A power information index extraction method, a computer device and a readable storage medium

Abstract

The present application belongs to the cross field of electric power artificial intelligence and smart grid big data analysis, and discloses a power intelligence index extraction method, a computer device and a readable storage medium. The event keyword extraction is performed on the power intelligence text to obtain an event anchor keyword set. Semantic retrieval is performed in a historical successful extraction case vector library to recall the historical successful extraction cases similar to the current power intelligence text in event type and business scene. The event type characteristics of the recalled cases are used to match and determine a target extraction template from a preset differentiated extraction template. The power intelligence text, the recalled historical successful extraction cases and the target extraction template are input into a large language model fine-tuned in the electric power field to drive the large language model to perform structured index extraction according to the field constraints defined by the target extraction template, and output the structured power index data conforming to the template field structure. The present application overcomes the problems of low artificial efficiency, incomplete coverage and auxiliary tool understanding deviation.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of power artificial intelligence and smart grid big data analysis, specifically involving a method for extracting power intelligence indicators, computer equipment, and a readable storage medium. Background Technology

[0002] Power system load forecasting and grid planning require extracting key indicators from massive amounts of unstructured electricity consumption data. This type of data is widely distributed in texts such as news reports, official announcements, industry research reports, and meteorological records, containing dynamic information about the impact of rule changes, industry events, and weather anomalies on regional electricity load. The ability to quickly and accurately extract structured indicators from these texts and transform event information into quantifiable load impact assessment criteria is a crucial foundation for intelligent decision-making in the power system.

[0003] Currently, electricity demand intelligence processing mainly relies on professionals manually screening unstructured text and extracting key information. Taking the electricity demand analysis of the new energy vehicle industry chain as an example, analysts need to identify reports related to capacity expansion from a large amount of industry news and corporate announcements, manually record elements such as the time, location, and capacity scale of the events, and combine experience to assess their impact on regional electricity load. In some scenarios, rule-based keyword matching tools are used for preliminary filtering.

[0004] The above-mentioned solutions have the following limitations in practical applications. In terms of processing efficiency, manually filtering and extracting complete information for a single complex event takes several weeks, making it difficult to match the speed of intelligence updates during periods of frequent rule changes or extreme weather seasons. Regarding information coverage, manual analysis typically focuses only on explicit fields, easily overlooking implicit relationships between events. For example, when extracting new energy vehicle production capacity data, it may ignore the synchronous expansion information of supporting enterprises in the industrial chain, or fail to establish a correlation between capacity release and the seasonal characteristics of regional load. In terms of quantitative analysis, existing solutions are mostly limited to qualitative judgments, making it difficult to establish a numerical mapping relationship between events and electricity demand, resulting in dispatch departments being unable to obtain accurate load forecasts. Furthermore, auxiliary tools based on general rules have comprehension biases when dealing with texts dense with power industry terminology and tightly coupled semantics, leading to a significant decrease in extraction accuracy when facing new types of events, still requiring substantial manual verification. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method for extracting power intelligence indicators, a computer device, and a readable storage medium. Its purpose is to quickly and accurately extract structured indicators from massive amounts of unstructured power consumption intelligence text, transforming event information into quantifiable load impact assessment criteria, and overcoming the problems of low efficiency in manual screening and processing, incomplete information coverage, difficulty in quantitative analysis, and misunderstanding and reduced extraction accuracy caused by rule-based auxiliary tools.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: According to a first aspect of the present invention, a method for extracting power information indicators is provided, comprising: Obtain the power intelligence text to be processed; The power information text is subjected to event keyword extraction to obtain a set of event anchor keywords; Based on the event anchor keyword set, semantic retrieval is performed in the pre-built historical successful extraction case vector library to recall at least one historical successful extraction case that is similar to the current power intelligence text in terms of event type and business scenario. Each historical successful extraction case contains corresponding historical original text and historical structured extraction results. Based on the event type characteristics of the historical successful extraction cases, the target extraction template is determined by matching from a plurality of preset differentiated extraction templates. The plurality of differentiated extraction templates correspond to different event types, and each extraction template contains a preset set of fields to be extracted. The power intelligence text, the recalled historical successful extraction cases, and the target extraction template are input into a large language model fine-tuned with small samples in the power field. The large language model is driven to extract structured indicators from the power intelligence text according to the field constraints defined by the target extraction template, under the demonstration guidance of the historical successful extraction cases, and output structured power indicator data that conforms to the field structure of the target extraction template.

[0007] In one possible implementation of the first aspect, the large language model fine-tuned with small samples from the power sector is obtained in the following manner: A small sample labeled dataset in the power sector is obtained. The small sample labeled dataset in the power sector contains multiple labeled samples. Each labeled sample consists of a power intelligence text fragment, an extraction instruction, and the structured extraction result corresponding to the power intelligence text fragment. A parameter-efficient fine-tuning strategy is adopted to insert trainable side-path adaptation matrices into specified network layers of the large language model while freezing the original parameters of the large language model. The large language model after inserting the bypass adaptation matrix is ​​incrementally trained using the small sample labeled dataset in the power field. Only the weight parameters of the bypass adaptation matrix are updated to obtain the large language model fine-tuned by the small sample in the power field.

[0008] In one possible implementation of the first aspect, the parameter efficient fine-tuning strategy is implemented using a low-rank adaptation algorithm, and the number of parameters in the trainable side-path adaptation matrix is ​​less than 1% of the original total number of parameters in the large language model.

