Review method, device and equipment for new energy access scheme

Abstract

This invention provides a method, apparatus, and equipment for reviewing renewable energy access schemes, relating to the field of power system planning technology. The method includes: acquiring the scheme text and industry standard clauses of the renewable energy access scheme; semantically slicing the scheme text to obtain multiple text segments; semantically slicing the industry standard clauses to obtain multiple minimum constraint units; for each text segment, determining the minimum constraint unit matching the text segment from the multiple minimum constraint units; and determining the review result record for the text segment based on the matching minimum constraint unit; inputting the review result records of each text segment into a large language model to obtain a comprehensive review result of the renewable energy access scheme. This invention can improve the review efficiency of renewable energy access schemes, avoid the influence of subjective experience, and enhance the objectivity and standardization of review conclusions.

Description

Technical Field

[0001] This invention relates to the field of power system planning technology, and in particular to a method, apparatus and equipment for reviewing new energy access schemes. Background Technology

[0002] With the rapid growth in the scale of distributed photovoltaic, wind power, and other new energy sources connected to the grid, grid planning and operation departments face an urgent need for efficient and standardized technical reviews of massive grid connection schemes. The review work needs to cover multiple dimensions, including grid connection point site selection, voltage level matching, grid capacity, protection configuration, compliance of electrical simulation results (such as power flow, short circuit, N-1 check), and policy and economic analysis. It is a comprehensive task involving complex technical specifications and empirical judgment.

[0003] Currently, the mainstream practice in this field still heavily relies on manual review and analysis by professional and technical personnel. Reviewers need to carefully study the renewable energy grid connection scheme reports and then compare them one by one with technical standards, design guidelines, and policy documents to make compliance judgments. This approach is not only time-consuming and labor-intensive with a long review cycle, but it is also difficult to ensure the consistency of review results among different reviewers or at different times, and it is easily influenced by subjective experience. Summary of the Invention

[0004] This invention provides a method, apparatus, and equipment for reviewing new energy access schemes, in order to solve the problems of low review efficiency and susceptibility to subjective experience in the review of new energy access schemes.

[0005] In a first aspect, embodiments of the present invention provide a method for reviewing new energy access schemes, including: Obtain the scheme text and industry standard clauses of the new energy access scheme, perform semantic segmentation on the scheme text to obtain multiple text fragments, and perform semantic segmentation on the industry standard clauses to obtain multiple minimum constraint units; For each text segment, determine the minimum constraint unit that matches the text segment from the plurality of minimum constraint units, and determine the review result record for the text segment based on the minimum constraint unit that matches the text segment; The review results of each text segment are recorded and input into the large language model to obtain the comprehensive review result of the new energy access scheme; the large language model is used to perform inductive reasoning on the review results of each text segment and output the comprehensive review result.

[0006] Optionally, the semantic slicing of the scheme text yields multiple text fragments, including: Based on the punctuation marks in the proposed solution text, the proposed solution text is sequentially divided into multiple smallest semantic units; For each pair of adjacent smallest semantic units, the existence of a semantic boundary between each pair of adjacent smallest semantic units is determined sequentially according to the preset structured tags, preset boundary keywords, and the semantic similarity between each pair of adjacent smallest semantic units. According to the semantic boundaries, all the smallest semantic units in the scheme text are divided into multiple text segments.

[0007] Optionally, for each pair of adjacent smallest semantic units, determining whether there is a semantic boundary between each pair of adjacent smallest semantic units according to preset structured tags, preset boundary keywords, and the semantic similarity between each pair of adjacent smallest semantic units includes: For every two adjacent smallest semantic units, if there is a preset structured marker between the two adjacent smallest semantic units, then it is determined that there is a semantic boundary between the two adjacent smallest semantic units. If there is no preset structured marker between the two adjacent smallest semantic units, then it is detected whether there is a preset boundary keyword in the two adjacent smallest semantic units; If a preset boundary keyword exists in two adjacent smallest semantic units, then it is determined that there is a semantic boundary between the two adjacent smallest semantic units. If there is no preset boundary keyword in the two adjacent smallest semantic units, then it is detected whether the semantic similarity between the two adjacent smallest semantic units is less than a first preset threshold. If the semantic similarity between two adjacent smallest semantic units is less than the first set threshold, then it is determined that there is a semantic boundary between the two adjacent smallest semantic units. If the semantic similarity between two adjacent smallest semantic units is greater than or equal to the first set threshold, then it is determined that there is no semantic boundary between the two adjacent smallest semantic units.

[0008] Optionally, the semantic slicing of the industry standard clauses to obtain multiple minimal constraint units includes: For each clause in the industry standard terms, check whether the clause meets the preset splitting rules; If the clause satisfies the preset splitting rules, then the clause is split into multiple minimum constraint units according to the preset splitting rules; If the clause does not meet the preset splitting rules, then the clause is determined as a minimum constraint unit.

[0009] Optionally, the preset splitting rules include: For each clause in the industry standard clause, if the clause contains multiple independently determinate constraints, the clause is divided into multiple minimum constraint units; each minimum constraint unit corresponds to one independently determinate constraint. Alternatively, for each clause in the industry standard clauses, if the clause contains at least two of the mandatory constraint keywords, advisory requirement keywords, and special instruction keywords, then the clause shall be divided into different minimum constraint units according to at least two of the mandatory constraint keywords, advisory requirement keywords, and special instruction keywords.

[0010] Optionally, determining the minimum constraint unit matching the text segment from the plurality of minimum constraint units for each text segment includes: For each text segment, it is detected whether the text segment hits a rule anchor point; the rule anchor point is a predefined semantic identifier, and each rule anchor point is associated with at least one minimum constraint unit; If the text fragment hits at least one rule anchor, then the smallest constraint unit associated with the at least one rule anchor is determined as the smallest constraint unit that matches the text fragment. If the text segment does not match the rule anchor point, the semantic similarity between the text segment and each minimum constraint unit is calculated, and the minimum constraint unit with a semantic similarity greater than the second set threshold is determined as the minimum constraint unit that matches the text segment.

