Decision-making target and multi-source heterogeneous data aided decision-making report generation method

Abstract

The application relates to the technical field of power system data analysis and artificial intelligence, and discloses a decision target and multi-source heterogeneous data auxiliary decision report generation method; the generation method realizes the automatic generation of the power auxiliary decision report through decision target driven information dimension analysis, multi-source text standardization processing, long text global semantic coding, key information screening, small sample driven optimization and structured rule mapping; the extraction result is highly matched with specific decision targets such as dispatching and fault disposal, and the decision adaptability is strong; the method can be migrated across scenes with few labeled samples, has strong generalization ability, and has low adaptive iteration cost; in combination with secondary error correction and semantic filtering, the method effectively resists the interference caused by oralization and non-standard format, is stable in extraction precision, retains the cross-paragraph correlation information in the long text through semantic blocking and cross-block attention fusion, and significantly improves the efficiency of power text analysis and the objectivity and timeliness of the decision report.

Description

Technical Field

[0001] This invention relates to the field of power system data analysis and artificial intelligence technology, specifically to a method for generating auxiliary decision reports based on decision objectives and multi-source heterogeneous data. Background Technology

[0002] With the deepening of the construction of new power systems, core power operations such as grid dispatching and equipment maintenance have accumulated massive amounts of unstructured text data, including regulations, fault reports, and dispatch instructions. These texts contain core information supporting scientific decision-making in power operations. As a core supporting component of intelligent auxiliary decision-making systems for power systems, the technology for extracting key text information directly determines the accuracy and efficiency of decision-making.

[0003] Currently, key text information extraction technologies used in the power industry are mainly divided into three categories: rule-based dictionary matching, traditional machine learning extraction, and fine-tuning of general pre-trained models. These methods can extract preset types of power text information in fixed scenarios, supporting basic decision support needs, but they generally suffer from the following shortcomings: First, insufficient decision adaptability; existing methods are mostly general semantic extraction, resulting in low matching degrees between the extraction results and specific decision objectives such as scheduling and fault handling. Second, weak scenario generalization ability; rules need to be redefined for cross-scenario applications, and general models require fine-tuning with a large number of labeled power samples, leading to high adaptation costs and long cycles. Third, insufficient long text processing capability; power regulations, fault analysis reports, and other documents are often long texts, and existing methods, when processed in segments, easily lose cross-segment related decision information, failing to achieve global semantic association extraction. Fourth, poor anti-interference capability; field operation and maintenance logs and fault records often contain colloquial language and non-standard formats, causing large fluctuations in the extraction accuracy of existing methods and making it difficult to guarantee the stability of decision support.

[0004] Meanwhile, traditional power decision support methods still rely heavily on manual experience analysis and report writing, which has significant drawbacks: First, they are inefficient and lack timeliness. Faced with massive amounts of data, manual reading, screening, summarizing, and report writing are time-consuming and laborious, making it impossible to quickly respond to changes in the power market and emergencies, resulting in delayed decision-making. Second, they are highly subjective and lack consistent standards. Differences in the knowledge background and understanding of different analysts make it difficult to guarantee the completeness of the report content, the depth of analysis, and the consistency of conclusions. Third, information mining is insufficient. Manual processing cannot fully capture potential correlations, trend changes, and abnormal signals in unstructured texts, easily overlooking key information. Fourth, they lack adaptability. Processes and rules need to be redesigned for different decision-making objectives, lacking universality and flexibility.

[0005] Therefore, there is an urgent need for a new method that can automatically and intelligently process multi-source heterogeneous power text data and efficiently and accurately generate auxiliary decision-making reports according to specific decision-making objectives, so as to improve the analysis efficiency and decision-making level of power systems. Summary of the Invention

[0006] To overcome the aforementioned technical problems, this invention provides a method for generating auxiliary decision-making reports based on decision objectives and multi-source heterogeneous data. This method achieves automated generation of power auxiliary decision-making reports through decision objective-driven information dimension analysis, multi-source text standardization processing, long text global semantic encoding, key information filtering, small-sample driven optimization, and structured rule mapping. The extracted results are highly matched with specific decision objectives such as scheduling and fault handling, demonstrating strong decision adaptability. It can be transferred across scenarios without a large number of labeled samples, exhibiting strong generalization ability and low adaptation iteration cost. Combined with two-level error correction and semantic filtering, it effectively resists interference from colloquialisms and non-standard formats, ensuring stable extraction accuracy. Through semantic segmentation and cross-block attention fusion, it fully preserves cross-paragraph correlation information in long texts, significantly improving the efficiency of power text analysis and the objectivity and timeliness of decision-making reports.

[0007] To achieve the above objectives, the present invention provides a method for generating auxiliary decision-making reports based on decision objectives and multi-source heterogeneous data, comprising: The decision-making objectives for acquiring power data reports and the multi-source heterogeneous power texts to be processed; Based on the decision-making objectives, analyze and obtain information dimension groups; A standardized text sequence to be processed is obtained based on the multi-source heterogeneous power text to be extracted; The text sequence to be processed is segmented and semantically encoded and fused to generate a global semantic representation matrix; Extract information related to information dimension groups from the global semantic representation matrix, filter and integrate it to form a candidate information key set; A small-sample-driven extraction optimization method is used to optimize the candidate information key set and generate the information key set. The information key set is structurally standardized, input into a pre-defined power auxiliary decision-making system, and the corresponding auxiliary decision-making report is output.

[0008] Preferably, the information dimension groups obtained based on the decision-making objectives include: Construct a labeling system for power auxiliary decision-making scenarios, including power grid dispatching, equipment operation and maintenance, fault handling, safety management and control, and power grid planning scenarios; Standardized extraction dimensions are preset for each power-assisted decision-making scenario; The ALBERT model is used to parse the text content of the decision target. Relevant extraction dimensions are obtained from all pre-defined standardized extraction dimensions to construct information dimension groups. At the same time, the task of constructing information dimension groups is used as the task instruction to generate an extraction task with task instructions, output requirements and instances.

[0009] Preferably, a standardized text sequence to be processed is obtained based on the multi-source heterogeneous power text to be extracted, including: The multi-source heterogeneous power text to be processed is cleaned and standardized preprocessed to obtain the format-normalized multi-source heterogeneous power text. A two-level text correction method is used to process the format-normalized multi-source heterogeneous power text and construct the corrected power text. The power industry-specific word segmentation tool is used to segment the corrected power industry text into sentences and words, and output a standardized sequence of text to be processed.

[0010] Preferably, a two-level text correction method is used to process the format-normalized multi-source heterogeneous power text, and the corrected power text is constructed, including: Map all candidate terms in the power field with the same semantics to the same standard terminology to construct a power error correction dictionary; Identify and locate all candidate terms in the power sector that need to be corrected; Based on the Levenshtein distance algorithm, formula (1) is used to calculate the editing distance between all candidate terms in the power field to be corrected and their corresponding standard terms. (1) in, These are candidate terms in the power sector that need correction. To correspond with the same standard terminology as the candidate terminology in the power sector that needs to be corrected, This is the minimum edit distance between the first few characters of a candidate term in the power sector to be corrected and the first few characters of the same standard term. These are the character index numbers, For boundary condition calculation terms, It is a minimum value function. The unit cost of editing operations, To replace the indicator cost function, The preceding text is a list of candidate terms for the power sector that need correction. One character, For the same standard term One character; When the edit distance is less than or equal to the first preset threshold, the corresponding standard term is output and included in the candidate error correction set to be corrected; The semantic matching degree of the error correction result is generated by formula (2) using the ALBERT model to correct the error correction result of the candidate error correction set. (2) in, The semantic matching degree function of the error correction result. For the purpose of error correction, the same standard terminology should be used. For the complete sentence containing the same standard term in electrical texts. This refers to the high-dimensional semantic vectors generated by processing the same standard term using the ALBERT model. This refers to the high-dimensional semantic vectors generated from the complete sentences containing the same standard term in electricity texts after processing by the ALBERT model. It is an L2 norm; Error correction is completed when the semantic matching degree of the error correction result is greater than or equal to the second preset threshold, and the corrected power text is constructed.