[0009] In one possible implementation of the first aspect, the step of performing semantic retrieval based on the event anchor keyword set in a pre-built historical successful extraction case vector library to recall at least one historical successful extraction case similar to the current power intelligence text in terms of event type and business scenario specifically includes: The event anchor keyword set is semantically expanded to obtain an enhanced query vector containing synonyms and related terms in the power field; Calculate the semantic similarity between the enhanced query vector and each candidate case vector in the historical successfully extracted case vector library; Based on the semantic similarity, the cases are sorted from high to low, and a predetermined number of candidate cases with the highest ranking are selected as historically successfully extracted cases for recall.

[0010] In one possible implementation of the first aspect, the step of inputting the power intelligence text, the recalled historical successful extraction cases, and the target extraction template into a large language model fine-tuned with small samples in the power field specifically includes: The original historical text and its corresponding structured extraction results from the recalled historical successful extraction cases are used to construct a small sample demonstration example. The set of fields to be extracted in the target extraction template is converted into structured output format constraints; The task instructions, the small sample demonstration examples, the structured output format constraints, and the power intelligence text are concatenated into a composite prompt instruction, which is then input into the large language model that has been fine-tuned with small samples from the power field.

[0011] In one possible implementation of the first aspect, the plurality of differential extraction templates include at least: The industry-related event template, wherein the set of fields to be extracted from the industry-related event template includes at least: project identifier, production time, spatial location, production capacity and estimated increase in electricity load; The rule-based event template, wherein the set of fields to be extracted from the rule-based event template includes at least: issuing organization, effective time, applicable geographical scope, associated industry category and expected impact period; The meteorological event template includes at least the following fields to be extracted: meteorological event type, affected period, affected area, preset meteorological physical quantity values, and historical year-on-year change data of electricity load during the same period.

[0012] In one possible implementation of the first aspect, after outputting structured power index data that conforms to the target extraction template field structure, the method further includes: Based on the entity information and relationships between entities contained in the structured power index data, a preset dynamic knowledge graph is updated. The update includes: mapping the identified entities to graph nodes, mapping the relationships between the entities to graph edges, and adjusting the corresponding node attributes and edge weights according to the quantified values ​​in the structured power index data.

[0013] In one possible implementation of the first aspect, after the dynamic knowledge graph update is completed, it further includes: Based on the updated dynamic knowledge graph, feature aggregation analysis is performed on preset geographical regions or industry types to generate electricity consumption profiles that include dynamic feature tags and associated quantitative indicators. The dynamic feature tags are used to characterize the power load sensitivity and development trend of the corresponding region or industry.

[0014] According to a second aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the aforementioned method for extracting power information indicators.

[0015] According to a third aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned method for extracting power information indicators.

[0016] According to a fourth aspect of the present invention, a computer program product is provided that, when executed by a processor, implements the aforementioned method for extracting power information indicators.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: This invention provides a method for extracting power intelligence indicators. It acquires power intelligence text and extracts event anchor keywords. Semantic retrieval is performed in a historical successful extraction case vector library to recall similar cases. Differential extraction templates are matched based on the event type characteristics of the recalled cases. The power intelligence text, recalled cases, and target templates are input into a large language model fine-tuned with small samples in the power field for structured indicator extraction. The final output is structured power indicator data conforming to the template field structure. By combining event anchor keyword extraction with semantic retrieval from the historical successful extraction case vector library, when faced with new power intelligence text, there is no need to traverse the entire knowledge base. Instead, the event anchor keywords are used as query anchors to quickly locate the historical successful extraction cases most similar to the current text in terms of event type and business scenario within the vector space. This mechanism compresses the information processing time window from several weeks in the traditional manual mode to hours, significantly improving the response speed and analysis timeliness of power intelligence. This enables power dispatching departments to obtain timely judgment basis during periods of frequent rule changes or peak seasons for extreme weather.

[0018] By introducing historically successful extraction cases as demonstrative references into the reasoning process of the large language model, the model no longer relies on open generation when faced with new power information text. Instead, it uses analogical reasoning based on the extraction paradigm of similar historically successful extraction cases. The corresponding historical original texts and historical structured extraction results in the historically successful extraction cases provide clear teaching demonstrations for the large language model, enabling it to complete the current task according to the validated extraction logic under the field constraints of the target extraction template. This constrained analogical reasoning mechanism effectively suppresses erroneous extraction and format deviations in professional contexts, resulting in consistent and standardized extraction results.

[0019] By matching differentiated extraction templates to the event type characteristics of recalled cases, information extraction can be performed using field frameworks adapted to different event types. The differentiated extraction templates correspond to different event types, and each template contains a preset set of fields to be extracted, limiting the specific fields that the extraction task needs to extract from the power intelligence text. This transforms information extraction from open-ended text understanding into a data-driven fill-in-the-blank reasoning task. When facing different types of events such as industry expansion, rule releases, and weather warnings, accurate extraction can be completed according to the field logic of the corresponding template, avoiding the problems of field omissions or format confusion that arise when using general extraction methods for heterogeneous events.

