Power grid standard difference analysis method and device, equipment and medium

Abstract

This application provides a method, apparatus, device, and medium for analyzing differences in power grid standards. The method includes: acquiring standard text comparison data; mining the standard text comparison data based on a pre-trained mining model to obtain multiple standard text pairs and the difference patterns of each standard text pair; performing difference analysis using a pre-trained inference model based on the multiple standard text pairs and the difference patterns of each standard text pair to obtain difference inference results; the difference inference results include the difference types of different standard text pairs; the inference model is an initial model trained on a Bi-LSTM architecture using a difference-sensitive knowledge distiller. This method aims to improve the efficiency of difference analysis for the same standard from different countries.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method, apparatus, device and medium for analyzing differences in power grid standards. Background Technology

[0002] With the development of the Global Energy Interconnection, cross-border power grid interconnection projects are increasing, and the compatibility of power grid standards in different countries has become a key factor restricting interconnection. Different countries may adopt different technical standards in the power grid field, or modify the same international standard for local use, resulting in various forms of differences between standards.

[0003] Traditional standard difference analysis mainly relies on manual comparison, which is labor-intensive and prone to oversights.

[0004] With the application of deep learning algorithms in various fields, there is an urgent need for an analysis technique based on intelligent algorithms to improve the efficiency of difference analysis. Summary of the Invention

[0005] This application provides a method, apparatus, equipment, and medium for analyzing the differences in power grid standards, in order to improve the efficiency of analyzing the differences in the same standards across different countries.

[0006] In a first aspect, embodiments of this application provide a method for analyzing differences in power grid standards, including:

[0007] Obtain standard text comparison data, which includes the standard text of the same power grid standard in different countries;

[0008] The standard text comparison data is mined based on a pre-trained mining model to obtain multiple standard text pairs and the difference patterns of each standard text pair. The difference patterns include literal differences and semantic differences between the standard text pairs. The mining model is built based on the Transformer algorithm.

[0009] Based on the multiple standard text pairs and the difference patterns of each standard text pair, a pre-trained inference model is used to perform difference analysis and obtain difference inference results; the difference inference results include the difference types of different standard text pairs; the inference model is obtained by pre-training an initial model of the Bi-LSTM architecture using a difference-sensitive knowledge distiller.

[0010] In one possible implementation, the method further includes:

[0011] A difference-sensitive knowledge distillation apparatus is used to process the pre-acquired sample set to obtain the difference-sensitive regions in standard text pairs corresponding to different difference patterns; the sample set includes multiple standard text pairs and difference patterns pre-annotated to the standard text pairs;

[0012] Distillation weight data of the pre-trained model of the difference-sensitive knowledge distiller, and difference-sensitive regions in standard text pairs corresponding to different difference patterns, are used to distill the initial model of the pre-constructed Bi-LSTM architecture to obtain the inference model; the distillation weight data of the pre-trained model includes the importance weight of each difference-sensitive region.

[0013] The inference model is used to identify the difference-sensitive regions of the standard text pairs based on the input standard text pairs and the difference patterns of the standard text pairs, and output the difference inference results after performing difference inference based on the text features corresponding to the difference-sensitive regions.

[0014] In one possible implementation, the method further includes:

[0015] The interactive interface displays the multiple standard text pairs and the difference inference results for each standard text pair;

[0016] Receive correction data input by the user on the interactive interface, the correction data including the user's correction content for the difference inference result of any standard text pair;

[0017] The inference model is updated based on the corrected data.

[0018] In one possible implementation, updating the inference model based on the difference inference result and the correction data includes:

[0019] The difference-sensitive knowledge distiller is optimized based on the corrected data to obtain the updated weights of the difference-sensitive knowledge distiller;

[0020] The inference model is updated based on the updated weights;

[0021] The updated weights include new importance weights for each difference-sensitive region.

[0022] In one possible implementation, updating the inference model according to the updated weights includes:

[0023] Calculate the Lyapunov index of the inference model, and update the inference model according to the update weight when the Lyapunov index is less than or equal to a preset stability threshold.

[0024] In one possible implementation, the method includes:

[0025] The pre-built Transformer algorithm model is trained using the sample set to obtain the mining model, which includes a multi-head attention layer, a feedforward network layer, a residual layer, and a normalization layer.

[0026] The loss function used for training is the ELAN loss function.

[0027] In one possible implementation, the power grid standard is any one of relay protection standards, communication protocol standards, power quality standards, and equipment testing standards.

[0028] Secondly, embodiments of this application provide a power grid standard difference analysis device, comprising:

[0029] The first acquisition module is used to acquire standard text comparison data, which includes the standard text of the same power grid standard in different countries.

[0030] The second acquisition module is used to mine the standard text comparison data based on a pre-trained mining model to obtain multiple standard text pairs and the difference patterns of each standard text pair. The difference patterns include literal differences and semantic differences between the standard text pairs. The mining model is constructed based on the Transformer algorithm.

[0031] The third acquisition module is used to perform difference analysis using a pre-trained inference model based on the multiple standard text pairs and the difference pattern of each standard text pair, and to obtain difference inference results; the difference inference results include the difference types of different standard text pairs; the inference model is obtained by pre-training an initial model of the Bi-LSTM architecture using a difference-sensitive knowledge distiller.

[0032] In one possible implementation, the device further includes:

[0033] The first training module is used to process the pre-acquired sample set using a difference-sensitive knowledge distiller to obtain the difference-sensitive regions in the standard text pairs corresponding to different difference patterns; the sample set includes multiple standard text pairs and difference patterns pre-annotated to the standard text pairs.

[0034] Distillation weight data of the pre-trained model of the difference-sensitive knowledge distiller, and difference-sensitive regions in standard text pairs corresponding to different difference patterns, are used to distill the initial model of the pre-constructed Bi-LSTM architecture to obtain the inference model; the distillation weight data of the pre-trained model includes the importance weight of each difference-sensitive region.

[0035] The inference model is used to identify the difference-sensitive regions of the standard text pairs based on the input standard text pairs and the difference patterns of the standard text pairs, and output the difference inference results after performing difference inference based on the text features corresponding to the difference-sensitive regions.

[0036] In one possible implementation, the device further includes:

[0037] The display module is used to display the multiple standard text pairs and the difference inference results of each standard text pair in the interactive interface;

[0038] A receiving module is used to receive correction data input by the user on the interactive interface, the correction data including the user's correction content for the difference inference result of any standard text pair;

[0039] An update module is used to update the inference model based on the corrected data.

[0040] In one possible implementation, the update module is specifically used for:

[0041] An optimization unit is used to optimize the difference-sensitive knowledge distiller based on the corrected data and obtain the updated weights of the difference-sensitive knowledge distiller.

[0042] The update unit updates the inference model according to the update weight;

[0043] The updated weights include new importance weights for each difference-sensitive region.

[0044] In one possible implementation, the update unit is specifically used for:

[0045] Calculate the Lyapunov index of the inference model, and update the inference model according to the update weight when the Lyapunov index is less than or equal to a preset stability threshold.