[0011] Preferably, after segmenting and semantically encoding the text sequence to be processed, a global semantic representation matrix is ​​generated, including: Obtain the text sequence to be processed; Sentences with semantic association are divided into the same semantic block based on the cosine similarity between adjacent sentences in the text sequence, and the maximum length of a single semantic block is limited to less than or equal to a first preset length. Furthermore, semantic content with a second preset length is set to overlap between adjacent semantic blocks. The Longformer model is selected, and a hybrid attention mechanism consisting of sliding window local attention and global attention is used. Formula (3) is used to encode attention for each semantic block separately. (3) in, The attention encoding output function for semantic blocks, The query vector matrix for semantic blocks. The key vector matrix of the semantic block. The value vector matrix of the semantic block. For normalization function, For attention mask matrix, This is the scaling factor; Each encoded semantic block is obtained, and the encoded features of each encoded semantic block are extracted and represented. The encoded features of all semantic blocks are globally fused using a cross-semantic block attention fusion layer to generate a unified semantic representation matrix for the entire text.

[0012] Preferably, information related to information dimension groups is extracted from the global semantic representation matrix and integrated to form a candidate information key set, including: The BERT-CRF model is used to extract core power entities related to information dimension groups and the relationships between core power entities from the global semantic representation matrix. Based on formula (4), the loss value of the BERT-CRF model is optimized to obtain the first candidate key information set. (4) in, This is the output of the loss value for the BERT-CRF model. It is the natural logarithm function. This is the global semantic representation matrix. A standard label sequence for core power entities. Given a global semantic representation matrix, this represents the conditional probability that the BERT-CRF model outputs the correct standard label sequence for core electricity entities. The sentence vector of each sentence in the power text is extracted from the global semantic representation matrix, and the weighted TextRank algorithm is used to obtain the weight score of each sentence according to formula (5). (5) in, The weight score for the sentence. The base score of a sentence is obtained iteratively based on the TextRank algorithm. The decision relevance coefficient. This represents the business importance coefficient. All sentence vectors are sorted from high to low according to their weight scores. The first preset number of sentence vectors with the highest scores are selected. A second candidate information key set is generated based on the sentence corresponding to each sentence vector. The first and second candidate information key sets are deduplicated and semantically merged to generate an initial candidate information key set. At the same time, based on the principles of global semantic consistency and decision target matching, the initial candidate information key set is processed to generate a strongly relevant candidate information key set.

[0013] Preferably, the candidate information key set is optimized using a small-sample driven extraction optimization method, including: Construct a small-sample support library for power-assisted decision-making scenarios; Obtain the task instructions from the current extraction task, generate the semantic vector of the current task instructions, and perform similarity matching calculation with the small sample instances in the small sample support library according to formula (6). (6) in, This is the semantic vector of the current task instruction. For the first in the small sample support library The semantic vectors corresponding to small sample instances The semantic vector of the current task instruction and the first The similarity between the semantic vectors corresponding to a small sample instance It is an L2 norm; Obtain all similarities between the current task and the small sample instances in the small sample support library, filter out the top-ranked second preset number of small sample instances, and output all the filtered small sample instances to the instance of the current extraction task to generate small sample task instructions. Using a small sample task instruction, the candidate key information set and the global semantic representation matrix are optimized based on the Llama3-8B model to generate an accurate key information set; After the process of generating accurate key information sets is revised, the updated records are stored in a small sample support library to achieve adaptive optimization.

[0014] Preferably, the information key set undergoes structural standardization processing, is input into a pre-defined power auxiliary decision-making system, and outputs a corresponding report for auxiliary decision generation, including: Obtain the optimized key information set and standardize it using the extracted dimensions from the information dimension group; Construct a power business rules database, set mapping rules between key information sets and decision recommendations, and form a power auxiliary decision-making system; The standardized information key set is input into the power auxiliary decision-making system, and the system outputs a report on the auxiliary decision-making based on the information key set.

[0015] Preferably, a power business rules database is constructed, mapping rules between key information sets and decision recommendations are set, forming a power auxiliary decision-making system, including: A rule database that supports visualization is constructed. The rule database adopts a five-layer structured design of "decision scenario - triggering condition - decision action - rule priority - constraint". The triggering conditions are used as the key information set, and the decision actions are used as decision suggestions. A mapping relationship between the key information set and the decision suggestions is constructed. Determine whether the data in the key information set meets the triggering conditions of a specific rule; When the data in the key information set meets the triggering conditions of a specific rule, the decision action of the specific rule with the highest priority is selected as the final decision recommendation based on the rule priority. Version management is implemented for the editing process of the rule database, multiple editing operation modes are set, and the data pairs of key information sets and output decisions generated each time are recorded, which are then converted into standardized rules. This completes the feedback optimization of the rule database and forms a power auxiliary decision-making system.

[0016] A second aspect of the present invention provides a system for generating auxiliary decision reports based on decision objectives and multi-source heterogeneous data, comprising: The target and data acquisition module is used to acquire the decision targets and multi-source heterogeneous power text to be processed from the power data report; The dimension parsing module is used to parse and obtain information dimension groups based on decision objectives; The text standardization module is used to obtain a standardized text sequence to be processed based on the multi-source heterogeneous power text to be extracted; The semantic encoding fusion module is used to perform segmented semantic encoding fusion on the text sequence to be processed to generate a global semantic representation matrix; The candidate information extraction module is used to extract information related to information dimension groups from the global semantic representation matrix, filter and integrate it to form a candidate information key set. The small sample optimization module is used to optimize the candidate information key set and generate the information key set by using a small sample-driven extraction optimization method. The report generation module is used to perform structural standardization processing on the information key set, input it into a preset power auxiliary decision-making system, and map out the corresponding auxiliary decision-making report.

[0017] Through the above technical solution, this generation method constructs a power scenario labeling system and uses the ALBERT model to analyze decision-making objectives, dynamically generating information dimension groups that match specific tasks. It extracts key information focused on decision-making, avoiding information redundancy or omissions caused by general extraction, and significantly improving the matching degree between the extraction results and decision-making objectives such as scheduling and fault handling. A two-level text correction method is employed, combining fast editing distance filtering from the power error correction dictionary with semantic verification from the ALBERT model, effectively correcting noise such as typos and non-standard abbreviations in field records. Simultaneously, through key information filtering and semantic consistency filtering, redundant and off-target content is further eliminated, ensuring the stability and reliability of the extraction results. Through semantic similarity segmentation and the Longformer hybrid attention mechanism, it achieves the processing of power regulations and fault analysis... The system employs global semantic encoding of long texts such as reports and retains causal and temporal relationships between paragraphs through a cross-semantic block attention fusion layer, overcoming the information fragmentation problem caused by traditional segmentation processing. It builds a small-sample support library and the context learning capability of the Llama3-8B model, requiring only a few historical instances to guide the model to adapt to new scenarios without large-scale labeled data or lengthy fine-tuning, significantly reducing the cost and cycle of cross-scenario adaptation. After structural standardization of key information sets, the system inputs them into a rule mapping library, automatically outputting auxiliary decision-making reports containing current situation analysis, risk warnings, and disposal suggestions. This achieves full-process automation from raw text to decision-making reports, compressing hours of manual work to minutes, realizing an end-to-end closed loop from unstructured text data to business decision support, and fully meeting the core needs of multi-scenario auxiliary decision-making in the power industry. Attached Figure Description