[0020] By performing structured indicator extraction using a large language model fine-tuned with small samples in the power sector, the model gains a deeper understanding of power sector terminology, implicit relationships between indicators, and business context while retaining its ability to understand general language. When faced with technically dense and semantically tightly coupled expressions in power intelligence texts, the large language model can accurately identify indicator boundaries and transform natural language descriptions into standardized expressions that conform to field type requirements. The fine-tuned large language model shows a significant improvement in the accuracy of power indicator boundary identification compared to general models, while effectively suppressing illusory outputs in technical contexts. By outputting structured power indicator data that conforms to the target extraction template field structure, power intelligence, originally existing in unstructured natural language form, is transformed into structured data conforming to a preset field structure. Each field contains indicator values ​​or text descriptions extracted from the power intelligence text and standardized, thus transforming discrete text information into standardized data. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the specific embodiments of the present invention, the drawings used in the description of the specific embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a method for extracting power intelligence indicators. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] like Figure 1 As shown, this invention provides a method for extracting power information indicators, specifically including the following steps: S1. Obtain the power intelligence text to be processed.

[0025] Specifically, power intelligence text refers to unstructured text data containing rule information, industry dynamics, and natural environmental factors related to changes in power load. Sources of power intelligence text include news reports, announcements from publishing agencies, industry research reports, publicly disclosed corporate information, and weather forecast records. Acquisition methods can include proactively collecting publicly available information from the internet or receiving specified documents pushed by business systems. The acquired power intelligence text is typically in paragraph form using natural language descriptions, such as a news report about the commissioning of a new energy vehicle factory in a certain location.

[0026] S2. Extract event keywords from the power information text to obtain a set of event anchor keywords.

[0027] Specifically, event anchor keywords refer to a number of words that can characterize the core event type and theme of power intelligence text. For example, from the power intelligence text about a car company investing in and building a production base for 100,000 new energy vehicles per year in Suzhou, Jiangsu Province, and starting production next month, event anchor keywords such as "starting production," "production base," "annual production," and "new energy vehicles" can be extracted. The extraction algorithm can use a combination of rule-based and statistical methods to quickly extract the most representative core words from the power intelligence text, providing focused query anchors for semantic retrieval.

[0028] S3. Based on the event anchor keyword set, perform semantic retrieval in the pre-built historical successful extraction case vector library to recall at least one historical successful extraction case that is similar to the current power intelligence text in terms of event type and business scenario. Each historical successful extraction case contains corresponding historical original text and historical structured extraction results.

[0029] Specifically, the historical successful extraction case vector library is a pre-built dedicated vector database that transforms unstructured power intelligence text into data points in a high-dimensional vector space through an embedding model. The library stores a large number of vectorized representations of historical successful extraction cases. Each historical successful extraction case contains a pair of corresponding data: the original historical text and a historical structured extraction result that has been manually or systematically verified. For example, a historical successful extraction case might contain the original historical text "A battery company is expanding its production capacity by 5 gigawatt-hours in Ningde," and the corresponding historical structured extraction result might show the event type as "capacity expansion," the subject as "a battery company," the location as "Ningde," and the production capacity as "5 gigawatt-hours per year." The goal of semantic retrieval is to recall at least one historical successful extraction case from the historical successful extraction case vector library that is similar to the current power intelligence text in terms of event type and business scenario. The recalled historical successful extraction cases should be isomorphic to the current power intelligence text in terms of event attributes. For example, if the current power intelligence text involves a factory commissioning event, then the recalled historical successful extraction cases should also primarily be industry expansion events. This process ensures that the system can find historical examples of how fields were successfully defined and extracted for such events, providing accurate reference samples for imitation learning.

[0030] S4. Based on the event type characteristics of the historical successful extraction cases, a target extraction template is determined by matching from a plurality of preset differentiated extraction templates. The plurality of differentiated extraction templates correspond to different event types, and each extraction template contains a preset set of fields to be extracted.

[0031] Specifically, the differentiated extraction template is a pre-designed set of structured fields for different types of power events. Each differentiated extraction template corresponds to one event type and contains a set of preset fields to be extracted. Based on the common characteristics of recalled historical successfully extracted cases, the system intelligently determines and locks the differentiated extraction template to be used for the current task. If the recalled historical successfully extracted cases are all factory expansion events, the industry event template is activated; if the recalled historical successfully extracted cases are mostly rule release events, the rule event template is activated; if the recalled historical successfully extracted cases involve weather warnings, the weather event template is activated. Each differentiated extraction template limits the specific field items that the extraction task needs to extract from the power intelligence text, transforming open-ended text understanding into a data-driven fill-in-the-blank reasoning task.

[0032] S5. Input the power intelligence text, the recalled historical successful extraction cases, and the target extraction template into a large language model fine-tuned with small samples in the power field. Drive the large language model to extract structured indicators from the power intelligence text according to the field constraints defined by the target extraction template, under the demonstration guidance of the historical successful extraction cases, and output structured power indicator data that conforms to the field structure of the target extraction template.

[0033] Specifically, the large language model fine-tuned for small-sample testing in the power sector has been refined to possess a deep understanding of power terminology, indicator logic, and business context. Upon receiving three inputs—power intelligence text, recalled historical successful extraction cases, and the target extraction template—the large language model performs a structured indicator extraction task on the power intelligence text, using the recalled historical successful extraction cases as examples and under the field constraints of the target extraction template. The extraction process is essentially a constrained semantic parsing and information filling process. Based on the field list defined in the target extraction template, the large language model locates text fragments in the power intelligence text that semantically match each field, transforms them into standardized expressions that conform to the field type requirements, and fills them into the corresponding field positions.

[0034] The output structured power index data is in key-value pair format, with each structured power index data corresponding to an event instance. The field names are consistent with the set of fields to be extracted defined in the target extraction template, and the field content is the index value or text description extracted from the power intelligence text and normalized. For example, the production capacity field in the industry event template outputs 100,000 vehicles per year, and the estimated electricity load increment field outputs 5%.