[0011] Optionally, the large language model is a pre-trained Transformer model; the review results of each text segment are all structured data, including: the standard constraint value in the minimum constraint unit, the scheme parameter value in the text segment, and the judgment result; The process of inputting the review results of each text segment into the large language model to obtain the comprehensive review results of the new energy access scheme includes: Each structured data point is input into a pre-trained Transformer model to obtain the review comments and conclusions output by the Transformer model. The Transformer model is used for: Check whether all the judgment results in the review result records show "pass"; For review result records that do not show as passed, infer whether there are corresponding modification suggestions for the review result records that do not show as passed; If there are corresponding modification suggestions for the review result records that are not displayed as passed, the modification suggestions and the reasoning of each review result record will be converted into natural language and output as review opinions, and the review conclusion of the new energy access scheme will be determined as "conditionally passed". If there is no corresponding modification suggestion for the review result record that is not displayed as passed, then the reasoning of each review result record is converted into natural language and output as review opinion, and the review conclusion of the new energy access scheme is determined to be "not passed". If all the judgment results in each review result record show that they are passed, then the reasoning of each review result record will be converted into natural language and output as review comments, and the review conclusion of the new energy access scheme will be determined as "passed".

[0012] Optionally, determining whether there are corresponding modification suggestions for the record of the review result that was not displayed as passed includes: Determine the constraint identifier of the corresponding minimum constraint unit; the constraint identifier includes: mandatory constraint identifier and advisory requirement identifier; If the constraint identifier of the corresponding minimum constraint unit is a mandatory constraint identifier, then it is determined that there is no corresponding modification suggestion for the review result record that has not been displayed as passed; If the constraint identifier of the corresponding minimum constraint unit is a suggested requirement identifier, then the modification suggestion is determined through chain reasoning based on the standard constraint value in the corresponding minimum constraint unit and the scheme parameter value in the text fragment.

[0013] Secondly, embodiments of the present invention provide a review device for new energy access schemes, comprising: The slicing module is used to obtain the scheme text and industry standard clauses of the new energy access scheme, perform semantic slicing on the scheme text to obtain multiple text fragments, and perform semantic slicing on the industry standard clauses to obtain multiple minimum constraint units. The matching module is used to determine the minimum constraint unit that matches the text segment from the plurality of minimum constraint units for each text segment, and to determine the review result record of the text segment based on the minimum constraint unit that matches the text segment. The review module is used to record the review results of each text segment and input them into the large language model to obtain the comprehensive review result of the new energy access scheme; the large language model is used to perform inductive reasoning on the review results of each text segment and output the comprehensive review result.

[0014] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0015] Compared to existing technologies, this invention performs semantic slicing on the text of the new energy access scheme and the clauses of industry standards to obtain text fragments and minimum constraint units. It then matches and reviews these text fragments and minimum constraint units, achieving precise matching and judgment between them. Furthermore, it utilizes a large language model to record, summarize, and reason about each review result, determining the comprehensive review result of the new energy access scheme text. Compared to the traditional review model of comparing each clause individually, this significantly improves review efficiency and avoids subjective judgment biases or inconsistent standard implementation standards caused by differences in experience and understanding among reviewers, thus ensuring the objectivity and standardization of the review conclusions. Attached Figure Description

[0016] Figure 1 This is an application scenario diagram of the review method for new energy access schemes provided in the embodiments of the present invention; Figure 2 This is a flowchart illustrating the implementation of semantic slicing of the solution text to obtain text fragments, as provided in an embodiment of the present invention. Figure 3 This is a system architecture diagram of the review method for new energy access schemes provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of the structure of the review device for the new energy access scheme provided in the embodiment of the present invention; Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] When reviewing renewable energy grid connection schemes, reviewers typically need to carefully study the scheme reports and then compare them one by one with technical standards, design guidelines, and policy documents to determine their compliance. This approach is not only time-consuming and labor-intensive, with a long review cycle, but it is also easily influenced by the reviewers' subjective experience.

[0019] To improve review efficiency, enhance the consistency of review results, and avoid the influence of subjective experience, this invention performs semantic slicing on the text of the new energy access scheme and the clauses of industry standards to obtain text fragments and minimum constraint units. These text fragments and minimum constraint units are then matched and reviewed, enabling precise matching and judgment between them. A large language model is then used to summarize and reason about each review result to determine the final comprehensive review result. Compared to the traditional review model of comparing each clause individually, this significantly improves review efficiency and avoids the influence of subjective experience, ensuring the objectivity and standardization of the review conclusions.

[0020] See Figure 1 The document illustrates a flowchart of the review method for the new energy access scheme provided in this embodiment of the invention, which is described in detail below: Step 101: Obtain the scheme text and industry standard clauses of the new energy access scheme, perform semantic slicing on the scheme text to obtain multiple text fragments, and perform semantic slicing on the industry standard clauses to obtain multiple minimum constraint units.

[0021] Considering the length of the renewable energy access scheme text, and that its scheme parameters and simulation results are all contained in the simulation and verification results section of the scheme text, this embodiment of the invention can employ a semantically driven hierarchical slicing method to semantically slice the simulation and verification results section of the scheme text, obtaining multiple text fragments. This embodiment of the invention can use structured extraction to extract the core indicator objects and their corresponding scheme parameter values ​​contained in each text fragment, and represent them in structured JSON format for subsequent review.

[0022] Here, the core indicators mainly include: installed capacity, grid connection voltage level, access point location, simulation results of power flow, short circuit, and stability, protection configuration, economic indicators such as investment scale, and policy indicators such as policy compliance.

[0023] In this embodiment of the invention, the above-mentioned multiple text fragments can be divided into the following five categories: 1. Basic information segment; 2. Electrical simulation results segment; 3. Economic indicators segment; 4. Policy compliance segment; 5. Conclusion segment.

[0024] The basic information segment mainly includes descriptive statements such as station name, voltage level, grid connection point location, and installed capacity. The electrical simulation results segment mainly includes Optical Character Recognition (OCR) content corresponding to text or images such as voltage value tables for each node, power flow distribution tables, and short-circuit currents, as well as explanatory text such as equipment N-1 verification results, protection configuration, and protection setting settings. The economic indicators segment mainly includes keywords such as investment scale, line cost, and construction scale. The policy compliance segment mainly includes keywords such as access area restrictions, absorption constraints, and local requirements. The conclusion segment mainly includes keywords such as final evaluation or recommendations.

[0025] For each text fragment, embodiments of the present invention can classify the text fragment into any of the five categories mentioned above by identifying whether the text fragment contains corresponding keywords, descriptive statements, or explanatory text. The classified text fragments can be output in structured JSON format for subsequent matching with the smallest constraint unit.