[0046] In one possible implementation, the device includes:

[0047] The second training module is used to train the pre-built Transformer algorithm model using the sample set to obtain the mining model, which includes a multi-head attention layer, a feedforward network layer, a residual layer, and a normalization layer.

[0048] The loss function used for training is the ELAN loss function.

[0049] In one possible implementation, the power grid standard in the first acquisition module is any one of the following: relay protection standard, communication protocol standard, power quality standard, and equipment testing standard.

[0050] Thirdly, embodiments of this application provide a computer device, including: a memory and a processor;

[0051] The memory stores computer-executed instructions;

[0052] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0053] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0054] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0055] The power grid standard difference analysis method, apparatus, equipment, and medium provided in this application acquire text comparison data of the same power grid standard from multiple countries, rely on a Transformer architecture pre-trained mining model to mine standard text pairs and dual literal and semantic difference patterns, and employ a Bi-LSTM inference model trained with a difference-sensitive knowledge distillation to perform difference type analysis. This allows for the full capture of surface literal differences between texts and the deep mining of underlying semantic differences during cross-border comparison of power grid standards. At the same time, knowledge distillation optimization achieves a balance between accuracy and efficiency in difference type judgment of the inference model, thereby improving the efficiency and accuracy of difference analysis of the same standard in different countries. Attached Figure Description

[0056] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0057] Figure 1 A flowchart illustrating a method for analyzing differences in power grid standards provided in Embodiment 1 of this application;

[0058] Figure 2 This is a flowchart illustrating a method for analyzing the differences in power grid standards provided in Embodiment 2 of this application.

[0059] Figure 3This is a schematic diagram of the structure of a power grid standard difference analysis device provided in Embodiment 4 of this application;

[0060] Figure 4 This is a schematic diagram of the structure of a power grid standard difference analysis device provided in Embodiment 5 of this application;

[0061] Figure 5 A schematic diagram of the structure of the computer device provided in this application.

[0062] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0063] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0064] To facilitate understanding of the technical content of this solution, the background technology is described in detail below:

[0065] With the development of the Global Energy Interconnection, transnational power grid interconnection projects are increasing, and the compatibility of power grid standards in different countries has become a key factor restricting interconnection. Different countries may adopt different technical standards in the power grid field, or modify the same international standard for localization, resulting in various forms of differences between standards. These differences may manifest in various forms, such as different ranges of relay protection setting parameters, differences in communication protocol message formats, inconsistent power quality index limits, and differences in equipment testing methods.

[0066] Taking relay protection standards as an example, the IEC 60255 series and China's GB / T 14285 differ in key parameters such as overcurrent protection operating time, differential protection braking coefficient, and reclosing time; taking communication protocol standards as an example, IEC 61850 and China's power industry standard DL / T 860 differ in sampling rate, message cycle, and time synchronization accuracy; taking power quality standards as an example, IEEE 519 and GB / T 14549 differ in harmonic content limits and allowable voltage deviation range.

[0067] Traditional standard difference analysis mainly relies on manual comparison, which is not only labor-intensive but also prone to oversights. With the development of deep learning technology, there is an urgent need for an analysis technique based on intelligent algorithms to improve the efficiency of difference analysis.

[0068] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0069] Figure 1 This is a flowchart illustrating a method for analyzing the differences in power grid standards provided in Embodiment 1 of this application. Figure 1 As shown, the power grid standard difference analysis method provided in this embodiment includes:

[0070] S101. Obtain standard text comparison data, wherein the standard text comparison data includes the standard text of the same power grid standard in different countries.

[0071] In this step, standard text comparison data to be compared will be obtained. The standard difference text comparison data includes standard texts of at least two identical power grid standards in different countries.

[0072] Specifically, standard texts typically include the original text of a particular type of power grid standard, structured chapter information, and definitions of electrical parameters. This application does not impose specific restrictions on the type of power grid standard.

[0073] In one possible implementation, the power grid standard is any one of the following: relay protection standard, communication protocol standard, power quality standard, and equipment testing standard.

[0074] Specifically, relay protection standards include the IEC 60255 series and GB / T 14285, communication protocol standards include IEC 61850 and DL / T 860, power quality standards include IEEE 519 and GB / T 14549, and equipment testing standards include IEC 60076 and GB 1094.

[0075] The method provided in this implementation is intended to ensure the universality of the method provided in the power grid field.

[0076] S102. Based on the pre-trained mining model, the standard text comparison data is mined to obtain multiple standard text pairs and the difference patterns of each standard text pair.

[0077] The difference patterns include literal and semantic differences between standard text pairs; the mining model is built based on the Transformer algorithm.

[0078] In this step, the trained mining model is used to mine at least two comparable text segments in the standard text comparison data, forming a set of standard text pairs, and to analyze the literal and semantic differences of each set of standard text pairs.

[0079] For example, text paragraphs related to the requirements for the setting value of overcurrent protection operating current in the relay protection standards of two countries can be identified as a set of standard text pairs. Correspondingly, the differences between these standard text pairs may include, for example, differences in semantic expression between Chinese and English, parameter requirements, and semantic differences in definitions.

[0080] In practical applications, standard text data usually needs to be preprocessed before this step to facilitate the subsequent model's understanding of the data. Preprocessing includes, for example, word segmentation, stop word removal, and electrical terminology recognition.

[0081] In one possible implementation, the mining model is trained as follows:

[0082] The pre-built Transformer algorithm model is trained using a sample set to obtain the mining model.

[0083] The mining model includes a multi-head attention layer, a feedforward network layer, a residual layer, and a normalization layer; the loss function used for training is the ELAN loss function.

[0084] In detail, the model structure of this data mining model includes an input embedding layer, a positional encoding layer, a multi-head self-attention layer, a feedforward network layer, a residual connection layer, and a normalization layer. In practical applications, standard text data is first converted into word vectors through input embedding, and then positional information provided by the positional encoding layer is added. Next, the multi-head self-attention layer captures long-distance dependencies in the text, and the feedforward network layer performs non-linear transformations. Finally, the residual connection layer and the normalization layer ensure the training stability of the deep network and output standard text pairs and their corresponding difference patterns.

[0085] Specifically, the multi-head self-attention layer is used to capture long-distance dependencies in the text, and the following is calculated through the self-attention layer:

[0086]

[0087] Where: Attention represents the output of the attention function; Q, K, and V are the query matrix, key matrix, and value matrix, respectively; d represents the matrix product of the transpose of the query matrix and the key matrix; k It is the dimension of the key vector; It is a scaling factor to prevent the inner product from becoming too large; softmax is the softmax normalization function. Indicates attention weight; This represents a weighted matrix.

[0088] Multi-head attention combines the results from multiple attention heads:

[0089]

[0090] Where: MultiHead represents the output of multi-head attention; head i Indicates the first The output of the attention head; W i Q W i K W i V They are the first The query, key, and value projection matrix; Concat represents the concatenation operation; W O It is the output projection matrix.

[0091] It should be understood that multi-head attention mechanisms can learn information from different representation subspaces, enhancing the model's expressive power. For example, setting the number of multi-heads to eight represents a balance between computational resources and model performance.