[0035] This implementation method introduces historically successful extraction cases as demonstration references into the reasoning process of the large language model. When faced with new power information text, the large language model no longer relies on open generation but instead uses analogical reasoning based on the extraction paradigm of similar historically successful extraction cases. This effectively constrains the standardization of the output format and significantly reduces erroneous extraction and format deviations in professional contexts. Information extraction efficiency is reduced from several weeks in traditional manual methods to hours, an improvement of more than tenfold. Simultaneously, the combination of event anchor keyword set extraction and semantic retrieval enables the system to quickly identify the historically successful extraction cases most relevant to the current task, avoiding the computational overhead of full database traversal and significantly improving processing efficiency while ensuring extraction accuracy.

[0036] In one possible implementation, the large language model fine-tuned with small samples from the power sector is obtained in the following manner: First, a small sample labeled dataset in the power sector is obtained. The small sample labeled dataset in the power sector contains multiple labeled samples. Each labeled sample consists of a power intelligence text fragment, an extraction instruction, and the structured extraction result corresponding to the power intelligence text fragment.

[0037] Specifically, the small-sample annotation dataset for the power sector was constructed by manually selecting typical fragments from massive industry reports, power grid regulations, and authoritative information, and contains multiple annotated samples. Each annotated sample consists of three parts: a power intelligence text fragment, an extraction instruction, and the corresponding structured extraction result. For example, the construction of the small-sample annotation dataset for the power sector emphasizes the diversity of business scenarios, covering key situations such as new energy consumption analysis, extreme weather emergency response, and regional industrial electricity consumption trend analysis, injecting rich domain prior knowledge into the large language model. Because each annotated sample requires manual annotation by professionals, the overall size of the small-sample annotation dataset for the power sector is relatively small.

[0038] Next, a parameter-efficient fine-tuning strategy is adopted to insert trainable side-path adaptation matrices into specified network layers of the large language model while freezing the original parameters of the large language model.

[0039] It should be understood that the core idea of ​​the efficient parameter fine-tuning strategy is to achieve domain adaptation by inserting a small number of trainable parameters into the large language model without changing its original parameters. Specifically, this involves freezing all the original parameters of the base large language model, keeping them unchanged during forward propagation; and inserting a trainable side-path adaptation matrix into a specified network layer of the large language model. The initial values ​​of the side-path adaptation matrix can be randomly initialized or zero-initialized.

[0040] In this embodiment, the basic large language model used is Qwen2.5-VL-7B-Instruct, which has strong general semantic understanding and multimodal processing capabilities.

[0041] Finally, the large language model after inserting the bypass adaptation matrix is ​​incrementally trained using the small sample labeled dataset in the power field. Only the weight parameters of the bypass adaptation matrix are updated to obtain the large language model fine-tuned by small samples in the power field.

[0042] Specifically, during training, each labeled sample from the small-sample labeled dataset in the power sector is sequentially or in batches input into the large language model inserted into the bypass adaptation matrix. For each labeled sample, the large language model receives power intelligence text fragments and extraction instructions, calculates the predicted structured extraction result through forward propagation, performs loss calculation on this prediction result and the structured extraction result labeled in the labeled sample, and updates only the weight parameters of the bypass adaptation matrix through the backpropagation algorithm. Since the original parameters of the large language model are frozen, the scope of gradient calculation and parameter updates is strictly limited to within the bypass adaptation matrix. After multiple rounds of iterative training, the weight parameters of the bypass adaptation matrix gradually converge, and the large language model gradually internalizes the domain paradigm and structured expression logic of indicator extraction, significantly enhancing its sensitivity to power terminology and basic extraction logic. After training is completed, the frozen original parameters of the large language model are merged and saved with the trained bypass adaptation matrix, thus obtaining the large language model fine-tuned by small samples in the power sector.

[0043] This implementation employs a highly efficient parameter fine-tuning strategy for domain adaptation, requiring only a minimal update to the general-purpose language model to achieve accurate semantic understanding within the power industry domain. Quantitative evaluation shows that the language model fine-tuned with small samples in the power industry domain improves the accuracy of power indicator boundary recognition by over 15% compared to the baseline, while fully retaining its general-purpose language understanding capabilities, achieving a harmonious balance between domain specialization and generalization. Compared to the full-scale fine-tuning approach, this implementation reduces computational resource consumption and training time costs, while significantly decreasing the dependence on labeled data, effectively suppressing the risk of overfitting under small sample conditions.

[0044] Preferably, the efficient parameter fine-tuning strategy is implemented using a low-rank adaptation algorithm, and the number of parameters in the trainable side-path adaptation matrix is ​​less than 1% of the original total number of parameters in the large language model.

[0045] In detail, the basic principle of the low-rank adaptation algorithm is as follows: A trainable side-path adaptation matrix is ​​embedded in the transformer layer. This side-path adaptation matrix is ​​decomposed into the product of two low-rank matrices. By constraining the rank of the matrices, the number of trainable parameters is significantly reduced. For a selected target network layer in a large language model, let the dimension of its original weight matrix be d x k. The low-rank adaptation algorithm inserts two matrices A and B into the side path of this layer, where the dimension of A is d x r and the dimension of B is r x k, where r is a preset low-rank value. During forward propagation, the output of the side path is the result of multiplying the input vector by A and B sequentially. This result is added to the output of the original network layer and used as the input to the next layer. The number of trainable parameters is determined by the sum of the parameter values ​​of A and B. By selecting a low-rank value r that is much smaller than d and k, the number of trainable parameters can be reduced by an order of magnitude compared to the original total number of parameters in the large language model.