[0026] Here, industry standard clauses can include clauses from technical guidelines, standards, and specifications. Each clause can be segmented at the clause level into at least one comparable minimum constraint unit. This minimum constraint unit can include: the minimum constraint unit number, constraint type (which can also be understood as the constraint object), the corresponding original clause text, core judgment parameter values, etc., and output in a structured JSON format. For example, the minimum constraint unit "10kV bus voltage deviation should be within ±7%" obtained from the industry standard clause segmentation can be represented as: { "clause_id": "DL-XXX-3.2.1", "clause_type": "voltage_limit", "clause_text": "The voltage deviation of the 10kV busbar should be within ±7%", "parameters": { "upper": 1.07, "lower": 0.93 } } In the example above, the smallest constraint unit number is DL-XXX-3.2.1, which can be directly determined from the clause number in the original clause. The constraint identifier type (i.e., the constraint object) is voltage limit (voltage_limit), and the corresponding original clause text is "10kV bus voltage deviation should be within ±7%", with the core judgment parameter values ​​being the upper limit of 1.07 and the lower limit of 0.93 (upper: 1.07, lower: 0.93).

[0027] In this embodiment of the invention, the smallest constraint unit can also be classified into constraint clauses, suggestion clauses, or special clauses by detecting whether there are mandatory constraint keywords, suggestion requirement keywords, or special explanation keywords in each smallest constraint unit.

[0028] Here, keywords for mandatory constraints can include mandatory words such as "should," "must," "must not," and "prohibited." Keywords for advisory requirements can include advisory words such as "appropriate," "inappropriate," "recommended," and "can be considered." Keywords for special instructions can include expressions such as "generally...but..." and "if...then..."

[0029] For example, "The short-circuit current shall not exceed XX" can be classified as a restrictive clause. "The distance between access points should not exceed X km" can be classified as a suggestive clause. "It may be implemented in accordance with the actual conditions in various regions" can be classified as a special clause.

[0030] In this embodiment of the invention, each minimum constraint unit can be mapped to three preset types of review constraints based on its constraint type (i.e., constraint object). The three preset types of review constraints include: Technical constraints: These are constraints whose types (i.e., the objects of constraints) are technical indicators, such as whether there are situations like voltage exceeding limits, power flow exceeding limits, or short-circuit current exceeding limits; whether the N-1 requirement is met; and whether the protection configuration is complete and reasonable. Economic constraints: These are constraints whose type (or object) is an economic indicator, such as whether the investment scale exceeds the upper limit. Policy constraints: The type of constraint (i.e. the object of constraint) is a policy indicator, such as whether it meets the requirements of the access area or does not exceed the red line of the absorption capacity.

[0031] Step 102: For each text segment, determine the minimum constraint unit that matches the text segment from multiple minimum constraint units, and determine the review result record for the text segment based on the minimum constraint unit that matches the text segment.

[0032] Understandably, text fragments mostly contain semantic identifiers to indicate the target objects, such as "short-circuit current," "circuit breaker rated breaking capacity," and "short-circuit capacity." Correspondingly, each minimum constraint unit also contains the corresponding constraint type (i.e., constraint object). If the target object in a text fragment is the same as the constraint object in a minimum constraint unit, then the text fragment is determined to match that minimum constraint unit. It is understood that the same text fragment can contain one or more target objects, and correspondingly, the same text fragment can match one or more minimum constraint units.

[0033] In this embodiment of the invention, the semantic identifiers used to indicate the index object / constraint object are determined as rule anchors, and the rule anchors are used to match text fragments and minimum constraint units.

[0034] After determining the minimum constraint unit that matches each text segment, embodiments of the present invention can use a structured comparison method to review the text segments using the corresponding minimum constraint unit to detect whether they meet the corresponding minimum constraint unit and generate a review result record. The review result record may include: the number of the minimum constraint unit, the constraint type of the minimum constraint unit, the standard constraint value in the minimum constraint unit, the scheme parameter value in the text segment, and the judgment result.

[0035] The review result record in this embodiment of the invention can also be output in structured JSON format. For example, the review result record for short-circuit current can be represented as follows: { "constraint_id": "short_circuit_limit", "matched_clause": "DL-XXX-4.3.2", "scheme_value": 30, "standard_value": 20, "result": "not_satisfied" } The constraint type of the minimum constraint unit is short-circuit current limit. The minimum constraint unit number (i.e., clause number) is "DL-XXX-4.3.2". The standard constraint value in the minimum constraint unit is 20. The scheme parameter value in the text fragment is 30. The judgment result is "not satisfied".

[0036] Considering that some regions may adopt non-standard solutions that do not conform to industry technical standards due to their regional characteristics, this embodiment of the invention, after determining that a text segment does not meet the corresponding minimum constraint unit, can retrieve the regional case library and check whether there are precedents in the regional case library. If there are no precedents, the final judgment result is determined to be "not satisfied". If there are precedents, the final judgment result is determined to be "request manual approval".

[0037] Taking the short-circuit current review result record as an example, after determining that the short-circuit current exceeds the standard, it is possible to retrieve whether there are cases in the regional case library where the short-circuit current exceeds the standard but still continues to operate safely. If such a case exists, the final judgment result will be determined as "request for manual approval and confirmation", and an exception prompt will be generated: there are historical operable cases in this region that need to be reviewed and confirmed by experts.

[0038] Step 103: Input the review results of each text segment into the large language model to obtain the comprehensive review results of the new energy access scheme. Here, the large language model is used to summarize and reason about the review results of each text segment and output the comprehensive review results.

[0039] Here, the large language model is a pre-trained Transformer model. This pre-trained Transformer model is used to integrate the various review result records, summarize the reasons, and generate suggestions.

[0040] In this embodiment of the invention, different review result records and their corresponding review opinions and conclusions, as well as relevant industry knowledge graphs, can be used in advance to train the Transformer model and obtain a trained Transformer model.

[0041] Based on the above, the review results of each text fragment are all structured data, including: the constraint type of the minimum constraint unit, the standard constraint value in the minimum constraint unit, the scheme parameter values ​​in the text fragment, and the judgment result; In this embodiment of the invention, when inputting the review results of each text fragment into a pre-trained Transformer model, descriptive information from the solution report can also be input into the Transformer model. Here, descriptive information refers to: the operational description, engineering assumptions, indicator calculation methods, risk warnings, and existing judgmental statements from the simulation system in the new energy access solution report. This descriptive information is used to limit the applicable premises of the review conclusions and support reasoning and interpretation, but it does not directly participate in numerical comparison or judgment logic.