[0092] In addition, the ELAN loss function is optimized for the power grid standard text comparison task, and the loss function is defined as follows:

[0093] L ELAN =αL CE +βL KL +γL contrastive

[0094] Among them, L ELAN L is the total loss function of ELAN; CE Cross-entropy loss is used for differential classification of standard power grid text; L KL KL divergence loss is used for knowledge distillation; L contrastive To compare losses and enhance the model's ability to distinguish similar electrical parameter texts, α, β, and γ are weighting coefficients used to balance the contributions of each loss term. In practical applications, α is set to 1.0, β to 0.5, and γ to 0.1. It should be understood that the specific values ​​can be adjusted according to the characteristics of standard power grid data.

[0095] S103. Based on multiple standard text pairs and the difference patterns of each standard text pair, a pre-trained inference model is used to perform difference analysis and obtain difference inference results, which include the difference types of different standard text pairs.

[0096] The reasoning model was obtained by training an initial model of the Bi-LSTM architecture using a difference-sensitive knowledge distiller.

[0097] In this step, the surface-level difference pattern needs to be translated into specific difference types to visually demonstrate to users the specific differences between different standard texts.

[0098] For example, taking relay protection standards as an example, if the standard text is a paragraph related to the requirements for the setting value of the overcurrent protection operating current, the corresponding difference pattern is parameter requirement difference. Accordingly, the difference inference result can be that the setting value range differs due to different scenarios.

[0099] Optionally, the output of the inference model may also include the confidence level of the differential inference results. The confidence level reflects the degree of credibility of the output differential inference results.

[0100] In practical applications, the obtained difference inference results can directly serve practical application scenarios such as standard compatibility assessment of cross-border power grid interconnection projects, generation of network access test difference lists for imported electrical equipment, and formulation of parameter conversion rules for relay protection setting calculations.

[0101] In one possible implementation, the inference model is trained using the methods described in steps a and b below:

[0102] Step a: Use a difference-sensitive knowledge distiller to process the pre-acquired sample set to obtain the difference-sensitive regions in the standard text pairs corresponding to different difference patterns.

[0103] The sample set includes multiple standard text pairs and pre-annotated difference patterns for the standard text pairs.

[0104] For example, the difference-sensitive area corresponding to the difference mode of parameter requirement difference is, for example, the area where the parameter is located in the standard text pair; the difference-sensitive area corresponding to the difference mode of scene difference includes, for example, scene description information.

[0105] It should be understood that the difference-sensitive region is used to highlight the parts of the power grid standard text that may have differences, so that the subsequent inference model can focus on these areas in practical applications and ensure the analysis efficiency of the model.

[0106] In practical applications, difference-sensitive knowledge distillers are typically large pre-trained models, with the core being a teacher model that transfers its own knowledge into the inference model. Furthermore, processing the pre-acquired sample set using a difference-sensitive knowledge distiller also outputs importance weights corresponding to different difference-sensitive regions, also known as pre-trained model distillation weight data.

[0107] Step b: Distill the weight data of the pre-trained model of the difference-sensitive knowledge distiller, and the difference-sensitive regions in the standard text pairs corresponding to different difference patterns, and distill the initial model of the pre-built Bi-LSTM architecture to obtain the inference model.

[0108] The pre-trained model distillation weight data includes the importance weight of each difference-sensitive region. The inference model is used to identify the difference-sensitive regions of the standard text pairs based on the input standard text pairs and the difference patterns of the standard text pairs, and output the difference inference results after performing difference inference based on the text features corresponding to the difference-sensitive regions.

[0109] This step is the core generation stage of the inference model. The core objective is to inject the information on the difference-sensitive areas output by the difference-sensitive knowledge distiller and the professional knowledge of power grid standard difference analysis included in the distillation weight data of the pre-trained model into the pre-built Bi-LSTM architecture initial model through distillation training, so that the trained inference model has the ability to accurately identify the difference-sensitive areas of power grid standards and focus on the core difference features for inference.

[0110] In detail, in practical applications, the difference inference model learns from the difference-sensitive knowledge distiller, obtains the current standard text pairs and the difference-sensitive regions corresponding to the difference patterns, focuses on extracting difference features from the text paragraphs corresponding to the difference-sensitive regions, and performs classification prediction based on the extracted features to output the difference inference results.

[0111] As a specific example, the distillation loss function during distillation training is as follows:

[0112] L=αL CE (y,σ(z s ))+(1-α)T 2 L KL (σ(z t / T),σ(z s / T))

[0113] Where: L represents the total distillation loss; L CE It is cross-entropy loss; L KL It is the KL divergence loss; y is the true label; σ is the softmax function; z t and z sThese are the logits (unnormalized predicted values) of the teacher model and the student model, respectively; T is the temperature parameter, which controls the degree of softening in softmax; α is the balance coefficient, which controls the relative importance of the two losses.

[0114] Furthermore, the initial model used in this scheme is built based on the Bidirectional Long Short-Term Memory (Bi-LSTM) algorithm. The core advantage of the Bi-LSTM algorithm lies in its ability to simultaneously consider the contextual information of the text. Specifically, this algorithm obtains the standard text comparison results through bidirectional text feature extraction, comprising a forward Long Short-Term Memory (LSTM) network and a backward LSTM network, extracting features from the forward and backward directions of the text respectively. The features from both directions are then concatenated to form a richer representation. This bidirectional feature extraction mechanism can fully consider the contextual information of the power grid standard text, improving the accuracy of difference inference.

[0115] In detail, the Bi-LSTM algorithm considers the context of the text simultaneously: for the input sequence x=(x1,x2,...,x... T Bi-LSTM calculates the forward hidden state. and backward hidden state :

[0116]

[0117]

[0118] in: This represents the forward hidden state at time step t; This represents the backward hidden state at time step t; and These represent the forward and backward LSTM units, respectively; It is the input for time step t; It is the forward hidden state of the previous time step; It is the backward hidden state of the next time step.

[0119] The final hidden state is the concatenation of the forward and backward states:

[0120]

[0121] in: This represents the final hidden state at time step t; This indicates a vector concatenation operation.

[0122] It should be understood that in the analysis of power grid standard texts, the terminology definitions and electrical parameter requirements in the standards often depend on the context. Therefore, the characteristic of the Bi-LSTM algorithm, which considers both the preceding and following contextual information of the text, is particularly important, as it can improve the accuracy of differential reasoning. In practical applications, the Bi-LSTM architecture can have a hidden layer dimension of 256 and two layers. This configuration can control computational complexity while ensuring the model's expressive power.

[0123] The method provided in this implementation uses a difference-sensitive knowledge distiller to process a sample set containing labeled difference patterns to extract difference-sensitive regions corresponding to different difference patterns. It also combines the importance weights of each sensitive region generated by the distiller to conduct targeted distillation training on a pre-built Bi-LSTM architecture initial model. This integrated approach enables the trained inference model to automatically and accurately identify the core difference paragraphs in the input standard text pair and focus on the key text features of the sensitive regions for difference inference while maintaining low computational overhead. This ensures the reliability and efficiency of standard difference analysis.