[0046] In the power sector small-sample fine-tuning scenario of this embodiment, the selected low-rank value ensures that the number of parameters in the trainable side-fitting matrix is ​​less than 1% of the original total number of parameters in the large language model. Taking a basic large language model with seven billion parameters as an example, when using the low-rank adaptation algorithm for power sector small-sample fine-tuning, the actual number of parameters that need to be updated can be controlled to tens of millions, i.e., less than 1% of the total number of parameters. Because the low-rank adaptation algorithm imposes a low-rank constraint on the side-fitting matrix, the large language model can learn incremental knowledge in the power sector in a more compact parameter form during training, while effectively suppressing the risk of overfitting.

[0047] In one possible implementation, based on the event anchor keyword set, semantic retrieval is performed in a pre-built historical successful extraction case vector library to recall at least one historical successful extraction case that is similar to the current power intelligence text in terms of event type and business scenario. The specific process is as follows: 1) Semantically expand the set of event anchor keywords to obtain an enhanced query vector containing synonyms and related terms in the power industry.

[0048] The purpose of semantic expansion is to compensate for the limitations of single event anchor keywords in terms of semantic coverage. In power information texts, the same concept may appear in multiple forms; for example, capacity expansion may be described as capacity improvement, capacity increase, new capacity, capacity ramp-up, etc. Directly performing semantic retrieval using the original event anchor keywords may miss historically successful extraction cases using different wording. To overcome this problem, the initial set of event anchor keywords is semantically enhanced and expanded.

[0049] The specific method of semantic expansion is as follows: Using a pre-trained word vector model or synonym knowledge base in the power industry domain, synonyms are retrieved for each event anchor keyword in the event anchor keyword set. Simultaneously, related power industry terms that are semantically closely related to the event anchor keyword are supplemented based on the power industry terminology association graph. The word vectors of the original event anchor keywords and the expanded terms are then weighted and fused to form an enhanced query vector.

[0050] 2) Calculate the semantic similarity between the enhanced query vector and each candidate case vector in the historical successfully extracted case vector library.

[0051] Specifically, each candidate case in the historical successfully extracted case vector library is pre-converted into a fixed-dimensional vector representation through an embedding model. The embedding model and the word vector model that generates the enhanced query vector use the same semantic space or undergo spatial alignment processing.

[0052] For example, semantic similarity can be calculated using metrics such as cosine similarity, Euclidean distance, or dot product. Similarity is determined by calculating the vector distance between the current event anchor keyword and the structured historical successfully extracted case vectors in the historical successfully extracted case vector library.

[0053] 3) Sort the cases from high to low according to the semantic similarity, and select a preset number of candidate cases with the highest ranking as historical successfully extracted cases for recall.

[0054] In other words, the semantic retrieval target here is not ordinary background knowledge, but rather a targeted recall of historically successfully extracted cases, selecting several historically successfully extracted cases that are most similar in terms of event type, business scenario, and data characteristics. The preset number can be flexibly set according to the actual application scenario. If there are cases with low similarity among the top-ranked candidate cases, the number of cases recalled can be appropriately reduced or a manual intervention mechanism can be triggered.

[0055] This implementation expands the set of event anchor keywords into an enhanced query vector that includes synonyms and power industry-related terms through semantic extension. This effectively overcomes the recall omission problem caused by differences in expression when retrieving single event anchor keywords. The introduction of power industry-related terms further enhances the domain semantic density of the enhanced query vector, making the recall results more closely aligned with actual needs in terms of event type and business scenario. This process ensures that the system can find how fields were successfully defined and extracted for such events in the past, providing accurate benchmark samples for imitation learning, thereby transforming an open-ended text understanding problem into a well-founded analogy retrieval problem.

[0056] In one possible implementation, the step of inputting the power intelligence text, the recalled historical successful extraction cases, and the target extraction template into a large language model fine-tuned with small samples in the power field specifically includes the following process: First, the original historical text and its corresponding structured extraction results from the successfully extracted historical cases are used to construct a small sample demonstration example.

[0057] In detail, each historical successful extraction case can be transformed into a small sample demonstration example unit. The system transforms the historical successful extraction cases recalled by semantic retrieval—that is, the historical original text and the corresponding historical structured extraction results—into dynamic small sample demonstration examples. The construction format of the small sample demonstration example is as follows: the historical original text is used as the input example part, and the corresponding historical structured extraction result is used as the output example part, forming a set of input-output correspondence demonstrations. When multiple historical successful extraction cases are recalled, the small sample demonstration examples corresponding to each historical successful extraction case are arranged in descending order of similarity to form a complete demonstration example sequence. At this time, the recalled historical successful extraction cases are no longer just background knowledge, but serve as explicit teaching demonstrations, showing the large language model, which has been fine-tuned with small samples in the power field, how to accurately map target data from unstructured power intelligence text under the constraints of the target extraction template.

[0058] Next, the set of fields to be extracted in the target extraction template is converted into structured output format constraints.

[0059] Specifically, the target extraction template defines a list of fields that need to be output for this extraction task. The purpose of the structured output format constraints is to clarify to the large language model, which has been fine-tuned with small samples in the power field, which fields should be included in the output, the data type of each field, and the expression specifications.

[0060] Finally, the task instructions, the small sample demonstration examples, the structured output format constraints, and the power intelligence text are concatenated into a composite prompt instruction, which is then input into the large language model that has been fine-tuned with small samples in the power field.

