoid the linear limitation of single weighting; The method for determining the corresponding optimization strategy based on the feedback score range is: If 80<=Grade<=100, it indicates that the method has high accuracy, high decision adoption rate, and meets the efficiency standards. The three indicators are balanced and perform well. The following strategies can be adopted: Include this method in the default analysis method for new feature groups; Scores in this range indicate that the analysis method has been verified in actual business scenarios, has stable technical effects and practical value, and does not require repeated debugging; If 60<=Grade<80, it indicates that at least one variable indicator of the method is significantly lower than the threshold, and the technical solution has systemic risks. The following strategies can be adopted: Check for anomalies in the independent variables, such as low accuracy due to incorrect data labels; Switch to an alternative method, such as switching from K-means++ to DBSCAN clustering; If the Grade is less than 60, it indicates that all three indicators fail to meet the standards and the existing method is completely unable to meet business needs. The following strategies can be adopted: Redefine feature group attributes, such as splitting mixed attribute labels; Initiate cross-method integration, such as joint training of policy models and transaction data; Comprehensive failure to meet standards indicates that the original design has fundamental flaws, such as incorrect division of feature group attributes and inappropriate selection of analysis methods, and the underlying logic must be reconstructed.

[0020] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0021] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A multi-dimensional classification and dynamic analysis management method for renewable energy green certificate data, characterized by: It includes: Step1. Collect green certificate data in real time, extract basic attribute tags, policy attribute tags, and transaction attribute tags, and form a set of tag pairs. Step2. Count the co-occurrence times and proportions of the set of tag pairs within the time window T, screen the set of high-frequency co-occurring tag pairs through double thresholds, and obtain a list of feature groups through clustering and semantic fusion. Step3. Obtain the attributes of each feature group in the list of feature groups, and match the analysis method based on the mapping table between the feature group attributes and the analysis methods. Step4. For the analysis method adopted for each feature group, obtain its accuracy rate, decision adoption rate, and efficiency score, calculate the feedback score, and determine the corresponding optimization strategy.

2. The multi-dimensional classification and dynamic analysis management method for renewable energy green certificate data according to claim 1 is characterized in that: The basic attribute tags specifically include the type of renewable energy, power generation, and generation time. The policy attribute tags specifically include the subsidy policy category, certification standard, and regional policy attributes. The transaction attribute tags specifically include the transaction area, transaction volume, and transaction entity type.

3. The multi-dimensional classification and dynamic analysis management method for renewable energy green certificate data according to claim 1 is characterized in that: Calculate the co-occurrence times C and the proportion of co-occurrence data P = C / N of each tag pair within the window T. Set the co-occurrence times threshold as Cth and the proportion of co-occurrence data threshold as Pth. If C > Cth and P > Pth, then screen it into the set of high-frequency co-occurring tag pairs, where N is the total number of record pairs.

4. The multi-dimensional classification and dynamic analysis management method for renewable energy green certificate data according to claim 3 is characterized in that: Calculate the proportion threshold of co-occurrence data according to the formula Pth = Pmin + (Pmax - Pmin) × exp(-k × Cth), where Pmin is the lower limit of the proportion, Pmax is the upper limit of the proportion, k is the attenuation coefficient, and exp() is the function with base e.

5. The multi-dimensional classification and dynamic analysis management method for renewable energy green certificate data according to claim 1 is characterized in that: The specific steps for obtaining the list of feature groups through clustering and semantic fusion are as follows: Use the K-means++ algorithm to initialize the clustering center of the set of high-frequency co-occurring tag pairs. Use the ISODATA algorithm to dynamically adjust the clusters obtained through clustering. The specific operation is as follows: If the number of samples in a certain cluster < Nmin or the dispersion exceeds the threshold miu, then split it into two sub-clusters. If the distance between the centers of two clusters < Dmerge, then merge the two clusters. Among them, Nmin is the minimum number of samples to avoid isolated small clusters, Dmerge is the merging distance threshold, based on the Euclidean distance of the feature vectors. Use the BERT pre-trained model to encode the label pairs in each cluster into a 384-dimensional semantic vector according to the formula Calculate the average cosine similarity between clusters, where m and n are the number of samples in the two clusters; If the average similarity between clusters >= Simth and the sum of co-occurrence times >= Cmerge, trigger the merging between the two clusters, where Simth is the similarity threshold and Cmerge is the co-occurrence times threshold. Recalculate the cluster centroid after merging and update the samples within the cluster. If the number of clusters remains unchanged for 3 consecutive iterations, or reaches the maximum number of iterations, then terminate the iteration. Finally, output the list of feature groups.

6. The method for multi-dimensional classification and dynamic analysis management of renewable energy green certificate data according to claim 5 is characterized in that: Convert the co-occurring tag pairs in the set of high-frequency co-occurring tag pairs into one-hot encoded vectors.

7. The multi-dimensional classification and dynamic analysis management method for renewable energy green certificate data according to claim 1 is characterized in that: The specific steps for obtaining the feature group attributes are as follows: Calculate the occurrence proportion of each attribute in all tag pairs of the feature group according to the formula Fu(Ai) = (the number of tag pairs containing attribute Ai / the total number of tag pairs in the feature group), where Fu(Ai) is the coverage of each attribute, i takes values of 1, 2, 3, A1 represents the basic attribute, A2 represents the policy attribute, and A3 represents the transaction attribute. If Fu(Ai)>=Futh, then the attribute is retained, otherwise it is removed. The specific value range of Futh is [30%, 50%].

8. The multi-dimensional classification and dynamic analysis management method for renewable energy green certificate data according to claim 1 is characterized in that: The specific contents of the analysis method mapping table include: If it is the basic attribute, select cluster analysis + principal component analysis PCA; If it is a policy attribute, select Policy Association Network Analysis + Compliance Verification; If it is a transaction attribute, select transaction pattern mining + predictive analysis; If it is basic attributes + policy attributes, select the policy impact assessment model; If it is basic attributes + transaction attributes, select market segmentation and pricing strategy analysis; If it is policy attributes + transaction attributes, select policy compliance risk assessment; If it is basic attributes + policy attributes + transaction attributes, choose a comprehensive value assessment model.

9. The method for multi-dimensional classification and dynamic analysis management of renewable energy green certificate data according to claim 1 is characterized in that: The specific steps for calculating feedback scores are: The three variables of method accuracy AA, decision adoption rate BB, and efficiency score CC are normalized to obtain the corresponding AA', BB', and CC'; According to the formula Calculate the entropy value of each variable, where , xij is the jth variable value of the i-th feature group, n is the sample size, j∈{AA',BB',CC'}; According to the formula Determine the weight of each variable; Map the feedback variable of each feature group to a three-dimensional space point X 反馈 =(AA',BB',CC'), pre-set X 理想 =(1,1,1); Calculate X 反馈 With X 理想 The cosine similarity Sim; According to the formula Grade=(X 反馈 The feedback score is calculated by using the formula (·W)×Sim×100, where · represents the dot product between two vectors and W is a vector consisting of the weights of the three variables.

10. The method for multi-dimensional classification and dynamic analysis management of renewable energy green certificate data according to claim 1 is characterized in that: The specific steps to determine the optimization strategy based on the feedback score are: If 80<=Grade<=100, then this method will be included in the default analysis method of the new feature group; If 60<=Grade<80, check for abnormal independent variables or switch to an alternative method; If Grade < 60, redefine the feature group attributes or start cross-method fusion.

Images