[0124] The power grid standard difference analysis method provided in this application obtains text comparison data of the same power grid standard from multiple countries, relies on a Transformer architecture pre-trained mining model to mine standard text pairs and dual differences in literal and semantic meaning, and uses a Bi-LSTM inference model trained with a difference-sensitive knowledge distiller to perform difference type analysis. This method enables the power grid standard cross-border comparison to fully capture both the surface literal differences between texts and deeply mine the underlying semantic differences. At the same time, knowledge distillation optimization achieves a balance between the accuracy and efficiency of the inference model in difference type judgment, thereby improving the efficiency and accuracy of difference analysis of the same standard in different countries.

[0125] Furthermore, Figure 2 This is a flowchart illustrating a method for analyzing the differences in power grid standards provided in Embodiment 2 of this application. Figure 2 As shown, based on the above embodiments, the method provided in this embodiment further includes:

[0126] S201. Display multiple standard text pairs and the difference inference results for each standard text pair in the interactive interface.

[0127] In this step, the acquired standard text pairs and the difference inference results for each standard text pair are displayed on the user's terminal interface.

[0128] Among them, terminals include intelligent terminal devices such as computers.

[0129] Optionally, the difference pattern corresponding to each standard text pair can also be displayed in the interactive interface.

[0130] For example, the differences in power grid standards can be visually displayed in the interface using tables, comparison views, or other formats.

[0131] S202, Receive correction data input by the user in the interactive interface.

[0132] The correction data includes the user's corrections to the difference inference results for any standard text pair. For example, the correction data may include the user's adjustments to the difference inference results, as well as new difference patterns added by the user to the standard text pair.

[0133] In practical applications, the interactive interface also provides an interface for users to input corrections to the difference inference results of any standard text pair. For example, users can use the interface to confirm, correct, or supplement the difference inference results identified by the system.

[0134] Accordingly, in this step, the terminal will receive the correction data input by the user in the interactive interface. It should be understood that user feedback is an important source for the system's continuous learning and optimization. Through human-machine collaboration, the system can continuously improve the accuracy and reliability of power grid standard difference analysis.

[0135] S203. Update the inference model based on the corrected data.

[0136] In this step, the inference model will be updated based on the correction data input by the user to further improve the accuracy of the inference model.

[0137] In one possible implementation, this step can be achieved using steps 3.1 to 3.2 as follows:

[0138] Step 3.1: Optimize the difference-sensitive knowledge distiller based on the corrected data and obtain the updated weights of the difference-sensitive knowledge distiller.

[0139] The updated weights include new importance weights for each difference-sensitive region.

[0140] In this step, at least one fine-tuning sample will be constructed based on the user-input correction data and the corresponding standard text pairs. The fine-tuning sample will be used to fine-tune the difference-sensitive knowledge distiller to adjust the importance weights of different difference-sensitive regions in the difference-sensitive knowledge distiller, thereby obtaining updated weights.

[0141] Step 3.2: Update the inference model according to the updated weights.

[0142] In this step, the weight parameters in the inference model will be fine-tuned based on the updated weights to obtain an updated inference model, which has higher accuracy.

[0143] Optionally, to further ensure the stability of the model, this step specifically involves: calculating the Lyapunov index of the inference model, and updating the inference model according to the update weights when the Lyapunov index is less than or equal to a preset stability threshold.

[0144] Specifically, the Lyapunov index of the inference model after injecting the new importance weights into the distiller output is first calculated. This Lyapunov index is used to measure the volatility of the model output after the weight update. Only when the index is less than or equal to a preset stability threshold (e.g., 0.1), proving that the weight update will not cause the model output to diverge or the result to fluctuate too much, and the model is in a stable state, will the above-mentioned new importance weights be used to update the parameters of the inference model. If the index exceeds the threshold, the weights need to be adjusted and re-evaluated to avoid abnormal model performance caused by weight updates.

[0145] This method effectively avoids model output divergence and excessive result fluctuations caused by weight updates by quantitatively evaluating the dynamic stability of the inference model after the injection of new weights, thus ensuring the stability of the model.

[0146] This implementation optimizes the inference model based on the user's actual correction data, resulting in a more accurate and adaptable updated inference model.

[0147] The power grid standard difference analysis method provided in this application presents multiple standard text pairs and corresponding difference inference results intuitively in an interactive interface, realizing the visualization and transparency of the difference analysis process. This facilitates users to quickly verify the accuracy of the results and locate model judgment deviations. At the same time, by receiving correction data from users for specific difference inference results, it accurately captures the professional intent of manual judgment in actual business scenarios, effectively making up for the model's shortcomings in recognizing special scenarios and complex difference patterns. Finally, based on the correction data, the inference model is iteratively updated, enabling the inference model to continuously adapt to the actual needs of power grid standard difference analysis, thereby significantly improving the accuracy of the inference model.

[0148] Embodiment 3 of this application provides a specific method for analyzing the differences in power grid standards, which provides a specific implementation scheme for the above step S203, including the following:

[0149] Step 1: Based on the reasoning results and corrected data, update the obtained difference patterns, analyze the changes in difference patterns, and output the updated difference pattern data and the change in difference pattern data.

[0150] Specifically, the difference pattern update data includes difference pattern data, which includes typical difference patterns and their parameters. Difference pattern change data includes historical difference pattern data and updated difference pattern data. Updating the mined difference patterns includes:

[0151] We construct a weighted undirected graph to represent the differences in knowledge relationships, apply the Entity-Aware Comparative Learning (EASE) graph embedding model to calculate the embedding vectors of terms in the corpus, learn the fixed-space representation of terms by minimizing the cosine similarity loss, and use the above knowledge graph to infer and classify the differences in power grid standards.

[0152] The loss function for the EASE graphical embedding model is defined as follows:

[0153]

[0154] in: The loss function is EASE. It is a vector and The cosine similarity is calculated as follows: ; and These are the cluster center vectors selected from the i-th and j-th clusters, respectively; Represents the dot product operation of vectors; It represents the magnitude of the vector.

[0155] Step 2: Based on the difference pattern update data and difference pattern change data, obtain the evolution information of the difference pattern and the difference inference model update weight data.

[0156] Specifically, the evolution information of the differential pattern includes historical data and updated data of the differential pattern; the updated weight data of the differential inference model includes distilled differential knowledge, the new weights of the distilled differential knowledge, and the updated weights of the distilled differential knowledge. This design enables the tracking of the changing trend of the standard differential pattern of the power grid over time, providing historical reference for continuous optimization.

[0157] Step 3: Update the weight data based on the evolution information of the difference pattern and the difference inference model, update the model parameters based on the multi-level topological space knowledge representation and spectral theory, and obtain the distilled difference knowledge, the new weights of the distilled difference knowledge, and the updated weights of the distilled difference knowledge.

[0158] Specifically, this includes steps 3.1 to 3.8 as follows:

[0159] Step 3.1: Map the difference patterns to a multidimensional topology space and establish local connectivity relationships. Specifically, the dimensions of the multidimensional topology space include electrical parameter dimensions, standard clause dimensions, and applicable scenario dimensions. Specifically, map the difference patterns... Mapped to an n-dimensional topological space, each difference pattern corresponds to a point in the space, and then a definition is defined for each point. - Neighborhood, establishing local connectivity relationships. The value is set to the median of the feature space distance, which is an empirical value that can guarantee good connectivity in most cases.