[0018] Figure 1 This is a connection block diagram of a method for extracting key text information based on decision objectives to assist in decision generation, according to an embodiment of the present invention. Figure 2 This is a connection diagram of the information dimension group of a method for extracting key text information based on decision objectives to assist in decision generation, according to an embodiment of the present invention. Figure 3 This is a connection diagram of a method for extracting key text information based on decision objectives to assist in decision generation, according to an embodiment of the present invention, for obtaining a standardized text sequence to be processed. Figure 4 This is a connection diagram of the construction of the corrected power text according to an embodiment of the present invention, which is a method for extracting key information from text based on decision objectives to assist in decision generation. Figure 5 This is a connection diagram of a method for generating a global semantic representation matrix based on a decision objective to extract key textual information for assisting decision generation, according to an embodiment of the present invention. Figure 6 This is a connection diagram of the generation of candidate information key sets in a method for extracting key text information based on decision objectives to assist in decision generation, according to an embodiment of the present invention. Figure 7 This is a connection diagram of the generated information key set of a method for extracting key text information based on decision objectives to assist in decision generation, according to an embodiment of the present invention. Figure 8 This is a connection diagram of the output report of a method for assisting decision generation based on extracting key textual information from a decision objective according to an embodiment of the present invention. Figure 9 This is a connection block diagram of a power auxiliary decision-making system based on a method for extracting key textual information based on decision objectives to assist in decision generation, according to an embodiment of the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0020] like Figure 1 The diagram shown is a connection block diagram of a method for extracting key textual information based on a decision objective to assist in decision generation, according to an embodiment of the present invention; Figure 1 The generation method includes the following steps: In step S10, the decision objectives of the power data report and the multi-source heterogeneous power text to be processed are obtained; In step S11, information dimension groups are obtained by parsing based on the decision objective; In step S12, a standardized text sequence to be processed is obtained based on the multi-source heterogeneous power text to be extracted; In step S13, the text sequence to be processed is segmented and semantically encoded and fused to generate a global semantic representation matrix; In step S14, information related to the information dimension group is extracted from the global semantic representation matrix and integrated to form a candidate information key set; In step S15, the candidate information key set is optimized and the information key set is generated using a small sample-driven extraction optimization method. In step S16, the information key set is structurally standardized, input into a preset power auxiliary decision-making system, and the corresponding auxiliary decision-making report is mapped and output.

[0021] Step S10 obtains the decision objectives of the power data report and the multi-source heterogeneous power text to be processed as the raw input for subsequent information extraction and report generation.

[0022] Step S11 involves analyzing and obtaining information dimension groups based on the decision objective, identifying preset extraction dimensions in the objective using the ALBERT model, and constructing targeted extraction task instructions to focus information extraction on the content required by the decision objective.

[0023] Step S12 obtains a standardized text sequence to be processed based on the multi-source heterogeneous power text to be extracted, including cleaning, error correction and word segmentation, to eliminate format differences and erroneous expressions, and to provide a clean and consistent input for subsequent semantic encoding.

[0024] Step S13 generates a global semantic representation matrix by performing segmented semantic encoding and fusion on the text sequence to be processed, and preserves the contextual association of the whole text by using a hybrid attention mechanism and cross-semantic block fusion technology.

[0025] Step S14 extracts information related to information dimension groups from the global semantic representation matrix, filters and integrates candidate information key sets, and preliminarily gathers core content highly relevant to the decision-making objective through entity relationship extraction and key sentence sorting.

[0026] Step S15 uses a small-sample-driven extraction optimization method to optimize the candidate information key set and generate an information key set. By using a small number of examples to eliminate redundancy and strengthen the key, the accuracy and robustness of information extraction are improved.

[0027] Step S16 involves standardizing the structure of the information key set, inputting it into a preset power auxiliary decision-making system, mapping and outputting the corresponding auxiliary decision-making report, thus completing the automated closed loop from data to decision.

[0028] This generation method acquires decision-making objectives and multi-source heterogeneous power text, and parses information dimension groups based on the objectives. The text is then cleaned, corrected, and segmented to obtain a standardized text sequence. Segmented semantic encoding and global fusion are used to generate a global semantic representation matrix. Entities, relationships, and key sentences related to the dimension groups are extracted from this matrix to form a candidate information key set. A few-sample-driven optimization method is then used to remove redundancy and strengthen key features, generating an accurate information key set. Finally, the set is structured and input into a power auxiliary decision-making system, mapping the output to a decision report. This method achieves fully automated report generation driven by decision-making objectives, significantly improving the decision adaptability and scenario generalization ability of information extraction, solving the problems of lost cross-paragraph associations and colloquial interference in long texts, significantly reducing manual costs, and improving decision-making efficiency and accuracy.

[0029] like Figure 2 The diagram shown is a connection block diagram of the information acquisition dimension group of a method for extracting key textual information based on a decision objective to assist in decision generation, according to an embodiment of the present invention; Figure 2 In order to meet the specific requirements of the power scenario and achieve accurate extraction of information dimension groups, in one embodiment of the present invention, obtaining information dimension groups based on decision objectives may include the following steps: In step S20, a labeling system for power auxiliary decision-making scenarios is constructed, including power auxiliary decision-making scenarios such as power grid dispatching, equipment operation and maintenance, fault handling, safety management and control, and power grid planning. In step S21, standardized extraction dimensions are preset based on each power-assisted decision-making scenario; In step S22, the ALBERT model is used to parse the text content of the decision target, and relevant extraction dimensions are obtained from all preset standardized extraction dimensions to construct information dimension groups. At the same time, the extraction task of task instruction-output requirements-instance is generated with the construction of information dimension groups as the task instruction.

[0030] A tagging system covering typical scenarios such as power grid dispatching, equipment operation and maintenance, fault handling, safety management, and power grid planning is constructed, and standardized extraction dimensions are preset for each scenario (for example, the standardized extraction dimensions for fault handling scenarios include: fault equipment type, fault location, fault occurrence time, fault cause, risk level, implemented handling measures, suggested handling plan, and historical similar cases). The ALBERT model is used to parse the decision target text and select matching dimensions from the preset extraction dimensions to construct information dimension groups. The ALBERT model contains a 12-layer Transformer encoder, which can perform cross-layer parameter sharing, embedding matrix factorization, and sentence order prediction. The ALBERT model can obtain the decision target text (such as "Please analyze the cause of this transformer fault and provide handling suggestions") as input. The ALBERT model encodes, integrates, and pools the input text to obtain the semantic vector of the decision target text. The semantic vector is then used to calculate the cosine similarity with the semantic vector of each preset extraction dimension (such as 'fault equipment type', 'fault cause', etc.), and the dimensions with similarity higher than a threshold (set to 0.75 in this embodiment) are selected to form information dimension groups. The ALBERT model takes the target text for power decision-making as input and outputs the matched power scenario extraction dimensions, achieving an automated transformation from decision intent to information extraction requirements. With its small parameter count and fast inference speed, the ALBERT model is suitable for real-time processing needs in power systems, capable of dimension parsing in milliseconds. By utilizing power-related text in the pre-training corpus, the ALBERT model accurately understands power terminology and scenario descriptions, improving the accuracy of dimension matching. Centered on information dimension groups, a structured extraction task instruction is automatically generated, containing "what to extract" (task instruction), "output format" (output requirements), and "an example" (instance). This avoids the rigidity of manually setting fixed dimensions in traditional methods, allowing information extraction to flexibly adapt to different decision objectives and significantly improving decision adaptability.