[0042] For each review result record corresponding to the technical constraints, the Transformer model performs semantic reasoning on the overall technical feasibility of the solution based on the judgment results and related descriptive information. For example, it determines whether there are problems such as voltage exceeding limits or voltage qualification rate not meeting the standards; whether there are power flow exceeding limits or equipment overload risks caused by power flow reversal; whether the short circuit current meets the equipment rated capacity requirements; and whether the safety constraints are met based on the N-1 verification results and protection configuration description.

[0043] It should be noted here that the Transformer model only performs comprehensive inductive reasoning on the input review result records and does not change any review result records.

[0044] For the review results records corresponding to economic constraints, the Transformer model can perform semantic-level analysis and summary of the economic rationality of the plan based on investment estimation indicators and corresponding judgment results. For example, whether the investment scale is reasonable, and whether there are significant cost risks exceeding expectations.

[0045] For each review record corresponding to policy constraints, the Transformer model can semantically summarize the policy compliance of the solution based on the judgment results of policy constraints such as access area restrictions and absorption capacity requirements, and determine whether the solution meets the current policy requirements or needs to be implemented under specific conditions.

[0046] Specifically, the inference logic of the Transformer model can be briefly described as follows: Check whether all the judgment results in the review result records show "pass"; For review result records that do not show as passed, determine whether there are corresponding modification suggestions for the review result records that do not show as passed based on the constraint identifier type of the corresponding minimum constraint unit; If there are corresponding modification suggestions for the review results that are not displayed, the modification suggestions and each review result record will be converted into natural language and output as review comments, and the review conclusion of the new energy access scheme will be determined as "conditionally passed". If no corresponding modification suggestions are found for the review results that are not displayed as passed, then each review result record will be converted into natural language and output as review comments, and the review conclusion for the new energy access scheme will be determined as "not passed". If all the judgment results in each review result record show "pass", then each review result record will be converted into natural language and output as review comments, and the review conclusion of the new energy access scheme will be determined as "pass".

[0047] Here, if the judgment result is "satisfied", then the judgment result is displayed as passed. If the judgment result is "not satisfied", then the judgment result is displayed as failed.

[0048] If all review result records show a "pass" result, the Transformer model can use descriptive information to transform the structured JSON-formatted review result records into natural language, and output the transformed natural language as the review opinion. The review conclusion is "pass".

[0049] In some embodiments, for review result records that do not show as passed, the Transformer model in this embodiment can determine the corresponding modification suggestions based on their constraint identifiers and in conjunction with pre-learned industry knowledge. Specifically, the constraint identifier of the smallest constraint unit corresponding to the review result record that does not show as passed can be determined first. Here, the constraint identifier can include: mandatory constraints and suggested requirements.

[0050] Based on the above, embodiments of the present invention can determine the corresponding constraint identifier by detecting whether there are mandatory constraint keywords or suggested requirement keywords in each minimum constraint unit.

[0051] If the constraint identifier of the corresponding minimum constraint unit is a mandatory constraint identifier, then it is determined that there is no corresponding modification suggestion for the review result record that has not been displayed as passed; if the constraint identifier of the corresponding minimum constraint unit is a suggested requirement identifier, then the modification suggestion is determined by chain reasoning based on the standard constraint value in the corresponding minimum constraint unit and the scheme parameter value in the text fragment.

[0052] For example, a text fragment displays "10kV bus voltage deviation is -7.9%". Clearly, the review result record corresponding to this text fragment does not meet the minimum constraint unit of "10kV bus voltage deviation should be within ±7%", and the review result record shows a failure. The Transformer model can generate a chain of inference based on pre-learned industry knowledge: voltage deviation -7.9% → does not meet ±7% requirement → requires additional reactive power compensation → recommended capacitor capacity 2Mvar.

[0053] In this embodiment of the invention, modification suggestions and records of each review result can be converted into natural language and output as review comments to determine the review conclusion of the new energy access scheme as "condition passed".

[0054] In addition, the Transformer model can also perform consistency checks, that is, check whether the review results records are consistent with the solution conclusions in the solution text. If there is a discrepancy, the review conclusion is determined to be "not passed", and the reason for the failure is output accordingly.

[0055] In other embodiments, the present invention can also score the new energy access scheme from three aspects: policy, economy and technology, to quantify the comprehensive review results of the new energy access scheme.

[0056] When scoring policy-related, economic, and technical aspects separately, a tiered and quantitative assessment can be conducted based on the following rules: 1. Map the judgment results such as "satisfied", "not satisfied", and "request manual approval" to different scores respectively; 2. Different impact weights are set for mandatory constraints, advisory requirements, and special instructions; 3. Adjust the scoring results based on the number of objects involved, the coverage, and the degree of deviation from the threshold.

[0057] Through the above-described method, the embodiments of the present invention can generate domain-specific scoring results for technical, economic, and policy-related reviews.

[0058] Ultimately, embodiments of the present invention can generate corresponding scheme review reports for new energy access schemes. The scheme review report may include the following: 1. Score results for technical, economic, and policy-related reviews; 2. Review conclusions (passed, condition met, failed); 3. Records of review results that did not show as passed, their corresponding minimum constraint units, and the original text of the relevant clauses; 4. Modification suggestions and explanations of optimization schemes.

[0059] Among them, the modification suggestions refer to the implementation of feasible adjustment measures for a single constraint problem, including but not limited to adjusting the location of the access point, increasing reactive power compensation, reducing access capacity, replacing or increasing the capacity of equipment, and adjusting the implementation sequence. An optimization solution refers to the collaborative handling of multiple constraint problems, forming an overall improvement path by combining multiple modification suggestions. When there is more than one modification suggestion, the Transformer model in this embodiment can combine multiple modification suggestions based on pre-learned industry knowledge to form an optimization solution.

[0060] Furthermore, embodiments of the present invention can also write review conclusions, scheme features, and optimization suggestions into a knowledge graph and case library to form a "scheme-review conclusion-optimization path" triple, which can be used for automatic reference in subsequent similar projects and fine-tuning training of large models to achieve self-learning and experience accumulation.