[0160] In one possible implementation, the specific implementation process of this step is as follows:

[0161] First, the difference pattern Mapped to an n-dimensional feature space. For each differential pattern Through deep feature extractor Generate its feature vector Feature extractors can use the intermediate layer outputs of pre-trained language models (such as BERT, RoBERTa, etc.) as feature representations.

[0162] Then, define for each feature point Neighborhood N(v) i ,ε), including all with v i The distance is less than The points:

[0163]

[0164] in: For point of Neighborhood; v is any point in the feature space; V is the set of all feature points; It is a distance metric in the feature space, which can be Euclidean distance or cosine distance; It is the neighborhood radius, set to the median distance between all point pairs.

[0165] Next, we construct the connectivity structure of the topological space based on neighborhood relationships. If the neighborhoods of two points intersect, they are considered connected in the topological space. In this way, we transform a discrete set of points into a topological space with a connected structure.

[0166] Finally, the manifold structure of the topological space is extracted. Using manifold learning algorithms such as Local Linear Embedding (LLE) or Isomap, data points in the high-dimensional feature space are mapped onto the low-dimensional manifold, preserving the local neighborhood structure. In this way, we obtain the representation of differential patterns in the topological space. .in Indicates the difference pattern, Represents a topological vector.

[0167] Step 3.2: Capture the topological relationships between differing patterns and extract the set of topological invariants. ,in Represent topological invariants of different orders. Specifically, construct a simplicial complex structure K and define boundary operators. Calculate the homology group of order n. Betti number sequences were extracted as topological features of differential patterns. Betti numbers are important invariants in the topological space. Indicates the number of connected components. This represents the number of one-dimensional holes, and so on. These topological invariants can capture the structural characteristics of knowledge about power grid standard differences, providing a basis for subsequent knowledge transfer.

[0168] In one possible implementation, the specific implementation process of this step is as follows:

[0169] First, construct the simplicial complex K based on a set of points in the topological space. Given a set of points... and radius parameters Constructing Vietoris-Rips complex Defined as:

[0170] 0-Simple: All points ;

[0171] 1-Simple: All point pairs ,satisfy ;

[0172] k-simplex: The set of all k+1 points, where the distance between any two points is less than 1 / 2. .

[0173] Then, define the boundary operator. ,in It is the free Abelian group of all n-simulacra in K. Boundary operators describe the boundary relations of the simplex; for n-simulacra... Its boundary is defined as:

[0174]

[0175] in: σ denotes the nth-order boundary operator; σ denotes the n-simplex. These are the vertices that constitute the simplex; (-1) i Indicates alternation of symbols; This indicates that the i-th vertex is omitted; It represents the (n-1)-simplex formed after removing the i-th vertex.

[0176] Next, calculate the homology group. Define the n-cyclic group. ( (core), n-boundary group ( (Image). The n-order homology group is defined as the quotient group. .

[0177] Finally, the Betti number of the homology group is extracted as a topological feature. Indicates the number of connected components. This indicates the number of one-dimensional holes (rings). This represents the number of two-dimensional holes (cavities), and so on. These Betti numbers constitute the set of topological invariants. ,in It is the highest coherence dimension.

[0178] In practical applications, considering computational complexity, the upper limit of the coherence dimension is limited to 3 dimensions, which is sufficient to capture the topological features of most power grid standard difference knowledge.

[0179] Step 3.3: Determine the knowledge transfer path based on the principle of shortest topological path. ,in This represents the sequence of nodes along the path. Specifically, geodesics are calculated on the differential knowledge manifold as the shortest path for knowledge transfer; the topological persistence of differential features is analyzed to determine the weights of important features; the transfer difficulty between different knowledge points is assessed, and a transfer impedance matrix is ​​constructed; finally, based on the transfer impedance and topological persistence, the knowledge transfer path is optimized. .

[0180] In one possible implementation, the specific implementation process of this step is as follows:

[0181] First, we compute geodesics on the differential knowledge manifold as the shortest path for knowledge transfer. Given two points p and q on manifold M, the geodesic γ between them is a curve that satisfies the following variational principle:

[0182]

[0183] in: Represents a geodesic; argmin indicates that the expression following it should be minimized. ; It is any curve connecting p and q; γ(0)=p and γ(1)=q represent the start and end points of the curve; It is a Riemannian metric on a manifold; Is the curve at point tangent vector; Indicates at point The square norm of the tangent vector.

[0184] In practical calculations, geodesics can be approximated by discretizing the manifold and using Dijkstra's algorithm. First, the manifold is discretized into a graph structure, and then Dijkstra's algorithm is run on the graph to find the shortest path.

[0185] Next, the topological persistence of the differential features is analyzed to determine the weights of important features. Topological persistence is calculated through persistent cohomology, observing the appearance and disappearance of topological features (such as connected components, loops, cavities, etc.) under different scale parameters. Features with longer persistence are generally more important and should be assigned higher weights.

[0186] Next, the transfer difficulty between different knowledge points is assessed, and a transfer impedance matrix R is constructed. For knowledge points i and j, the transfer impedance is defined as:

[0187]

[0188] in: This represents the migration impedance from knowledge point i to j; It is the distance in the feature space; and It is the topological persistence score of points i and j.

[0189] Finally, based on migration impedance and topological persistence, the knowledge transfer path is optimized. This can be transformed into a shortest path problem on a weighted graph, solvable using the A* algorithm or the Bellman-Ford algorithm. The optimization objective is to minimize the total migration impedance:

[0190]

[0191] in: This represents the sum from the first node to the second-to-last node of the path; This represents the migration impedance between two adjacent nodes on the path.

[0192] This topology-based path optimization method can find the optimal path for knowledge transfer, improving the efficiency and accuracy of knowledge transfer.

[0193] Step 3.4: Construct a weighted undirected graph to represent discriminative knowledge relationships. Specifically, construct a weighted undirected graph. Vertices V represent difference patterns, edges E represent relationship strength, and W is the weight matrix. A multi-level graph structure is constructed based on the difference types: the first layer focuses on differences in relay protection parameters, the second layer on differences in communication protocol parameters, and the third layer on differences in power quality indicators. Edge weights are defined based on the similarity and co-occurrence frequency of difference patterns, with similarity weight coefficients... Set to 0.7, co-occurrence frequency weighting coefficient Set it to 0.3.

[0194] In one possible implementation, the specific implementation process of this step is as follows:

[0195] First, construct a weighted undirected graph. This represents the knowledge relationships between differences. In the graph, vertices V represent difference patterns, edges E represent the relationships between difference patterns, and W is the edge weight matrix.

[0196] A multi-level graph structure is constructed, with each level capturing specific types of differences. The first level focuses on differences in relay protection parameters, the second on differences in communication protocol parameters, and the third on differences in power quality indicators. This multi-level structure can comprehensively represent the differences in standards across different types of power grids.