[0031] like Figure 3 The diagram shown is a connection block diagram illustrating the acquisition of a standardized text sequence to be processed, according to an embodiment of the present invention, of a method for extracting key text information based on a decision objective to assist in decision generation; in Figure 3 In order to standardize the text preprocessing process and improve the consistency and extraction stability of subsequent encoding inputs, in one embodiment of the present invention, obtaining a standardized text sequence to be processed based on the multi-source heterogeneous power text to be extracted may include the following steps: In step S30, the multi-source heterogeneous power text to be processed is cleaned and standardized preprocessed to obtain the format-normalized multi-source heterogeneous power text. In step S31, a two-level text correction method is used to process the format-normalized multi-source heterogeneous power text and construct the corrected power text. In step S32, a dedicated word segmentation tool for the power industry is used to segment the corrected power text into sentences and words, and output a standardized sequence of text to be processed.

[0032] The multi-source heterogeneous power text to be processed undergoes cleaning and standardization preprocessing, such as removing HTML tags, special symbols, and redundant spaces, unifying the encoding format, and converting tables and lists from different sources into plain text format, resulting in format-normalized multi-source heterogeneous power text. A two-stage text correction method is then employed. The first stage, based on a power error correction dictionary and the Levenshtein distance algorithm, maps common synonymous terms (such as "circuit breaker" and "switch," "trip" and "tripped") to standard terms. When the edit distance is less than or equal to a first preset threshold (e.g., 2), the term is directly replaced. The second level evaluates the semantic matching degree of the replaced result. If the matching degree is greater than or equal to the second preset threshold (e.g., 0.85), the error correction is confirmed; otherwise, the original word is retained to avoid miscorrection. Power-specific word segmentation tools (e.g., word segmenters that integrate power domain lexicons) are used to segment the corrected text into sentences and words, outputting a standardized sequence of text to be processed. This effectively corrects common colloquialisms, typos, and formatting issues in on-site operation and maintenance logs and fault records, freeing subsequent semantic encoding from noise interference. At the same time, power-specific word segmentation ensures the correct segmentation of professional terms and improves the accuracy of entity recognition.

[0033] like Figure 4 The diagram shown is a connection block diagram of a method for extracting key textual information based on decision-making objectives to assist in decision generation, according to an embodiment of the present invention; in Figure 4 In order to enhance error correction capabilities and reduce the interference of colloquial and non-standard expressions on extraction accuracy, in one embodiment of the present invention, a two-level text error correction method is used to process the format-normalized multi-source heterogeneous power text. Constructing the error-corrected power text may include the following steps: In step S40, all candidate terms in the power field with the same semantics are mapped to the same standard terms to construct a power error correction dictionary; In step S41, all candidate terms in the power sector to be corrected are identified and located; In step S42, based on the Levenshtein distance algorithm, formula (1) is used to calculate the editing distance between all candidate terms in the power field to be corrected and their corresponding standard terms. (1) in, These are candidate terms in the power sector that need correction. To correspond with the same standard terminology as the candidate terminology in the power sector that needs to be corrected, This is the minimum edit distance between the first few characters of a candidate term in the power sector to be corrected and the first few characters of the same standard term. These are the character index numbers, which are the prefixes of the candidate terms in the power sector to be corrected. The characters are converted to the same standard terminology before The minimum number of operations required to generate one character For boundary condition calculation terms, when or When one of them is 0 (i.e., one of the strings is empty), the minimum edit distance is equal to the length of the non-empty string, corresponding to the number of operations for inserting or deleting all characters. The minimum value function selects the editing operation with the lowest cost as the current editing operation. The unit cost of editing operations, For the edit distance calculation item corresponding to the character deletion operation, the preceding part of the candidate terms in the power field to be corrected will be... Deleting one character increments the cost counter by one per operation. For the edit distance calculation item corresponding to the character insertion operation, in the candidate terms of the power field to be corrected Insert the same standard term after the character. For each character, the cost count for a single operation increases by one. To replace the indicator cost function, For the edit distance calculation item corresponding to the character replacement operation, when the preceding term in the power field candidate term to be corrected... The character and the preceding part of the same standard term When the characters are not equal, the replacement operation count cost increases by one; In step S43, when the edit distance of the candidate term in the power field to be corrected is less than or equal to the first preset threshold, the corresponding standard term is output and included in the candidate correction set to be corrected. In step S44, the semantic matching degree of the error correction result is generated by formula (2) using the ALBERT model on the error correction result of the candidate error correction set. (2) in, The semantic matching degree function of the error correction result. For the purpose of error correction, the same standard terminology should be used. For the complete sentence containing the same standard term in electrical texts. This refers to the high-dimensional semantic vectors generated by processing the same standard term using the ALBERT model. This refers to the high-dimensional semantic vectors generated from the complete sentences containing the same standard term in electricity texts after processing by the ALBERT model. It is an L2 norm; In step S45, error correction is completed if the semantic matching degree of the error correction result is greater than or equal to the second preset threshold, and the corrected power text is constructed.

[0034] A power error correction dictionary is constructed, containing a large number of candidate terms in the power field and their corresponding standard terms. For example, "knife switch" and "disconnect switch" are both mapped to "disconnect switch". All entries belonging to the candidate terms in the text to be corrected are identified and located. The input to the Levenshtein distance algorithm is the candidate terms and their corresponding standard terms, both of which are sequences of Chinese characters. The operation cost of the algorithm is set to 1 for each of deletion / insertion / replacement. The output is the minimum edit distance. A first preset threshold of 2 can be set. This threshold is based on the fact that the edit distance of typical candidate terms is 1~2, accounting for over 90%. Therefore, setting the threshold to 2 can achieve a balance between recalling near-miss errors and avoiding excessive input to the secondary level, adapting to the error word distribution characteristics of the power scenario. When the edit distance is ≤2, replacement pairs of "candidate term → standard term" are included in the candidate error correction set. However, relying solely on edit distance may lead to misjudgments (e.g., "voltage" and "current" have a small edit distance but different semantics). A secondary verification is performed using the ALBERT model to calculate the semantic matching degree, thereby improving the error correction accuracy. The ALBERT model contains a 12-layer Transformer encoder. The words before and after replacement are placed in the original text context to form the original sentence and the corrected sentence, respectively. The inputs to the ALBERT model include the original sentence and the sentence corrected by the replacement words, and the outputs are the semantic representation vectors of the original sentence and the corrected sentence, respectively. The semantic matching degree is calculated by taking the cosine similarity of the semantic representation vectors of two sentences. Only when the semantic matching degree is greater than or equal to a second preset threshold (e.g., the second preset threshold is 0.85) is the error finally confirmed and written into the corrected power text. The input of the ALBERT model is the complete sentence containing colloquial misspellings in power operation and maintenance, scheduling and other scenarios. The output is the semantic consistency score of the sentence pair before and after the error, which is used to determine whether the first-level edit distance error correction is correct. ALBERT's bidirectional encoding feature can simultaneously detect the power context before and after the misspelling (such as equipment model, parameter unit, and handling action), so that the semantic matching degree calculation can reflect whether the original sentence's power semantics are incorrect after the error correction. The first level of the two-level text error correction method quickly filters obvious errors, while the second level of fine verification avoids false positives, which greatly improves the error correction accuracy. It is especially suitable for situations that frequently occur in field records, such as writing "trip" as "trip-smash" or "transformer" as "transformer pressure device", which significantly enhances the method's anti-interference ability.