[0061] Compared to existing technologies, this invention performs semantic slicing on the text of the new energy access scheme and the clauses of industry standards to obtain text fragments and minimum constraint units. It then matches and reviews these text fragments and minimum constraint units, achieving precise matching and judgment between them. Furthermore, it utilizes a large language model to record, summarize, and reason about each review result, determining the comprehensive review result of the new energy access scheme text. Compared to the traditional review model of comparing each clause individually, this significantly improves review efficiency and avoids subjective judgment biases or inconsistent standard implementation standards caused by differences in experience and understanding among reviewers, thus ensuring the objectivity and standardization of the review conclusions.

[0062] The following is combined with Figure 2 This section elaborates on the method of obtaining text fragments through semantic slicing of the solution text. See details below. Figure 2 : Step 201: Based on the punctuation marks in the scheme text, the scheme text is sequentially divided into multiple smallest semantic units.

[0063] Here, punctuation marks can include paragraph marks, line breaks, and periods. This embodiment of the invention can sequentially divide the scheme text into multiple smallest semantic units according to the paragraph marks, line breaks, and periods in the scheme text. Furthermore, the position index and source attribute of each smallest semantic unit in the original text can be recorded. The source attribute here can include, but is not limited to, body text, title text, table description text, and text content obtained by OCR conversion of images.

[0064] Step 202: For each pair of adjacent smallest semantic units, determine whether there is a semantic boundary between each pair of adjacent smallest semantic units according to the preset structured tags, preset boundary keywords and the semantic similarity between each pair of adjacent smallest semantic units.

[0065] Based on the determination of the smallest semantic unit, embodiments of the present invention can determine whether a semantic boundary is formed between any two adjacent smallest semantic units based on preset structured tags, boundary keywords, and semantic similarity.

[0066] Here, the priority of structured markers, boundary keywords, and semantic similarity decreases sequentially. For each pair of adjacent smallest semantic units, this embodiment of the invention can determine whether a semantic boundary exists between them based on structured markers, boundary keywords, and semantic similarity in sequence.

[0067] Specifically, for each pair of adjacent smallest semantic units, if there is a preset structured marker between the two adjacent smallest semantic units, then it is determined that there is a semantic boundary between the two adjacent smallest semantic units. If there is no preset structured marker between two adjacent smallest semantic units, then check whether there is a preset boundary keyword in the two adjacent smallest semantic units. If there are predefined boundary keywords in two adjacent smallest semantic units, then it is determined that there is a semantic boundary between the two adjacent smallest semantic units. If there are no preset boundary keywords in two adjacent smallest semantic units, then check whether the semantic similarity between the two adjacent smallest semantic units is less than the first preset threshold. If the semantic similarity between two adjacent smallest semantic units is less than a first set threshold, then it is determined that there is a semantic boundary between the two adjacent smallest semantic units. If the semantic similarity between two adjacent smallest semantic units is greater than or equal to a first set threshold, then it is determined that there is no semantic boundary between the two adjacent smallest semantic units.

[0068] Here, the pre-defined structured markers can include chapter number variations, header rows, table start or end markers, etc. For each pair of adjacent smallest semantic units, when the presence of the aforementioned structured markers is detected, a semantic boundary can be determined between them.

[0069] Predefined boundary keywords can include keywords or phrases used to identify analysis results, calculation processes, or conclusions, such as "after calculation," "verification results," "as seen in Table ××," and "in summary." For each pair of adjacent smallest semantic units, a semantic boundary can be determined to exist between them when the aforementioned boundary keywords are detected.

[0070] For each pair of adjacent smallest semantic units, if there are neither structured markers nor boundary keywords between them, their semantic similarity is calculated. In calculating semantic similarity, both units can be semantically encoded separately to obtain corresponding semantic vectors, and the semantic similarity between these vectors is then calculated. If the semantic similarity is less than a first predetermined threshold, a semantic boundary is determined to exist between them; if the semantic similarity is greater than or equal to the first predetermined threshold, no semantic boundary is determined to exist between them.

[0071] Step 203: Divide all the smallest semantic units in the scheme text according to semantic boundaries to form multiple text segments.

[0072] Based on the determination of semantic boundaries, embodiments of the present invention can divide all the smallest semantic units into multiple text segments according to the semantic boundaries.

[0073] This invention uses punctuation marks to break down text into the smallest semantic units, providing a structured foundation for subsequent slicing and reducing processing complexity. Furthermore, by employing a three-level determination mechanism of structured markers > boundary keywords > semantic similarity, it balances explicit document formatting rules with the semantic coherence of text content, ensuring that the slicing results conform to industry document standards while accurately reflecting the logical division of the content itself.

[0074] In some embodiments, when semantically slicing industry standard clauses to obtain multiple minimum constraint units, for each clause in the industry standard clauses, it can be detected whether the clause meets a preset splitting rule; if the clause meets the preset splitting rule, the clause is split into multiple minimum constraint units according to the preset splitting rule; if the clause does not meet the preset splitting rule, the clause is determined as a single minimum constraint unit.

[0075] Before semantically slicing industry standard clauses, this invention can first perform semantic deconstruction on each clause to identify its core semantic elements. Here, the core semantic elements may include: Binding action elements: used to characterize the binding strength and nature of the clause, including but not limited to "shall", "shall not", "shall not", "may", etc.; Constraint object elements: used to characterize the constrained object, including indicators such as voltage, current, capacity, distance, and short-circuit current; Decision parameter elements: used to characterize the numerical values, ranges, or enumerations in the constraints, including upper limits, lower limits, intervals, or discrete values; Applicable conditions elements: These are used to characterize the prerequisites for the application of the terms, including voltage level, operating mode, access scenario, or regional conditions.

[0076] For each clause, embodiments of the present invention can split it according to a preset splitting rule based on the core semantic elements in the clause, thereby obtaining at least one minimum constraint unit.

[0077] In some embodiments, the preset splitting rules include: For each clause in the industry standard, if the clause contains multiple independently determinate constraints, the clause is broken down into multiple minimum constraint units. Here, each minimum constraint unit corresponds to one independently determinate constraint. Among them, multiple independently determinate constraints refer to: multiple clauses connected by coordinating conjunctions (e.g., "and", "as well as", "at the same time", "and", etc.), and each clause imposes constraints on different constraint objects; or, multiple enumerated or list-type requirements imposed on the same constraint object.

[0078] For example, the clause “the voltage deviation at the connection point shall meet the requirements and the short-circuit current shall not exceed the breaking capacity of the circuit breaker” can be split into two minimum constraint units, “the voltage deviation at the connection point shall meet the requirements” and “the short-circuit current shall not exceed the breaking capacity of the circuit breaker”, based on the coordinating conjunction “and”.