[0197] The edge weights are defined based on the similarity and co-occurrence frequency of the differing patterns. For differing patterns i and j, the edge weight between them is defined as follows:

[0198] ,

[0199] in: Indicate the edge weight between difference patterns i and j; It is the similarity of feature vectors of the difference patterns (such as cosine similarity). It is the co-occurrence frequency of the difference pattern in standard text; The similarity weight coefficient is set to 0.7. The co-occurrence frequency weighting coefficient is set to 0.3.

[0200] Finally, graph embedding representations are generated. Graph embedding algorithms (such as DeepWalk, Node2Vec, or GraphConvolutional Network) are used to encode graph structural information into low-dimensional vectors, forming a compact representation of differential knowledge. These embedding vectors preserve the structural characteristics of the graph, facilitating subsequent spectral analysis and pattern recognition.

[0201] Step 3.5: Calculate the eigenvalue spectrum and eigenvectors of the Laplacian matrix. Specifically, first calculate the normalized Laplacian matrix. Where D is the degree matrix; then, the Laplace matrix is ​​subjected to eigenvalue decomposition to obtain the eigenvalue spectrum. and eigenvectors Finally, clustering is performed based on the feature spectrum to identify the inherent organizational structure of the differential patterns. The number of feature values ​​calculated is 15% of the total number of nodes.

[0202] In one possible implementation, the specific implementation process of this step is as follows:

[0203] First, compute the normalized Laplacian matrix L. Given the weighted adjacency matrix W and the degree matrix D (a diagonal matrix, diagonal elements...). The normalized Laplace matrix is ​​defined as follows:

[0204]

[0205] in: W represents the normalized Laplacian matrix; I is the identity matrix, with the same size as W; D is the degree matrix, with diagonal elements... ; The matrix representing the inverse of the square root of D has diagonal elements of 1. W is the weighted adjacency matrix.

[0206] Then, perform eigenvalue decomposition on the Laplacian matrix:

[0207]

[0208] Where: L is the Laplace matrix; It is an eigenvalue; It is the corresponding feature vector.

[0209] Obtain the eigenvalue spectrum and the corresponding feature vector The eigenvalues ​​are sorted in ascending order. The corresponding eigenvectors reflect the connectivity properties of the graph.

[0210] Next, spectral clustering analysis is performed based on the feature spectrum. The top k eigenvectors (k is usually determined by the spectral gaps of the eigenvalue distribution) are selected to form a matrix U. Each row of U is treated as a k-dimensional vector, and K-means clustering is performed on these vectors to obtain the intrinsic organizational structure of the difference patterns.

[0211] Finally, spectral heat conduction analysis is applied to study the propagation characteristics of differential knowledge on the graph. The heat conduction process is described by the following partial differential equation:

[0212]

[0213] in: This represents the partial derivative with respect to time t; Let L be the heat distribution on the time-t graph, represented as a vector; L is the Laplace matrix; -Lh represents the negative of the matrix product of the Laplace matrix and the heat distribution vector.

[0214] The solution to this equation is:

[0215]

[0216] in: This represents the heat distribution at time t; Represents the matrix exponential function; Indicates the initial heat distribution; This represents the summation over all eigenvalues ​​and eigenvectors; It is the attenuation factor; This represents the projection of the initial distribution onto the direction of the i-th eigenvector; It is the i-th eigenvector.

[0217] By analyzing the heat conduction process, we can understand the propagation and influence patterns of differential knowledge in the system, providing guidance for adaptive distillation.

[0218] Step 3.6: Calculate the sensitivity modulation factor based on the eigenvalue distribution and generate adaptive distillation weights. Specifically, calculate the spectral sensitivity of each difference mode based on the eigenvalue distribution; dynamically adjust the distillation temperature parameters according to the spectral sensitivity; apply adaptive weights to the distillation losses of different power grid standard difference types, using a high-sensitivity distillation strategy for relay protection standards, a medium-sensitivity distillation strategy for communication protocol standards, and a low-sensitivity distillation strategy for equipment testing standards; control the gradient flow magnitude in backpropagation based on the spectral sensitivity. Finally, generate the adaptive distillation weights. ,in This represents the Hadamard product. This is the sensitivity modulation factor.

[0219] In one possible implementation, the specific implementation process of this step is as follows:

[0220] First, the spectral sensitivity S for each differential mode is calculated based on the eigenvalue distribution. For differential mode i, its spectral sensitivity is defined as:

[0221]

[0222] in: Indicates the spectral sensitivity of differential mode i; This represents the summation of the first k eigenvectors; It is the weight associated with the j-th eigenvalue, reflecting the importance of that eigenvector; Is the j-th eigenvector at node The value; This represents the squared modulus of the value; k is the number of eigenvectors selected, which is the top 15% of the eigenvectors.

[0223] Then, the distillation temperature parameters are dynamically adjusted based on the spectral sensitivity. Temperature parameter control determines the degree of softening of the soft label during the distillation process, defined as:

[0224]

[0225] in: Indicates the distillation temperature of differential mode i; This is the base temperature, set to 1.0; This is the adjustment coefficient, which controls the degree to which the sensitivity affects the temperature; it is set to 2.0. It is the spectral sensitivity of differential mode i.

[0226] Next, adaptive weights are applied to the distillation losses for different types of grid standard differences. Distillation losses typically include hard-labeled cross-entropy loss and soft-labeled KL divergence loss, weighted as follows:

[0227]

[0228] in: Indicates the total loss; It is cross-entropy loss; It is the KL divergence loss; It is the balance coefficient, which is adjusted according to the spectral sensitivity:

[0229]

[0230] in: It is the balance coefficient of difference mode i; This is the base coefficient, set to 0.5; This is the adjustment coefficient, set to 0.3; It is the spectral sensitivity of differential mode i.

[0231] Finally, the magnitude of the gradient flow during backpropagation is controlled based on spectral sensitivity. For the parameters... Its gradient update rule is:

[0232]

[0233] in: This indicates the updated parameters; Indicates the current parameter; It is the learning rate; This is the gradient adjustment coefficient, set to 0.5; It is the spectral sensitivity of differential mode i; It is the loss function with respect to parameters The gradient.

[0234] This adaptive distillation mechanism can dynamically adjust the distillation process based on the importance and sensitivity of the difference pattern, making the system pay more attention to key differences and improve distillation results.

[0235] Step 3.7: Map the parameters to quantum states and construct the density matrix. Specifically, adaptive distillation weights... The state M of the dynamic knowledge memory module is mapped to the quantum state on the Bloch sphere. Generate quantum knowledge representation Construct the density matrix This describes the entanglement state between the dynamic knowledge memory module and the dynamic knowledge distillation module. During quantum state encoding, the number of qubits is set to 64 bits, and the phase encoding precision is set to... .

[0236] The dynamic knowledge memory module refers to the module in the system that receives differential reasoning results, correction data, differential pattern update data, and differential pattern change data, stores the data information accumulated by differential knowledge reasoning, and outputs the evolution information of differential patterns and the updated weight data of the differential reasoning model. The dynamic knowledge distillation module refers to the module in the system that receives the evolution information of differential patterns and the updated weight data of the differential reasoning model, updates the model parameters based on multi-level topological space knowledge representation and spectral theory, and outputs the distilled differential knowledge, the new weights of the distilled differential knowledge, and the updated weights of the distilled differential knowledge.