[0035] like Figure 5The diagram shown is a connection block diagram for generating a global semantic representation matrix in a method for extracting key textual information based on a decision objective to assist in decision generation, according to an embodiment of the present invention; Figure 5 In order to optimize long text encoding strategies, preserve cross-paragraph semantic relationships, and improve global understanding capabilities, in one embodiment of the present invention, generating a global semantic representation matrix after segmented semantic encoding fusion of the text sequence to be processed may include the following steps: In step S50, the text sequence to be processed is obtained; In step S51, semantically related sentences are divided into the same semantic block based on the cosine similarity between adjacent sentences in the text sequence, and the maximum length of a single semantic block is limited to be less than or equal to a first preset length, and the semantic content of an overlap of a second preset length is set between adjacent semantic blocks. In step S52, the Longformer model is selected, and a hybrid attention mechanism consisting of sliding window local attention and global attention is used to encode and output attention for each semantic block using formula (3). (3) in, The attention encoding output function for semantic blocks, The query vector matrix for semantic blocks. The key vector matrix of the semantic block. The value vector matrix of the semantic block. For normalization function, For attention mask matrix, This is the scaling factor; In step S53, each encoded semantic block is obtained, and the encoded features of each encoded semantic block are extracted and represented. The encoded features of all semantic blocks are globally fused using a cross-semantic block attention fusion layer to generate a unified semantic representation matrix for the entire text.

[0036] A standardized text sequence to be processed is obtained. Based on the cosine similarity between adjacent sentences, sentences with close semantic connections are grouped into the same semantic block. The maximum length of a single semantic block is limited to a first preset length (e.g., 3072 tokens) to avoid exceeding the model's processing limit. A second preset length (e.g., 128 tokens) of semantic content overlaps between adjacent semantic blocks to ensure contextual semantic coherence. The Longformer model is used to encode the semantic blocks. The Longformer model employs a hybrid attention mechanism combining sliding window local attention and global attention. The Longformer model contains a 12-layer encoder. The sliding window local attention mechanism performs attention calculations only on each token and its neighboring tokens within a fixed window size (512 tokens in this embodiment). The global attention mechanism automatically sets the first sentence of each semantic block, and sentences containing keywords related to power decision dimensions (such as fault equipment type, risk level, etc.), as global attention tokens. These global attention tokens can interact with all tokens in the entire text to capture semantic connections across paragraphs. The input to the Longformer model is a semantic block, and each semantic block is converted into a token after processing. The system takes an ID sequence and appends a positional encoding; the output is a context encoding vector for each token, forming an encoded semantic block. The Longformer model's input semantic blocks come from power industry texts (such as maintenance logs and fault records), which, after standardization, contain a large number of power industry terms and long paragraph descriptions. The Longformer model's sliding window size of 512 tokens adapts to the common paragraph length in power industry texts (typically about 300-500 words per paragraph), ensuring semantic integrity within a single paragraph. The Longformer model's global attention automatically locates sentences containing keywords related to power industry decision-making, allowing the model to focus on key information relevant to the decision-making objective and ignore irrelevant descriptions. The sliding output encoding features of the Mer model are directly used to generate the subsequent global semantic representation matrix, preserving the causal relationships across paragraphs in the power text (e.g., a fault in device A causes a tripping of line B). The Longformer model, based on power scenarios, achieves efficient and accurate encoding of long power texts, laying the foundation for subsequent information extraction. After encoding each semantic block, its encoding features are extracted. Using a cross-semantic block attention fusion layer, the encoding features of all semantic blocks undergo secondary attention interaction to generate a unified semantic representation matrix for the entire text. This matrix contains local information from each segment and also incorporates long-range dependencies across segments. This effectively overcomes the shortcomings of traditional segmentation processing that fragments the context, significantly improving the quality of information extraction in long text scenarios.

[0037] like Figure 6The diagram shown is a connection block diagram for generating a candidate information key set in a method for extracting key text information based on a decision objective to assist in decision generation, according to an embodiment of the present invention; Figure 6 In order to comprehensively capture content related to the decision-making objective from different dimensions by utilizing entities, relationships, and key sentences, and to improve the coverage and accuracy of the candidate set, in one embodiment of the present invention, extracting information related to information dimension groups from the global semantic representation matrix and integrating it to form a candidate information key set may include the following steps: In step S60, the BERT-CRF model is used to extract the core power entities related to the information dimension group and the relationships between the core power entities from the global semantic representation matrix; In step S61, the loss value of the BERT-CRF model is optimized according to formula (4) to obtain the first candidate key information set. (4) in, This is the output of the loss value for the BERT-CRF model. It is the natural logarithm function. This is the global semantic representation matrix. A standard label sequence for core power entities. Given a global semantic representation matrix, this represents the conditional probability that the BERT-CRF model outputs the correct standard label sequence for core electricity entities. In step S62, the sentence vector of each sentence in the power text is extracted from the global semantic representation matrix, and the weighted TextRank algorithm is used to obtain the weight score of each sentence according to formula (5). (5) in, The weight score for the sentence. The base score of a sentence is obtained iteratively based on the TextRank algorithm. The decision relevance coefficient. This represents the business importance coefficient. In step S63, all sentence vectors are sorted from high to low according to their weight scores, and the first preset number of sentence vectors with the highest scores are selected. A second candidate information key set is generated based on the sentence corresponding to each sentence vector. In step S64, the first candidate information key set and the second candidate information key set are deduplicated and semantically merged to generate an initial candidate information key set. At the same time, based on the principles of global semantic consistency and decision target matching degree, the initial candidate information key set is processed to generate a strongly related candidate information key set.

[0038] The BERT-CRF model is used to extract core power entities (such as equipment names, parameter values, time, and status) and their relationships (such as "circuit breaker A connects to bus B" and "transformer C overload causes line D to trip") related to information dimension groups from the global semantic representation matrix. The BERT-CRF model consists of a BERT layer network and a CRF layer network. The BERT layer network contains 12 Transformer encoder layers, with the global semantic representation matrix as input. Each Transformer encoder layer incorporates a multi-head attention mechanism and a feedforward neural network. The output of the BERT layer network is the input to the CRF layer. The output of the BERT layer network undergoes dimensionality transformation through a fully connected network, mapping it to the power entity label space, and then inputs it into the CRF layer network. The CRF layer network learns the transition probability matrix between labels and performs global optimal path decoding on the output of the BERT layer network, outputting the optimal entity label sequence and the relationships between entities. The BERT-CRF model is optimized by combining the cross-entropy loss function with the negative log-likelihood loss function of CRF. The training dataset contains no fewer than 1000 power dispatch logs and fault reports. Through the above network layer structure settings and parameter training, the model obtains more accurate power core entity boundaries and relationship classifications, resulting in the first candidate information key set. Sentence vectors for each sentence are extracted from the global semantic representation matrix. The input to the weighted TextRank algorithm is the sentence vector for each sentence, and the output is the weight score for each sentence through iteration. Weighted TextRank introduces relevance weights between sentences and the decision objective on top of traditional PageRank. As a decision relevance coefficient, the information dimension group of the decision objective in step S11 is ordered and coupled with the sentences in step S62, so that the extracted sentences always revolve around the decision objective. This adapts to the characteristics of the power scenario where the focus of attention varies greatly under different decision objectives such as "fault handling", "equipment operation and maintenance" and "grid dispatch". The business importance coefficient is determined based on the entity extraction results of the BERT-CRF model in step S60, according to sentences in the power business scenario. It can be configured according to five major power auxiliary decision-making scenarios, and the scenario switching can be performed. Synchronous switching is performed to rank sentences more relevant to the target. All sentences are then sorted from highest to lowest weight, and a first predetermined number (e.g., 10) of the top-ranking sentences are selected to form the second candidate key set (a set of key sentences). The first and second candidate key sets are then deduplicated and semantically merged. Based on the principles of global semantic consistency and target-matching accuracy, the merged initial candidate key set undergoes a second screening to remove content irrelevant to the target or semantically contradictory, generating a strongly relevant candidate key set.