[0079] "Photovoltaic power plants should have the following protection functions: (1) instantaneous overcurrent protection; (2) overvoltage protection; (3) islanding protection." This clause is an enumerated type of multiple requirements for the same constraint object, which can be divided into three minimum constraint units.

[0080] Alternatively, for each clause in an industry standard clause, if the clause contains at least two of the mandatory constraint keywords, advisory requirement keywords, and special instruction keywords, then the clause shall be divided into different minimum constraint units according to at least two of the mandatory constraint keywords, advisory requirement keywords, and special instruction keywords.

[0081] Here, mandatory constraint keywords can be words like "should," "must," "must not," or "prohibited" from the constraint action elements. Suggestive requirement keywords can be words like "appropriate," "inappropriate," "recommended," or "can be considered" from the constraint action elements. Special explanation keywords can be words like "may," "permitted," "generally...but..." or "if...then..." from the constraint action elements.

[0082] For example, the clause “The grid connection point voltage should be within ±7% of the rated voltage, and preferably within ±5%” contains both the mandatory constraint keyword “should” and the advisory requirement keyword “preferably”. Therefore, the clause can be split into two minimum constraint units according to these two keywords: “The grid connection point voltage should be within ±7% of the rated voltage” and “The grid connection point voltage should preferably be controlled within ±5%”.

[0083] This invention, through preset splitting rules, deconstructs complex and lengthy original standard clauses into multiple single, clear minimum constraint units. Each minimum constraint unit contains only one core judgment object and condition, which provides a precise and unambiguous operation object for subsequent automated comparison.

[0084] This invention can automatically identify and process complex clauses containing multiple meanings or mixed requirements based on preset splitting rules, avoiding the need for manual preprocessing. By splitting clauses into the smallest comparable units, the system can achieve more refined and accurate compliance matching and judgment, significantly improving the coverage and reliability of automated review.

[0085] In embodiments of the present invention, after determining the text segment and the minimum constraint unit, the two are further matched. In some embodiments, when determining the minimum constraint unit matched by each text segment, the following can be performed: determining the minimum constraint unit that matches the text segment from a plurality of minimum constraint units, including: For each text segment, it is checked whether the text segment matches a rule anchor. If the text segment matches at least one rule anchor, the smallest constraint unit associated with that rule anchor is determined as the smallest constraint unit that matches the text segment; if the text segment does not match a rule anchor, the semantic similarity between the text segment and each smallest constraint unit is calculated, and the smallest constraint unit with a semantic similarity greater than a second set threshold is determined as the smallest constraint unit that matches the text segment.

[0086] Here, rule anchors are predefined semantic identifiers. Essentially, rule anchors are semantic identifiers used to indicate indicator / constraint objects. Examples include keywords such as "short-circuit current," "circuit breaker rated breaking capacity," and "short-circuit capacity." It can be understood that each rule anchor is associated with at least one minimum constraint unit.

[0087] In this embodiment of the invention, the smallest constraint unit corresponding to a rule anchor point matched by a text fragment can be determined as the smallest constraint unit for matching the text fragment. If the text fragment does not match any rule anchor points, semantic encoding can be performed on the text fragment and each smallest constraint unit to obtain the corresponding semantic vector, and the semantic similarity between the text fragment and each smallest constraint unit can be calculated. The smallest constraint unit with a semantic similarity greater than a second set threshold is determined as the smallest constraint unit for matching the text fragment.

[0088] Here, the value of the second set threshold can be determined according to the actual situation. For example, the second set threshold can be 0.75.

[0089] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0090] The following is combined with Figure 3 The overall process of the review method for the new energy access scheme provided in the embodiments of the present invention is introduced.

[0091] Business personnel upload the aforementioned new energy access solution reports to the automatic review system for new energy access via user terminals such as workstations and web clients. Furthermore, business personnel can also view the review results through these user terminals. The review system includes an access layer (also known as an access module), which provides a file upload interface and API / Web services for receiving files. Additionally, the review system can include a report display and interactive interface module to display review reports and provide warning information or exception notifications.

[0092] The review system has a built-in semantic parsing and text slicing module, which is used to perform semantic boundary recognition and scheme element extraction operations on new energy access schemes based on industry dictionaries, templates, etc., to achieve semantic-level text slicing, obtain text fragments, and output them in structured JSON format.

[0093] The review system also has a built-in indicator mapping and constraint comparison module, which maps the scheme elements in the text fragment to different minimum constraint units such as technology, economy, and policy, so as to achieve the matching between the text fragment and the minimum constraint unit and obtain the corresponding judgment result.

[0094] The review system also includes a built-in large-scale model reasoning and review decision-making module. This module uses a large language model to perform multi-dimensional reasoning, including technical, economic, and policy considerations, and outputs review opinions and conclusions. The large language model can also perform consistency checks to verify whether the recorded review results are consistent with the final conclusions of the access solution.

[0095] The review system also has a built-in comprehensive scoring and report generation module, which is used to calculate the weight and score of each review result record to generate the final domain score and output the final review report.

[0096] The review system also includes built-in modules for internal and external identification and expert collaboration, which are used to retrieve historical cases in the region to generate exception tags that "require manual review".

[0097] The review system also has a built-in knowledge writing and case accumulation module, which is used to write review conclusions, solution features and optimization suggestions into the knowledge graph and case library to form a "solution-review conclusion-optimization path" triple, so as to be used for automatic reference in subsequent similar projects and fine-tuning training of large models, so as to achieve self-learning and experience accumulation.

[0098] Furthermore, embodiments of this invention also include an industry standard clause library, a knowledge graph and case library, and expert terminals or expert workstations. The industry standard clause library mainly includes a knowledge base of technical guidelines / standards / specifications, etc., and by segmenting clauses, multiple minimal constraint units can be obtained for matching and judging individual text fragments. The expert terminals or expert workstations are used for manual review, viewing solutions and review results, and submitting manual review results.

[0099] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0100] Figure 4 A schematic diagram of the structure of the review device for the new energy access scheme provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 4 As shown, the review device 4 for the new energy access scheme includes: a slicing module 41, a matching module 42, and a review module 43.