[0061] In other words, the system performs a context-based isomorphic fusion operation of prompt words. It should be noted that the task instruction is a general description of what operation to perform on the large language model fine-tuned with small samples from the power sector. The concatenation order is as follows: the task instruction is placed first, followed by the sequence of small sample demonstration examples, structured output format constraints, and finally the power information text to be processed. The concatenated composite prompt instruction serves as the input for a single inference, and is input into the large language model fine-tuned with small samples from the power sector. Upon receiving this complete composite prompt instruction, the large language model fine-tuned with small samples from the power sector can, under the analogy guidance of the small sample demonstration examples, strictly follow the field logic defined by the target extraction template to perform deep semantic analysis and fill-in-the-blank extraction on the power information text.

[0062] This implementation constructs historical successful extraction cases as small-sample demonstration examples and directly embeds them into the input context. This allows the large language model, fine-tuned with small samples in the power sector, to understand the specific requirements and output specifications of the current extraction task through analogical learning during the inference process. The large language model fine-tuned with small samples in the power sector, as a capability base, already possesses sensitive recognition and basic extraction logic for power terminology. When semantic retrieval recalls historical successful extraction cases, these cases are dynamically injected into the inference process as contextual learning examples. By analogy with the format specifications and semantic structure of historical successful extraction cases and integrating its internalized domain knowledge, the large language model fine-tuned with small samples in the power sector accurately completes the indicator extraction task for the current event. This dual-constraint framework, combining small-sample fine-tuning in the power sector with real-time historical successful extraction cases, fundamentally overcomes the problems of insufficient extraction accuracy and output format deviation in subdivided domains for general models.

[0063] In one possible implementation, the plurality of differentiated extraction templates include at least industry-related event templates, rule-related event templates, and meteorological event templates.

[0064] The set of fields to be extracted for industry-related event templates must include at least: project identifier, commissioning time, spatial location, production capacity, and estimated electricity load increment. Industry-related event templates are applicable to industry-related events such as the evolution of the new energy vehicle industry chain, industrial project construction, and production capacity ramp-up. The project identifier uniquely identifies the construction project or production capacity project involved; the commissioning time indicates the time node for project commissioning or capacity release; the spatial location indicates the geographical region or industrial park where the project is located; the production capacity indicates the annual output or designed capacity after the project reaches full production; and the estimated electricity load increment indicates the increase in electricity load after the project commissions, estimated based on the production capacity and historical electricity consumption data. Industry-related event templates construct accurate quantitative analysis benchmarks by forcibly constraining a large language model fine-tuned with small samples in the power sector to locate numerical characteristics highly correlated with production capacity release and electricity load increments.

[0065] The set of fields to be extracted for rule-based event templates must include at least: issuing organization, effective date, applicable geographical scope, related industry category, and expected impact period. Rule-based event templates are applicable to rule-based events such as announcements by issuing organizations, electricity price adjustments, and changes in industry access standards. The issuing organization represents the entity that issued the rule; the effective date represents the start time of the rule's official implementation; the applicable geographical scope represents the geographical area or industrial park covered by the rule; the related industry category represents the industry sectors affected by the rule; and the expected impact period represents the time span over which the rule will affect electricity demand. Rule-based event templates achieve semantic understanding beyond surface text matching through deep semantic analysis of the rule's transmission path.

[0066] The set of fields to be extracted for meteorological event templates must include at least: meteorological event type, affected period, affected area, preset meteorological physical quantity values, and historical year-on-year change data for electricity load. Meteorological event templates are applicable to meteorological events such as typhoons, extreme high temperatures, and cold wave warnings. The meteorological event type indicates the category name of the meteorological event; the affected period indicates the start and end times of the meteorological event's impact on the region; the affected area indicates the geographical area affected by the meteorological event; the preset meteorological physical quantity values ​​represent the measured or forecasted values ​​of key meteorological parameters related to the event; and the historical year-on-year change data for electricity load represents a reference value for the magnitude of changes in regional electricity load under similar historical meteorological conditions. Meteorological event templates achieve a direct correlation between meteorological data and power grid load characteristics.

[0067] It should be noted that the aforementioned industry-related event templates, rule-related event templates, and meteorological event templates can be expanded or refined according to actual business needs. Through the construction of these differentiated extraction templates, the system transforms the processing of unstructured power intelligence text into a directed fill-in-the-blank reasoning process based on specific differentiated extraction templates. This effectively guides the attention allocation of the large language model, fine-tuned with small samples in the power field, in complex contexts, reducing interference from redundant information.

[0068] This implementation method designs differentiated extraction templates for different event types, enabling information extraction tasks to match the most suitable field framework based on the event type. Industry-related event templates mandate the identification of quantitative indicators such as production capacity and estimated electricity load increments, achieving a mapping from qualitative analysis to quantitative indicators. Rule-related event templates focus on transmission elements such as the issuing agency and the expected impact cycle, while meteorological event templates associate preset meteorological physical quantity values ​​with historical year-on-year changes in electricity load for the same period.

[0069] In one possible implementation, after outputting structured power index data that conforms to the target extraction template field structure, the following steps are also included: S6. Update the preset dynamic knowledge graph based on the entity information and the relationships between entities contained in the structured power index data. The update includes: mapping the identified entities to graph nodes, mapping the relationships between entities to graph edges, and adjusting the corresponding node attributes and edge weights according to the quantized values ​​in the structured power index data.