[0039] like Figure 7 The diagram shown is a connection block diagram of the generated information key set in a method for extracting key textual information based on a decision objective to assist in decision generation, according to an embodiment of the present invention; Figure 7 In order to reduce the cost of fine-tuning and achieve accurate optimization of the candidate information key set with a very small number of examples, thereby improving the adaptive iteration capability, in one embodiment of the present invention, optimizing the candidate information key set using a small-sample driven extraction optimization method may include the following steps: In step S70, a small sample support library for power-assisted decision-making scenarios is constructed; In step S71, the task instructions in the current extraction task are obtained, the semantic vector of the current task instructions is generated, and similarity matching calculation is performed with the small sample instances in the small sample support library according to formula (6). (6) in, This is the semantic vector of the current task instruction. For the first in the small sample support library The semantic vectors corresponding to small sample instances The semantic vector of the current task instruction and the first The similarity between the semantic vectors corresponding to a small sample instance It is an L2 norm; In step S72, obtain all similarities between the current task and the small sample instances in the small sample support library, filter out the top-ranked second preset number of small sample instances, and output all the filtered small sample instances to the instance of the current extraction task to generate a small sample task instruction. In step S73, the candidate key information set and the global semantic representation matrix are optimized using the small sample task instruction based on the Llama3-8B model to generate an accurate key information set; In step S74, the process of generating the accurate key information set is revised and updated in the small sample support library to achieve adaptive optimization.

[0040] A small-sample support library for power-assisted decision-making scenarios is constructed, including successfully extracted instances of various historical decision-making tasks. Each instance contains task instructions, input text fragments, and key information about the expected output. The task instructions for the current extracted task are obtained, converted into semantic vectors using a semantic encoding model, and then compared with the instruction vectors of all instances in the small-sample support library for similarity calculation. The top two preset number (e.g., 3-5) of similarity-ranked small-sample instances are selected and assembled with the current task instructions into a small-sample task instruction (containing the format "task description + several instances + current input"). The Llama3-8B model is then used to implement this small-sample support library. The sample task instructions optimize the candidate key information set. Llama3-8B is a large language model based on an autoregressive Transformer architecture, containing a 32-layer Transformer decoder. The input of the Llama3-8B model includes small sample task instructions and context composed of candidate key information sets and a global semantic representation matrix. The Llama3-8B model, guided by its pre-trained knowledge and small sample instances, identifies and removes redundant information in the candidate set through autoregression, fills in missing key points, unifies the expression style, and finally outputs a structured and accurate key information set, including entities, relationships, and key sentences. The instances in the small sample task instructions input to the Llama3-8B model all come from power industry auxiliary decision-making scenarios (such as fault handling and equipment operation and maintenance), enabling the model to quickly adapt to the professional terminology and expression habits of the power industry. The input candidate key information set and global semantic representation matrix contain contextual information of the power text, allowing the model to determine which information is truly relevant to the decision-making objective. The output format of the Llama3-8B model strictly follows the requirements of information dimension groups (such as JSON format), ensuring that the optimized key information set can be directly used for downstream decision report generation. The process of successfully generating the accurate key information set (including task instructions, inputs and outputs, and the instances used) was recorded and updated in the small-sample support library. This enabled the library to expand and evolve, allowing subsequent similar tasks to find more suitable instances. Thus, the effect of extensive fine-tuning can be achieved with only a few labeled examples, significantly reducing the cost and time required for scenario migration. Simultaneously, the dynamic updates to the support library give the method continuous learning capabilities, becoming more accurate with use and significantly improving its long-term deployment practicality.

[0041] like Figure 8 The diagram shown is a connection block diagram of the output report of a method for assisting decision generation based on extracting key textual information according to a decision objective, as per an embodiment of the present invention. Figure 8 In order to generate decision recommendations, in one embodiment of the present invention, the information key set is structurally standardized, input into a preset power auxiliary decision system, and the corresponding auxiliary decision generation report is mapped and output, which may include the following steps: In step S80, the optimized information key set is obtained and standardized using the extracted dimensions in the information dimension group; In step S81, a power business rules database is constructed, mapping rules between key information sets and decision recommendations are set, and a power auxiliary decision-making system is formed. In step S82, the standardized information key set is input into the power auxiliary decision-making system, and a report on auxiliary decision generation based on the information key set is output.

[0042] The optimized key information set undergoes standardized field mapping according to extraction dimensions, unifying field names, data formats, and enumeration values ​​(e.g., unifying "risk level" as "risk grade" and "serious" as "major"). It supports three output formats: JSON format using a key-value pair structure, with one field corresponding to each extraction dimension, facilitating system interface calls; Excel format generating tables according to "extraction dimension - extraction content - confidence level - original source," supporting batch export and statistics; and tabular text format based on a plain text structured table, adaptable to manual viewing scenarios without system integration. A power business rules database stores the mapping rules between key information and decision recommendations. The standardized key information set is input into the power auxiliary decision-making system, which automatically matches trigger conditions and generates the final auxiliary decision-making report. The report content can include a problem overview, key evidence, recommended measures, and risk warnings. This fully automates the data input to report output process, significantly improving decision response speed; simultaneously, standardized output ensures consistent report format and objective content, avoiding human bias.

[0043] like Figure 9 The diagram shows a connection block diagram for constructing a power auxiliary decision-making system based on a method for extracting key textual information based on decision objectives to assist in decision generation, according to an embodiment of the present invention; a structured rule base is constructed to achieve automatic reasoning and version traceability management of decision suggestions. Figure 9 To ensure the reliability of generated decision recommendations, a maintainable and updatable power auxiliary decision-making system is constructed. In one embodiment of the invention, constructing a power business rules database and setting mapping rules between key information sets and decision recommendations to form a power auxiliary decision-making system may include the following steps: In step S90, a rule database that supports visualization is constructed. The rule database adopts a five-layer structured design of "decision scenario - triggering condition - decision action - rule priority - constraint". In step S91, the triggering conditions are used as a key information set, and the decision actions are used as decision suggestions to construct a mapping relationship between the key information set and the decision suggestions; In step S92, it is determined whether the data of the key information set meets the triggering conditions of a specific rule; In step S93, if the data in the key information set meets the triggering conditions of a specific rule, the decision action of the specific rule with the highest priority is selected as the final decision suggestion based on the rule priority. In step S94, version management is performed on the editing process of the rule database, multiple editing operation modes are set, and the basis and time of each modification are recorded to complete the feedback optimization of the rule database and form a power auxiliary decision-making system.

[0044] A rule database supporting visualization is constructed, employing a five-layer structured design: "decision scenario - trigger condition - decision action - rule priority - constraint". The decision scenario represents the applicable business context (e.g., "transformer fault handling"). The trigger condition is a set of logical expressions, such as "temperature > 80 AND..." The duration is >30 minutes; the decision action is the specific suggested text; rule priority is used to resolve conflicts when multiple rules are triggered simultaneously, with higher values ​​indicating higher priority; constraints define the preconditions or side effects of executing the action (e.g., "If other load reduction operations have already been performed, this rule will be temporarily suspended"); the triggering conditions are mapped to fields in the key information set, and the decision action is used as the output suggestion to establish a mapping relationship; when the decision output process is running, it checks whether each key information set meets the triggering condition of a rule. If it does, all rules that meet the conditions are collected, sorted by priority, and the highest priority rule is selected for output. The rule database supports a visual editing interface, allowing business personnel to add, modify, or delete rules. Each modification records the operator, modification basis, and time, forming a complete version history for easy backtracking and auditing; based on actual application feedback (e.g., decision adoption rate, incident handling effectiveness), the rules are statistically analyzed regularly, and optimization suggestions are proposed to form a closed-loop feedback optimization. The rule base structure is clear and easy to maintain. Version management ensures the traceability of changes, and feedback optimization enables the rule base to continuously evolve with business development, thereby ensuring the long-term stable and intelligent operation of the auxiliary decision system.