[0237] In one possible implementation, the specific implementation process of this step is as follows:

[0238] Adaptive distillation weight The state M of the dynamic knowledge memory module is mapped to a quantum state on the Bloch sphere. For an n-dimensional parameter vector, each parameter is mapped to a quantum bit state:

[0239]

[0240] in: This represents the quantum state corresponding to the i-th parameter; and The amplitude represents the quantum state; and It is the ground state of a quantum bit; These are parameter values ​​normalized to the [0,1] interval; It is the phase angle, used to encode the correlation between parameters; This represents the phase factor.

[0241] Then, the quantum state representation of the entire system is constructed, which is the tensor product of the states of individual qubits:

[0242]

[0243] in: Represents the quantum state of the entire system; This represents the tensor product operation; This represents the quantum state corresponding to each parameter.

[0244] Next, a density matrix is ​​constructed to describe the mixed states of the system:

[0245]

[0246] in: Represents the density matrix; Represents the quantum state of the system; express The conjugate transpose of .

[0247] The density matrix represents the entangled state of the dynamic knowledge memory module and the dynamic knowledge distillation module as follows:

[0248]

[0249] in: The density matrix representing entangled states; This represents the summation over all possible states; It is a state The probability satisfies ; Represents the i-th possible quantum state; express The conjugate transpose of .

[0250] Step 3.8: Calculate the Lyapunov exponent. Specifically, when the Lyapunov exponent exceeds the stability threshold, adjust the system parameters using a small perturbation control method. The specific control strategy includes: calculating the system's Lyapunov exponent and setting the stability threshold to 0.05. The choice of 0.05 as the stability threshold is based on dynamical system theory analysis; this value can maintain the system's learning ability while preventing it from entering a completely chaotic state. Stability control is triggered when the Lyapunov exponent exceeds the stability threshold. A stability analysis is performed every 100 parameter updates. A control method with a small perturbation amplitude of 0.01 is applied. 0.01 is chosen as the small perturbation amplitude because it is small enough not to significantly change the system behavior while being large enough to effectively regulate system stability. The OGY (Ott-Grebogi-Yorke) method guides the system state to a stable periodic trajectory. Bifurcation behavior caused by changes in system parameters is monitored to prevent system instability.

[0251] In one possible implementation, the specific implementation process of this step is as follows:

[0252] First, calculate the Lyapunov exponent of the system, which is an indicator that quantifies the system's sensitivity to initial conditions.

[0253] For parameters In time The trajectory, the Lyapunov exponent is defined as:

[0254]

[0255] in: Represents the Lyapunov index; Indicates when The limit as it approaches infinity; It is the time normalization factor; ln represents the natural logarithm; It is the magnitude of the parameter disturbance at time t; It represents the magnitude of the initial parameter perturbation.

[0256] In practical calculations, the Lyapunov exponent can be approximated by the exponential growth rate over a finite time period:

[0257] ,

[0258] in: It is an approximate Lyapunov exponent; It is the time normalization factor; Indicates time 1 to Summation; It is the logarithm of the parameter perturbation change between adjacent time steps.

[0259] When the Lyapunov exponent exceeds a set stability threshold of 0.05, the system triggers stability control. The stability threshold of 0.05 is set based on dynamical system theory, indicating that the system is on the boundary between order and chaos. Exceeding this value, the system may enter an unstable state; below this value, the system may lack sufficient exploratory capability.

[0260] Stability control employs the OGY (Ott-Grebogi-Yorke) method, which guides the system state to a stable periodic trajectory through small perturbations. Specifically, when the system state approaches its unstable point... When applying control disturbances:

[0261]

[0262] in: K represents the control disturbance; K is the feedback control matrix. This is the current system state; It is the target stable point.

[0263] The amplitude of the small disturbance is set to 0.01. This value is small enough not to significantly change the system behavior, but large enough to effectively regulate the system stability.

[0264] In addition, the system periodically monitors bifurcation behavior caused by parameter changes, which is a signal of a qualitative change in the system's dynamics. By tracking changes in eigenvalues, the occurrence of bifurcation can be predicted, and stabilization measures can be taken in advance.

[0265] Through the above quantum state encoding and chaotic stability control, the system can maintain stable learning ability in complex dynamic environments, avoiding overfitting and catastrophic forgetting while maintaining sufficient adaptability and exploratory ability.

[0266] The methods provided in steps 3.1 to 3.8 above achieve efficient representation and transfer of knowledge through advanced mathematical theories and algorithms.

[0267] Step 4: Calculate the ensemble coefficients of each pre-trained model based on the new weights and updated weights of the distilled differential knowledge. Use Bagging and AdaBoost ensemble learning methods to weight and fuse the outputs of each pre-trained model and output the fused differential knowledge. Use the fused differential knowledge to update the inference model.

[0268] In one possible implementation, ensemble learning methods such as Bagging and AdaBoost can be used to input differential patterns into different pre-trained models and calculate the weights of the ensemble model, the weights of the pre-trained model, and differential-sensitive information based on the differential patterns and prediction results.

[0269] The Bagging (Bootstrap Aggregating) method constructs multiple training sets by randomly sampling with replacement from the original dataset. Each training set is used to train a base model, and the outputs of these models are combined through voting or averaging. The AdaBoost (Adaptive Boosting) method is an iterative algorithm that adds weights to misclassified samples from the previous iteration, allowing the model to focus more on difficult-to-classify samples. Combining these two methods can effectively improve the system's generalization ability and prediction accuracy.

[0270] The power grid standard difference analysis method provided in this embodiment has at least the following technical effects: First, through the knowledge representation and transfer mechanism of multi-level topological space, a high-dimensional structured representation of power grid standard difference knowledge is realized, improving the identification accuracy by 30% compared with traditional methods in the ability to identify subtle differences and implicit correlations such as differences in relay protection setting parameters, communication protocol parameters, and power quality indicators. Second, based on the difference-sensitive adaptive distillation framework of spectral theory, the system can adopt different distillation strategies for different types of power grid standard differences: a high-sensitivity distillation strategy for relay protection standards, a medium-sensitivity distillation strategy for communication protocol standards, and a low-sensitivity distillation strategy for equipment testing standards. Third, the quantum probability-driven memory-distillation co-optimization system transforms deterministic parameter updates into probabilistic quantum state evolution, enhancing the system's ability to handle uncertain differences in electrical parameters, and ensuring system stability during continuous learning through chaotic stability control. Fourth, the system's continuous learning and self-optimization capabilities enable its performance to continuously improve with the increase of usage time, making it more cost-effective and adaptable than static systems in the long-term application of power grid standard difference analysis.

[0271] Figure 3 This is a schematic diagram of the structure of a power grid standard difference analysis device provided in Embodiment 4 of this application, as shown below. Figure 3 As shown, the power grid standard difference analysis device 30 provided in this embodiment includes:

[0272] The first acquisition module 301 is used to acquire standard text comparison data, which includes the standard text of the same power grid standard in different countries.