[0101] The slicing module 41 is used to obtain the scheme text and industry standard clauses of the new energy access scheme, perform semantic slicing on the scheme text to obtain multiple text fragments, and perform semantic slicing on the industry standard clauses to obtain multiple minimum constraint units. The matching module 42 is used to determine the minimum constraint unit that matches the text segment from multiple minimum constraint units for each text segment, and to determine the review result record of the text segment based on the minimum constraint unit that matches the text segment. The review module 43 is used to record the review results of each text segment and input them into the large language model to obtain the comprehensive review results of the new energy access scheme; the large language model is used to perform inductive reasoning on the review results of each text segment and output the comprehensive review results.

[0102] In one possible implementation, the slice module 41 is specifically used for: Based on the punctuation marks in the scheme text, the scheme text is sequentially divided into multiple smallest semantic units; For each pair of adjacent smallest semantic units, the existence of a semantic boundary between each pair of adjacent smallest semantic units is determined sequentially according to the preset structured tags, preset boundary keywords, and semantic similarity between each pair of adjacent smallest semantic units. Based on semantic boundaries, all the smallest semantic units in the scheme text are divided into multiple text segments.

[0103] In one possible implementation, the slice module 41 is specifically used for: For every two adjacent smallest semantic units, if there is a preset structured marker between the two adjacent smallest semantic units, then it is determined that there is a semantic boundary between the two adjacent smallest semantic units. If there is no preset structured marker between two adjacent smallest semantic units, then check whether there is a preset boundary keyword in the two adjacent smallest semantic units. If there are predefined boundary keywords in two adjacent smallest semantic units, then it is determined that there is a semantic boundary between the two adjacent smallest semantic units. If there are no preset boundary keywords in two adjacent smallest semantic units, then check whether the semantic similarity between the two adjacent smallest semantic units is less than the first preset threshold. If the semantic similarity between two adjacent smallest semantic units is less than a first set threshold, then it is determined that there is a semantic boundary between the two adjacent smallest semantic units. If the semantic similarity between two adjacent smallest semantic units is greater than or equal to a first set threshold, then it is determined that there is no semantic boundary between the two adjacent smallest semantic units.

[0104] In one possible implementation, the slice module 41 is specifically used for: For each clause in the industry standard terms, check whether the clause meets the preset splitting rules; If the clause meets the preset splitting rules, then the clause will be split into multiple minimum constraint units according to the preset splitting rules; If the clause does not meet the preset splitting rules, then the clause will be determined as a minimum constraint unit.

[0105] In one possible implementation, the preset splitting rules include: For each clause in the industry standard clause, if the clause contains multiple independently determinate constraints, the clause is broken down into multiple minimum constraint units; each minimum constraint unit corresponds to one independently determinate constraint. Alternatively, for each clause in an industry standard clause, if the clause contains at least two of the mandatory constraint keywords, advisory requirement keywords, and special instruction keywords, then the clause shall be divided into different minimum constraint units according to at least two of the mandatory constraint keywords, advisory requirement keywords, and special instruction keywords.

[0106] In one possible implementation, the matching module 42 is specifically used for: For each text segment, check whether the text segment hits a rule anchor; the rule anchor is a predefined semantic identifier, and each rule anchor is associated with at least one minimum constraint unit; If a text fragment matches at least one rule anchor, then the smallest constraint unit associated with that at least one rule anchor is determined as the smallest constraint unit that matches the text fragment. If a text segment does not match a rule anchor point, the semantic similarity between the text segment and each minimum constraint unit is calculated, and the minimum constraint unit with a semantic similarity greater than the second set threshold is determined as the minimum constraint unit that matches the text segment.

[0107] In one possible implementation, the large language model is a pre-trained Transformer model; the evaluation results of each text segment are recorded as structured data, including: the standard constraint values ​​in the minimum constraint unit, the scheme parameter values ​​in the text segment, and the judgment results; Review module 43 is specifically used for: Each structured data point is input into a pre-trained Transformer model to obtain the review comments and conclusions output by the Transformer model. The Transformer model is used for: Check whether all the judgment results in the review result records show "pass"; For review result records that do not show as passed, infer whether there are corresponding modification suggestions for those records. If there are corresponding modification suggestions for the review results that are not displayed, the modification suggestions and the reasoning of each review result record will be converted into natural language and output as review comments, and the review conclusion of the new energy access scheme will be determined as "conditionally passed". If no corresponding modification suggestions are found for the review result record that has not been approved, the reasoning of each review result record will be converted into natural language and output as review comments, and the review conclusion of the new energy access scheme will be determined as "not approved". If all the judgment results in each review result record show that they are passed, then the reasoning of each review result record will be converted into natural language and output as review comments, and the review conclusion of the new energy access scheme will be determined as "passed".

[0108] In one possible implementation, review module 43 is specifically used for: Determine the constraint identifier of the corresponding smallest constraint unit; constraint identifiers include: mandatory constraint identifiers and recommended requirement identifiers; If the constraint identifier of the corresponding minimum constraint unit is a mandatory constraint identifier, then it is determined that there is no corresponding modification suggestion for the review result record that has not been displayed as passed; If the constraint identifier of the corresponding minimum constraint unit is a suggested requirement identifier, then the modification suggestion is determined through chain reasoning based on the standard constraint value in the corresponding minimum constraint unit and the scheme parameter value in the text fragment.

[0109] This device embodiment can be used to implement the above method embodiment, and its technical principle and implementation effect are the same as those of the above method embodiment, so they will not be repeated here.

[0110] Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Figure 5 As shown, the electronic device 5 of this embodiment includes a processor 50 and a memory 51. The memory 51 stores a computer program 52. When the processor 50 executes the computer program 52, it implements the steps in the various method embodiments described above. Alternatively, when the processor 50 executes the computer program 52, it implements the functions of each module / unit in the various device embodiments described above.

[0111] For example, computer program 52 may be divided into one or more modules / units, which are stored in memory 51 and executed by processor 50 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 52 in electronic device 5.

[0112] Electronic device 5 may include, but is not limited to, processor 50 and memory 51. Those skilled in the art will understand that... Figure 5 This is merely an example of electronic device 5 and does not constitute a limitation on electronic device 5. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 5 may also include input / output devices, network access devices, buses, etc.

[0113] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0114] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0115] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for reviewing new energy access schemes, characterized in that, include: Obtain the scheme text and industry standard clauses of the new energy access scheme, perform semantic segmentation on the scheme text to obtain multiple text fragments, and perform semantic segmentation on the industry standard clauses to obtain multiple minimum constraint units; For each text segment, determine the minimum constraint unit that matches the text segment from the plurality of minimum constraint units, and determine the review result record for the text segment based on the minimum constraint unit that matches the text segment; The review results of each text segment are recorded and input into the large language model to obtain the comprehensive review result of the new energy access scheme; the large language model is used to perform inductive reasoning on the review results of each text segment and output the comprehensive review result.