[0045] In one embodiment of the present invention, the construction of the rule database requires the setting and example of fields; its core fields may include the following: the rule ID is a globally unique identifier (such as RULE-FAULT-001) used as the primary key index for each decision rule; the relevant scenario categorizes the rules into five core business domains through enumeration values ​​(such as "fault handling", "equipment maintenance", etc.); the triggering condition is a Boolean logical expression (supporting AND, OR, NOT). The combination describes the objective conditions that the triggering rule must meet (e.g., fault equipment type = "main transformer" AND risk level = "critical"); the decision action is a pre-set standardized text array of handling suggestions (e.g., ["Immediate execution: shut down the equipment", "Within 1 hour: conduct insulating oil chromatography analysis"]); the rule priority is set with integer priority weights of 1-5 levels (level 1 is the highest) to handle conflict scenarios where multiple rules are triggered concurrently; the applicable constraints are the physical boundaries and technical parameters that limit the rule's effectiveness (e.g., "main transformers with voltage levels of 110kV and above"); the rule source is the record of the national standard, industry standard, or enterprise procedure number on which the rule is based (e.g., Q / GDW1799.1-2013).

[0046] In one embodiment of the present invention, the execution process of mapping key information sets to decisions can be as follows: In step S100, field alignment verification is performed, comparing the structured key information set with the trigger condition fields of the decision mapping library one by one, and automatically correcting differences in field names, units, and enumeration values ​​(such as converting "10kV" to "10 kilovolts" and "knife switch" to "disconnect switch") to ensure matching accuracy.

[0047] In step S101, two levels of trigger condition matching are performed. Precise matching prioritizes traversing all rules in the current scenario. If the key information set fully meets the trigger condition of a certain rule, the corresponding decision action is directly triggered. Fuzzy matching calculates the semantic similarity between the key information set and the rule trigger condition if there is no precise matching rule. When the similarity is ≥0.8, the rule with the highest similarity is triggered, and "Reference similar rule generation" is marked in the decision suggestion. Combined condition enhanced matching supports multi-dimensional combined triggering, such as "failure time = night AND failure location = remote mountainous area AND risk level = relatively high", and automatically adds special suggestions such as night operation safety requirements and material allocation in remote areas.

[0048] In step S102, multi-rule conflict resolution is performed. When multiple rules are triggered simultaneously, the resolution is executed in the following priority order, retaining the highest priority decision action and using the others as alternatives. It can include the following rule priorities, matching accuracy (exact match > fuzzy match), risk level (major > significant > moderate > low), and timeliness (immediate execution > within 1 hour > within 24 hours).

[0049] In step S103, decision recommendations are integrated and optimized. The triggered decision actions are classified into immediate execution (within 1 hour), follow-up (within 24 hours), and long-term prevention (within 7 days). Details such as the implementing entity, safety precautions, and resource requirements are added to form a complete decision recommendation system.

[0050] To ensure the standardization of output reports, multiple text formats are supported for export. In one embodiment of the present invention, a template for power business auxiliary decision-making reports can be set. This template is the final output carrier of the technical solution of the present invention, which aims to present the automatically extracted and optimized key information to decision-makers in a structured and highly readable manner, and directly generate hierarchical decision-making suggestions, thereby realizing a closed loop from "data input" to "decision output".

[0051] The first part of the template, "Basic Report Information," records metadata such as the report's unique identifier, generation time, decision-making scenario, and text source, ensuring that each report is traceable and auditable. It also provides quantitative evidence for subsequent performance evaluation and optimization by recording the total text length and the number of matched small samples. The second part, "Structured Key Information Summary," is the core of this template. It displays extracted content item by item according to defined information dimension groups, along with confidence scores and original text fragments. The confidence level mechanism (1.0 to <0.6) intuitively reflects the reliability of the information, helping decision-makers determine whether manual review is needed, addressing the pain point of uninterpretable output from traditional black-box models. The third part, "Auxiliary Decision-Making Suggestions," maps key information to four levels of measures—immediate execution, follow-up, long-term prevention, and alternative references—based on a rule database. This achieves refined classification of decision-making suggestions, ensuring rapid response to urgent issues while also considering medium- and long-term hazard management. Part Four, “Information Quality Statement,” quantitatively demonstrates the completeness of extraction, the effectiveness of noise reduction, and the results of long text processing, providing transparency for users and data support for system iteration. The appendix provides small sample examples of original text excerpts and matching, as well as the basis for decision-making rules, further enhancing the credibility and auditability of the report.

[0052] A second aspect of this invention provides a system for generating auxiliary decision-making reports based on decision objectives and multi-source heterogeneous data. This system includes an objective and data acquisition module, a dimension parsing module, a text standardization module, a semantic encoding fusion module, a candidate information extraction module, a small-sample optimization module, and a report generation module. Specifically, the objective and data acquisition module acquires the decision objectives of the power data report and the multi-source heterogeneous power text to be processed. The dimension parsing module acquires information dimension groups based on the decision objectives. The text standardization module acquires a standardized text sequence to be processed based on the multi-source heterogeneous power text to be extracted. The semantic encoding fusion module performs segmented semantic encoding fusion on the text sequence to be processed to generate a global semantic representation matrix. The candidate information extraction module extracts information related to the information dimension groups from the global semantic representation matrix, filters and integrates it to generate a candidate information key set. The small-sample optimization module optimizes the candidate information key set using a small-sample driven extraction optimization method to generate an information key set. The report generation module performs structural standardization processing on the information key set, inputs it into a preset power auxiliary decision-making system, and maps and outputs the corresponding auxiliary decision-making report.

[0053] Through the above technical solution, this generation method constructs a power scenario labeling system and uses the ALBERT model to analyze decision-making objectives, dynamically generating information dimension groups that match specific tasks. It extracts key information focused on decision-making, avoiding information redundancy or omissions caused by general extraction, and significantly improving the matching degree between the extraction results and decision-making objectives such as scheduling and fault handling. A two-level text correction method is employed, combining fast editing distance filtering from the power error correction dictionary with semantic verification from the ALBERT model, effectively correcting noise such as typos and non-standard abbreviations in field records. Simultaneously, through key information filtering and semantic consistency filtering, redundant and off-target content is further eliminated, ensuring the stability and reliability of the extraction results. Through semantic similarity segmentation and the Longformer hybrid attention mechanism, it achieves the processing of power regulations and fault analysis... The system employs global semantic encoding of long texts such as reports and retains causal and temporal relationships between paragraphs through a cross-semantic block attention fusion layer, overcoming the information fragmentation problem caused by traditional segmentation processing. It builds a small-sample support library and the context learning capability of the Llama3-8B model, requiring only a few historical instances to guide the model to adapt to new scenarios without large-scale labeled data or lengthy fine-tuning, significantly reducing the cost and cycle of cross-scenario adaptation. After structural standardization of key information sets, the system inputs them into a rule mapping library, automatically outputting auxiliary decision-making reports containing current situation analysis, risk warnings, and disposal suggestions. This achieves full-process automation from raw text to decision-making reports, compressing hours of manual work to minutes, realizing an end-to-end closed loop from unstructured text data to business decision support, and fully meeting the core needs of multi-scenario auxiliary decision-making in the power industry.