[0070] Specifically, structured power index data records event elements extracted from single power intelligence texts in key-value pair format. However, discrete single structured power index data is insufficient to support overall judgment and correlation analysis of regional power consumption trends. Dynamic knowledge graphs are databases that organize domain knowledge in a graph structure, where nodes represent various entity objects, edges represent semantic relationships between entities, and node attributes and edge weights are used to store quantitative features and relationship strength.

[0071] Entity information originates from the entity type field in structured power index data. Examples include company names extracted from the project identifier field of industry-related event templates, region names extracted from the spatial location field, and issuing agency names extracted from the issuing agency field of rule-related event templates. The core nodes of the dynamic knowledge graph include event nodes, subject nodes, indicator nodes, and result nodes. For each identified entity, the dynamic knowledge graph is queried to see if a corresponding node already exists: if it does, the node is located; otherwise, a new node is created and assigned an entity type label.

[0072] The relationships between entities originate from the semantic connections implicit or explicitly expressed in structured electricity index data. For example, in industry-related events, there is a relationship of being located in or investing in a business entity and a location entity; similarly, there is a causal relationship of causing or leading to a change in production capacity and an estimated increase in electricity load. Edges represent the causal and transmission relationships between entities, particularly the transmission path from an event to electricity demand. These relationships are transformed into directed or undirected edges in a dynamic knowledge graph, with edge type labels corresponding to the semantic meaning of the relationships.

[0073] The capacity scale value is written into the capacity attribute field of the corresponding project node, and the estimated electricity load increment value is written into the weight value of the edge connecting the project node and the regional load node. For multiple updates involving the same node or the same edge, the values ​​are updated using a weighted average or incremental overlay method to reflect the cumulative effect of multiple events.

[0074] Unlike static knowledge graphs, the dynamic knowledge graph in this implementation is dynamically updated. As new electricity intelligence text is continuously extracted and updated to the dynamic knowledge graph, the node states and edge weights in the dynamic knowledge graph are adjusted in real time. For example, the dynamic knowledge graph also introduces meteorological factors as environmental context nodes to demonstrate how meteorological factors regulate the transmission process between events and electricity demand.

[0075] This implementation transforms discrete, structured electricity index data into dynamic knowledge assets with a relational network through dynamic knowledge graph updates. Entity mapping and relation mapping weave isolated event information into a traceable and reasonable semantic network, while continuous updates of quantified values ​​enable the dynamic knowledge graph to dynamically reflect the evolving trends of regional electricity consumption. The system not only focuses on single events but also models rule transmission paths, upstream and downstream connections in the industrial chain, and the synergistic effects of meteorological factors, ensuring comprehensive electricity demand analysis.

[0076] In one possible implementation, after the dynamic knowledge graph is updated, the method further includes: based on the updated dynamic knowledge graph, performing feature aggregation analysis on a preset geographical region or industry type to generate an electricity consumption profile result containing dynamic feature tags and associated quantitative indicators, wherein the dynamic feature tags are used to characterize the power load sensitivity characteristics and development trends of the corresponding region or industry.

[0077] It should be understood that the preset geographical area can be any level of area, such as an industrial park or a power grid partition unit, and the preset industry type can be a specific category in the national economic industry classification or a power system-defined electricity consumption category.

[0078] The feature aggregation analysis process involves extracting entity nodes and associated edges falling within a preset geographical region or industry type from the updated dynamic knowledge graph, and then statistically analyzing the attribute values ​​of the nodes and the weights of the edges. Analysis dimensions include: the cumulative frequency of various events within the region, the total scale and distribution density of capacity expansion projects, the coverage intensity of rule influence, and the regional differentiation characteristics of meteorological sensitivity.

[0079] It should be noted that the electricity consumption profile integrates specific correlated quantitative indicators, such as the expected load baseline increase due to new capacity and seasonal fluctuation coefficients. These correlated quantitative indicators are derived from cumulative calculations of specific events. Dynamic feature labels are concise and general descriptions of the electricity consumption characteristics of a region or industry. The dimensions of the electricity consumption profile not only include basic geographical location and industry classification, but also deeply integrate dynamic features extracted from structured electricity indicator data. For example, for a region, the electricity consumption profile will be labeled with dynamic feature labels such as new energy vehicle industry clusters, sensitivity to high-temperature weather, and favorable regulations for capacity expansion. Each dynamic feature label is accompanied by correlated quantitative indicators as data support.

[0080] For example, electricity consumption profiles can be presented to power grid dispatchers and planners in the form of tables, statistical charts, or geographic information layers. Through these profiles, dispatchers can quickly identify regions or industries with high growth potential and optimize power grid resource allocation in advance. As new electricity intelligence texts are continuously extracted and updated to the dynamic knowledge graph, the dynamic feature labels and associated quantitative indicators in the electricity consumption profiles are also dynamically adjusted.

[0081] This implementation further processes the underlying dynamic knowledge graph data into electricity consumption profiles oriented towards business decision-making. Dynamic feature tags lower the understanding threshold of dynamic knowledge graph data in an intuitive semantic label format, while correlation quantification indicators provide verifiable data basis for each dynamic feature tag. This layer-by-layer abstraction process—from power intelligence text to structured power indicator data, then to dynamic knowledge graph correlation, and finally to electricity consumption profile results—enables power system decision-makers to obtain the information they need at different granularities. This invention transforms implicit data correlations into explicit decision-making basis, providing high-dimensional dynamic data for power grid planning, load forecasting, and energy decision-making, supporting the intelligent transformation of the power system from passive response to proactive forecasting.