2. The review method for new energy access schemes according to claim 1, characterized in that, The semantic slicing of the proposed text yields multiple text fragments, including: Based on the punctuation marks in the proposed solution text, the proposed solution text is sequentially divided into multiple smallest semantic units; For each pair of adjacent smallest semantic units, the existence of a semantic boundary between each pair of adjacent smallest semantic units is determined sequentially according to the preset structured tags, preset boundary keywords, and the semantic similarity between each pair of adjacent smallest semantic units. According to the semantic boundaries, all the smallest semantic units in the scheme text are divided into multiple text segments.

3. The method for reviewing new energy access schemes according to claim 2, characterized in that, For each pair of adjacent smallest semantic units, the existence of a semantic boundary between them is determined sequentially based on preset structured tags, preset boundary keywords, and the semantic similarity between the two adjacent smallest semantic units. This includes: For every two adjacent smallest semantic units, if there is a preset structured marker between the two adjacent smallest semantic units, then it is determined that there is a semantic boundary between the two adjacent smallest semantic units. If there is no preset structured marker between the two adjacent smallest semantic units, then detect whether there is a preset boundary keyword in the two adjacent smallest semantic units; If a preset boundary keyword exists in two adjacent smallest semantic units, then it is determined that there is a semantic boundary between the two adjacent smallest semantic units. If there is no preset boundary keyword in the two adjacent smallest semantic units, then it is detected whether the semantic similarity between the two adjacent smallest semantic units is less than a first preset threshold. If the semantic similarity between two adjacent smallest semantic units is less than the first set threshold, then it is determined that there is a semantic boundary between the two adjacent smallest semantic units. If the semantic similarity between two adjacent smallest semantic units is greater than or equal to the first set threshold, then it is determined that there is no semantic boundary between the two adjacent smallest semantic units.

4. The method for reviewing new energy access schemes according to any one of claims 1-3, characterized in that, The semantic slicing of the industry standard clauses yields multiple minimal constraint units, including: For each clause in the industry standard clauses, check whether the clause meets the preset splitting rules; If the clause satisfies the preset splitting rules, then the clause is split into multiple minimum constraint units according to the preset splitting rules; If the clause does not meet the preset splitting rules, then the clause is determined as a minimum constraint unit.

5. The method for reviewing new energy access schemes according to claim 4, characterized in that, The preset splitting rules include: For each clause in the industry standard clause, if the clause contains multiple independently determinate constraints, the clause is divided into multiple minimum constraint units; each minimum constraint unit corresponds to one independently determinate constraint. Alternatively, for each clause in the industry standard clauses, if the clause contains at least two of the mandatory constraint keywords, advisory requirement keywords, and special instruction keywords, then the clause shall be divided into different minimum constraint units according to at least two of the mandatory constraint keywords, advisory requirement keywords, and special instruction keywords.

6. The method for reviewing new energy access schemes according to any one of claims 1-3, characterized in that, For each text segment, determining the minimum constraint unit that matches the text segment from the plurality of minimum constraint units includes: For each text segment, it is detected whether the text segment hits a rule anchor point; the rule anchor point is a predefined semantic identifier, and each rule anchor point is associated with at least one minimum constraint unit; If the text fragment hits at least one rule anchor, then the smallest constraint unit associated with the at least one rule anchor is determined as the smallest constraint unit that matches the text fragment. If the text segment does not match the rule anchor point, the semantic similarity between the text segment and each minimum constraint unit is calculated, and the minimum constraint unit with a semantic similarity greater than the second set threshold is determined as the minimum constraint unit that matches the text segment.

7. The method for reviewing new energy access schemes according to any one of claims 1-3, characterized in that, The large language model is a pre-trained Transformer model; the review results of each text segment are all structured data, including: the standard constraint value in the minimum constraint unit, the scheme parameter value in the text segment, and the judgment result; The process of inputting the review results of each text segment into the large language model to obtain the comprehensive review results of the new energy access scheme includes: Each structured data point is input into a pre-trained Transformer model to obtain the review comments and conclusions output by the Transformer model. The Transformer model is used for: Check whether all the judgment results in the review result records show "pass"; For review result records that do not show as passed, infer whether there are corresponding modification suggestions for the review result records that do not show as passed; If there are corresponding modification suggestions for the review result records that are not displayed as passed, the modification suggestions and the reasoning of each review result record will be converted into natural language and output as review opinions, and the review conclusion of the new energy access scheme will be determined as "condition passed". If there is no corresponding modification suggestion for the review result record that is not displayed as passed, then the reasoning of each review result record is converted into natural language and output as review opinion, and the review conclusion of the new energy access scheme is determined to be "not passed"; If all the judgment results in each review result record show "pass", then the reasoning of each review result record is converted into natural language and output as review opinion, and the review conclusion of the new energy access scheme is determined to be "pass".

8. The method for reviewing new energy access schemes according to claim 7, characterized in that, Determining whether there are corresponding modification suggestions for the review results that were not displayed as passed includes: Determine the constraint identifier of the corresponding minimum constraint unit; the constraint identifier includes: mandatory constraint identifier and advisory requirement identifier; If the constraint identifier of the corresponding minimum constraint unit is a mandatory constraint identifier, then it is determined that there is no corresponding modification suggestion for the review result record that has not been displayed as passed; If the constraint identifier of the corresponding minimum constraint unit is a suggested requirement identifier, then the modification suggestion is determined through chain reasoning based on the standard constraint value in the corresponding minimum constraint unit and the scheme parameter value in the text fragment.

9. A device for evaluating new energy access schemes, characterized in that, include: The slicing module is used to obtain the scheme text and industry standard clauses of the new energy access scheme, perform semantic slicing on the scheme text to obtain multiple text fragments, and perform semantic slicing on the industry standard clauses to obtain multiple minimum constraint units. The matching module is used to determine the minimum constraint unit that matches the text segment from the plurality of minimum constraint units for each text segment, and to determine the review result record of the text segment based on the minimum constraint unit that matches the text segment. The review module is used to record the review results of each text segment and input them into the large language model to obtain the comprehensive review result of the new energy access scheme; the large language model is used to perform inductive reasoning on the review results of each text segment and output the comprehensive review result.

10. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 8.

Images