[0054] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various simple modifications can be made to the technical solutions of the present invention, and these simple modifications all fall within the protection scope of the present invention. Furthermore, it should be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

Claims

1. A method for generating auxiliary decision-making reports based on decision objectives and multi-source heterogeneous data, characterized in that, include: The decision-making objectives for acquiring power data reports and the multi-source heterogeneous power texts to be processed; Based on the decision-making objectives, analyze and obtain information dimension groups; A standardized text sequence to be processed is obtained based on the multi-source heterogeneous power text to be extracted; The text sequence to be processed is segmented and semantically encoded and fused to generate a global semantic representation matrix; Extract information related to information dimension groups from the global semantic representation matrix, filter and integrate it to form a candidate information key set; A small-sample-driven extraction optimization method is used to optimize the candidate information key set and generate the information key set. The information key set is structurally standardized, input into a preset power auxiliary decision-making system, and the corresponding auxiliary decision-making report is output. Based on the decision-making objectives, information dimension groups are obtained, including: Construct a labeling system for power auxiliary decision-making scenarios, including power grid dispatching, equipment operation and maintenance, fault handling, safety management and control, and power grid planning scenarios; Standardized extraction dimensions are preset for each power-assisted decision-making scenario; The ALBERT model is used to parse the text content of the decision target, and relevant extraction dimensions are obtained from all preset standardized extraction dimensions to construct information dimension groups. At the same time, the extraction task of task instruction-output requirements-instance is generated with the construction of information dimension groups as the task instruction. The candidate information key set is optimized using a few-sample-driven extraction optimization method, including: Construct a small-sample support library for power-assisted decision-making scenarios; Obtain the task instructions from the current extraction task, generate the semantic vector of the current task instructions, and perform similarity matching calculation with the small sample instances in the small sample support library according to formula (6). ，（6） in, This is the semantic vector of the current task instruction. For the first in the small sample support library The semantic vectors corresponding to small sample instances The semantic vector of the current task instruction and the first The similarity between the semantic vectors corresponding to a small sample instance It is an L2 norm; Obtain all similarities between the current task and the small sample instances in the small sample support library, filter out the top-ranked second preset number of small sample instances, and output all the filtered small sample instances to the instance of the current extraction task to generate small sample task instructions. Using a small sample task instruction, the candidate key information set and the global semantic representation matrix are optimized based on the Llama3-8B model to generate an accurate key information set; After the process of generating accurate key information sets is revised, the updated records are stored in a small sample support library to achieve adaptive optimization.

2. The generation method according to claim 1, characterized in that, Based on the multi-source heterogeneous power data to be extracted, a standardized sequence of text to be processed is obtained, including: The multi-source heterogeneous power text to be processed is cleaned and standardized preprocessed to obtain the format-normalized multi-source heterogeneous power text. A two-level text correction method is used to process the format-normalized multi-source heterogeneous power text and construct the corrected power text. The power industry-specific word segmentation tool is used to segment the corrected power industry text into sentences and words, and output a standardized sequence of text to be processed.

3. The generation method according to claim 2, characterized in that, A two-level text correction method is used to process the format-normalized multi-source heterogeneous power text, and the corrected power text is constructed, including: Map all candidate terms in the power field with the same semantics to the same standard terminology to construct a power error correction dictionary; Identify and locate all candidate terms in the power sector that need to be corrected; Based on the Levenshtein distance algorithm, formula (1) is used to calculate the editing distance between all candidate terms in the power field to be corrected and their corresponding standard terms. ，（1） in, These are candidate terms in the power sector that need correction. To correspond with the same standard terminology as the candidate terminology in the power sector that needs to be corrected, This is the minimum edit distance between the first few characters of a candidate term in the power sector to be corrected and the first few characters of the same standard term. These are the character index numbers, For boundary condition calculation terms, It is a minimum value function. The unit cost of editing operations, To replace the indicator cost function, The preceding text is a list of candidate terms for the power sector that need correction. One character, For the same standard term One character; When the edit distance is less than or equal to the first preset threshold, the corresponding standard term is output and included in the candidate error correction set to be corrected; The semantic matching degree of the error correction result is generated by formula (2) using the ALBERT model to correct the error correction result of the candidate error correction set. ，（2） in, The semantic matching degree function of the error correction result. For the purpose of error correction, the same standard terminology should be used. For the complete sentence containing the same standard term in electrical texts. This refers to the high-dimensional semantic vectors generated by processing the same standard term using the ALBERT model. This refers to the high-dimensional semantic vectors generated from the complete sentences containing the same standard term in electricity texts after processing by the ALBERT model. It is an L2 norm; Error correction is completed when the semantic matching degree of the error correction result is greater than or equal to the second preset threshold, and the corrected power text is constructed.

4. The generation method according to claim 1, characterized in that, After segmenting and semantically encoding the text sequence to be processed, a global semantic representation matrix is ​​generated, including: Obtain the text sequence to be processed; Sentences with semantic association are divided into the same semantic block based on the cosine similarity between adjacent sentences in the text sequence, and the maximum length of a single semantic block is limited to less than or equal to a first preset length. Furthermore, semantic content with a second preset length is set to overlap between adjacent semantic blocks. The Longformer model is selected, and a hybrid attention mechanism consisting of sliding window local attention and global attention is used. Formula (3) is used to encode attention for each semantic block separately. ，（3） in, The attention encoding output function for semantic blocks, The query vector matrix for semantic blocks. The key vector matrix of the semantic block. The value vector matrix of the semantic block. For normalization function, For attention mask matrix, This is the scaling factor; Each encoded semantic block is obtained, and the encoded features of each encoded semantic block are extracted and represented. The encoded features of all semantic blocks are globally fused using a cross-semantic block attention fusion layer to generate a unified semantic representation matrix for the entire text.

5. The generation method according to claim 4, characterized in that, Information related to information dimension groups is extracted from the global semantic representation matrix, filtered, and integrated to form a candidate information key set, including: The BERT-CRF model is used to extract core power entities related to information dimension groups and the relationships between core power entities from the global semantic representation matrix. Based on formula (4), the loss value of the BERT-CRF model is optimized to obtain the first candidate key information set. ，（4） in, This is the output of the loss value for the BERT-CRF model. It is the natural logarithm function. This is the global semantic representation matrix. A standard label sequence for core power entities. Given a global semantic representation matrix, this represents the conditional probability that the BERT-CRF model outputs the correct standard label sequence for core electricity entities. The sentence vector of each sentence in the power text is extracted from the global semantic representation matrix, and the weighted TextRank algorithm is used to obtain the weight score of each sentence according to formula (5). ，（5） in, The weight score for the sentence. The base score of a sentence is obtained iteratively based on the TextRank algorithm. The decision relevance coefficient. This represents the business importance coefficient. All sentence vectors are sorted from high to low according to their weight scores. The first preset number of sentence vectors with the highest scores are selected. A second candidate information key set is generated based on the sentence corresponding to each sentence vector. The first and second candidate information key sets are deduplicated and semantically merged to generate an initial candidate information key set. At the same time, based on the principles of global semantic consistency and decision target matching, the initial candidate information key set is processed to generate a strongly relevant candidate information key set.

6. The generation method according to claim 1, characterized in that, The information key set is structurally standardized, input into a pre-defined power auxiliary decision-making system, and the corresponding auxiliary decision-making report is output, including: Obtain the optimized key information set and standardize it using the extracted dimensions from the information dimension group; Construct a power business rules database, set mapping rules between key information sets and decision recommendations, and form a power auxiliary decision-making system; The standardized information key set is input into the power auxiliary decision-making system, and the system outputs a report on the auxiliary decision-making based on the information key set.

7. The generation method according to claim 6, characterized in that, Construct a power business rules database, set mapping rules between key information sets and decision recommendations, and form a power auxiliary decision-making system, including: A rule database that supports visualization is constructed. The rule database adopts a five-layer structured design of "decision scenario - triggering condition - decision action - rule priority - constraint". The triggering conditions are used as the key information set, and the decision actions are used as decision suggestions. A mapping relationship between the key information set and the decision suggestions is constructed. Determine whether the data in the key information set meets the triggering conditions of a specific rule; When the data in the key information set meets the triggering conditions of a specific rule, the decision action of the specific rule with the highest priority is selected as the final decision recommendation based on the rule priority. Version management is implemented for the editing process of the rule database, multiple editing operation modes are set, and the data pairs of key information sets and output decisions generated each time are recorded. These are then converted into standardized rules, and the rule database feedback optimization is completed to form a power auxiliary decision-making system.

Images