[0082] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a power information indicator extraction method.

[0083] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be Random Access Memory (RAM) or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the power information indicator extraction method in the above embodiments.

[0084] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0085] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0086] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0087] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0088] This invention also provides a computer program product for executing any of the above-described methods for extracting power information indicators. Since the computer program product provided by this invention belongs to the same inventive concept as the power information indicator extraction method described above, it possesses all the advantages of the aforementioned method. Therefore, the beneficial effects of the computer program product provided by this invention will not be elaborated upon here.

[0089] In this invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0090] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, 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 covered within the scope of protection of the present invention.

Claims

1. A method for extracting power information indicators, characterized in that, include: Obtain the power intelligence text to be processed; The power information text is subjected to event keyword extraction to obtain a set of event anchor keywords; Based on the event anchor keyword set, semantic retrieval is performed in the pre-built historical successful extraction case vector library to recall at least one historical successful extraction case that is similar to the current power intelligence text in terms of event type and business scenario. Each historical successful extraction case contains corresponding historical original text and historical structured extraction results. Based on the event type characteristics of the historical successful extraction cases, the target extraction template is determined by matching from a plurality of preset differentiated extraction templates. The plurality of differentiated extraction templates correspond to different event types, and each extraction template contains a preset set of fields to be extracted. The power intelligence text, the recalled historical successful extraction cases, and the target extraction template are input into a large language model fine-tuned with small samples in the power field. The large language model is driven to extract structured indicators from the power intelligence text according to the field constraints defined by the target extraction template, under the demonstration guidance of the historical successful extraction cases, and output structured power indicator data that conforms to the field structure of the target extraction template.

2. The method for extracting power information indicators according to claim 1, characterized in that, The large language model, fine-tuned with small samples from the power sector, was obtained in the following way: A small sample labeled dataset in the power sector is obtained. The small sample labeled dataset in the power sector contains multiple labeled samples. Each labeled sample consists of a power intelligence text fragment, an extraction instruction, and the structured extraction result corresponding to the power intelligence text fragment. A parameter-efficient fine-tuning strategy is adopted to insert trainable side-path adaptation matrices into specified network layers of the large language model while freezing the original parameters of the large language model. The large language model after inserting the bypass adaptation matrix is ​​incrementally trained using the small sample labeled dataset in the power field. Only the weight parameters of the bypass adaptation matrix are updated to obtain the large language model fine-tuned by the small sample in the power field.

3. The method for extracting power information indicators according to claim 2, characterized in that, The efficient parameter fine-tuning strategy is implemented using a low-rank adaptation algorithm, and the number of parameters in the trainable side-path adaptation matrix is ​​less than 1% of the original total number of parameters in the large language model.

4. The method for extracting power information indicators according to claim 1, characterized in that, The step of performing semantic retrieval based on the event anchor keyword set in a pre-built historical successful extraction case vector library to recall at least one historical successful extraction case similar to the current power intelligence text in terms of event type and business scenario specifically includes: The event anchor keyword set is semantically expanded to obtain an enhanced query vector containing synonyms and related terms in the power industry; Calculate the semantic similarity between the enhanced query vector and each candidate case vector in the historical successfully extracted case vector library; Based on the semantic similarity, the cases are sorted from high to low, and a predetermined number of candidate cases with the highest ranking are selected as historically successfully extracted cases for recall.

5. The method for extracting power information indicators according to claim 1, characterized in that, The process of inputting the power intelligence text, the recalled successful extraction cases, and the target extraction template into a large language model fine-tuned with small samples in the power field specifically includes: The original historical text and its corresponding structured extraction results from the recalled historical successful extraction cases are used to construct a small sample demonstration example. The set of fields to be extracted in the target extraction template is converted into structured output format constraints; The task instructions, the small sample demonstration examples, the structured output format constraints, and the power intelligence text are concatenated into a composite prompt instruction, which is then input into the large language model that has been fine-tuned with small samples from the power field.

6. The method for extracting power information indicators according to claim 1, characterized in that, The plurality of differential extraction templates include at least: The industry-related event template, wherein the set of fields to be extracted from the industry-related event template includes at least: project identifier, production time, spatial location, production capacity and estimated increase in electricity load; The rule-based event template, wherein the set of fields to be extracted from the rule-based event template includes at least: issuing organization, effective time, applicable geographical scope, associated industry category and expected impact period; The meteorological event template includes at least the following fields to be extracted: meteorological event type, affected period, affected area, preset meteorological physical quantity values, and historical year-on-year change data of electricity load during the same period.

7. The method for extracting power information indicators according to claim 1, characterized in that, After outputting structured power index data that conforms to the target extraction template field structure, the process also includes: Based on the entity information and relationships between entities contained in the structured power index data, a preset dynamic knowledge graph is updated. The update includes: mapping the identified entities to graph nodes, mapping the relationships between the entities to graph edges, and adjusting the corresponding node attributes and edge weights according to the quantified values ​​in the structured power index data.

8. The method for extracting power information indicators according to claim 7, characterized in that, After the dynamic knowledge graph is updated, it also includes: Based on the updated dynamic knowledge graph, feature aggregation analysis is performed on preset geographical regions or industry types to generate electricity consumption profiles that include dynamic feature tags and associated quantitative indicators. The dynamic feature tags are used to characterize the power load sensitivity and development trend of the corresponding region or industry.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a power information indicator extraction method as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a method for extracting power information indicators as described in any one of claims 1 to 8.

Images