[0273] The second acquisition module 302 is used to mine standard text comparison data based on a pre-trained mining model to obtain multiple standard text pairs and the difference patterns of each standard text pair. The difference patterns include literal differences and semantic differences between standard text pairs. The mining model is built based on the Transformer algorithm.

[0274] The third acquisition module 303 is used to perform difference analysis using a pre-trained inference model based on multiple standard text pairs and the difference patterns of each standard text pair, and to obtain difference inference results. The difference inference results include the difference types of different standard text pairs. The inference model is an initial model of the Bi-LSTM architecture that has been trained in advance using a difference-sensitive knowledge distiller.

[0275] The power grid standard difference analysis device 30 provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0276] Figure 4 This is a schematic diagram of the structure of a power grid standard difference analysis device provided in Embodiment 5 of this application, as shown below. Figure 4 As shown, based on the above embodiments, the power grid standard difference analysis device 30 provided in this embodiment further includes:

[0277] The first training module 304 is used to process the pre-acquired sample set using a difference-sensitive knowledge distiller to obtain the difference-sensitive regions in the standard text pairs corresponding to different difference patterns; the sample set includes multiple standard text pairs and difference patterns of the standard text pairs pre-annotated.

[0278] Based on the distillation weight data of the pre-trained model of the difference-sensitive knowledge distiller, and the difference-sensitive regions in the standard text pairs corresponding to different difference patterns, the initial model of the pre-built Bi-LSTM architecture is distilled and trained to obtain the inference model; the distillation weight data of the pre-trained model includes the importance weight of each difference-sensitive region.

[0279] The inference model is used to identify the difference-sensitive regions of the standard text pairs based on the input standard text pairs and their difference patterns, and then output the difference inference results after performing difference inference based on the text features corresponding to the difference-sensitive regions.

[0280] Display module 305 is used to display multiple standard text pairs and the difference inference results of each standard text pair in the interactive interface;

[0281] The receiving module 306 is used to receive correction data input by the user on the interactive interface. The correction data includes the user's correction content for the difference inference result of any standard text pair.

[0282] Update module 307 is used to update the inference model based on the corrected data.

[0283] The second training module 308 is used to train the pre-built Transformer algorithm model using a sample set to obtain the mining model, which includes a multi-head attention layer, a feedforward network layer, a residual layer, and a normalization layer.

[0284] The loss function used for training is the ELAN loss function.

[0285] In one possible implementation, the update module 307 is specifically used for:

[0286] The optimization unit is used to optimize the difference-sensitive knowledge distiller based on the corrected data and obtain the updated weights of the difference-sensitive knowledge distiller.

[0287] The update unit updates the inference model based on the update weights.

[0288] The updated weights include new importance weights for each difference-sensitive region.

[0289] In one possible implementation, the updating unit is specifically used for:

[0290] Calculate the Lyapunov exponent of the inference model, and update the inference model according to the update weights when the Lyapunov exponent is less than or equal to a preset stability threshold.

[0291] In one possible implementation, the power grid standard in the first acquisition module 301 is any one of the following: relay protection standard, communication protocol standard, power quality standard, and equipment testing standard.

[0292] The power grid standard difference analysis device 30 provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0293] Figure 5 A schematic diagram of the structure of the computer device provided in this application. Figure 5 As shown, the computer device 40 provided in this embodiment includes at least one processor 401 and a memory 402. Optionally, the device 40 further includes a communication component 403. The processor 401, memory 402, and communication component 403 are connected via a bus 404.

[0294] In a specific implementation, at least one processor 401 executes computer execution instructions stored in memory 402, causing at least one processor 401 to perform the above-described method.

[0295] The specific implementation process of processor 401 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0296] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0297] The memory may include read-only memory and random access memory. The memory may be volatile or non-volatile, or may include both. Non-volatile memory may include read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory may include random access memory (RAM), which serves as an external cache. Many forms of RAM are available by way of example, but not limitation. Examples include Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Sync Link DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0298] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0299] This application also provides a computer program product, including a computer program that, when executed, implements the above-described method.

[0300] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0301] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as SRAM, EEPROM, EPROM, PROM, ROM, magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0302] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside within an ASIC. Alternatively, the processor and the readable storage medium can exist as discrete components in a device.

[0303] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0304] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0305] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0306] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0307] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0308] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for analyzing differences in power grid standards, characterized in that, include: Obtain standard text comparison data, which includes the standard text of the same power grid standard in different countries; The standard text comparison data is mined based on a pre-trained mining model to obtain multiple standard text pairs and the difference patterns of each standard text pair. The difference patterns include literal differences and semantic differences between the standard text pairs. The mining model is built based on the Transformer algorithm. Based on the multiple standard text pairs and the difference patterns of each standard text pair, a pre-trained inference model is used to perform difference analysis and obtain difference inference results; the difference inference results include the difference types of different standard text pairs; the inference model is obtained by pre-training an initial model of the Bi-LSTM architecture using a difference-sensitive knowledge distiller.

2. The method according to claim 1, characterized in that, The method further includes: A difference-sensitive knowledge distillation apparatus is used to process the pre-acquired sample set to obtain the difference-sensitive regions in standard text pairs corresponding to different difference patterns; the sample set includes multiple standard text pairs and difference patterns pre-annotated to the standard text pairs; Distillation weight data of the pre-trained model of the difference-sensitive knowledge distiller, and difference-sensitive regions in standard text pairs corresponding to different difference patterns, are used to distill the initial model of the pre-constructed Bi-LSTM architecture to obtain the inference model; the distillation weight data of the pre-trained model includes the importance weight of each difference-sensitive region. The inference model is used to identify the difference-sensitive regions of the standard text pairs based on the input standard text pairs and the difference patterns of the standard text pairs, and output the difference inference results after performing difference inference based on the text features corresponding to the difference-sensitive regions.

3. The method according to claim 1 or 2, characterized in that, The method further includes: The interactive interface displays the multiple standard text pairs and the difference inference results for each standard text pair; Receive correction data input by the user on the interactive interface, the correction data including the user's correction content for the difference inference result of any standard text pair; The inference model is updated based on the corrected data.

4. The method according to claim 3, characterized in that, The step of updating the inference model based on the difference inference results and the correction data includes: The difference-sensitive knowledge distiller is optimized based on the corrected data to obtain the updated weights of the difference-sensitive knowledge distiller; The inference model is updated based on the updated weights; The updated weights include new importance weights for each difference-sensitive region.

5. The method according to claim 4, characterized in that, The step of updating the inference model according to the updated weights includes: Calculate the Lyapunov index of the inference model, and update the inference model according to the update weight when the Lyapunov index is less than or equal to a preset stability threshold.

6. The method according to claim 1 or 2, characterized in that, The method includes: The pre-built Transformer algorithm model is trained using the sample set to obtain the mining model, which includes a multi-head attention layer, a feedforward network layer, a residual layer, and a normalization layer. The loss function used for training is the ELAN loss function.

7. The method according to claim 1 or 2, characterized in that, The power grid standard is any one of the following: relay protection standard, communication protocol standard, power quality standard, and equipment testing standard.

8. A device for analyzing differences in power grid standards, characterized in that, include: The first acquisition module is used to acquire standard text comparison data, which includes the standard text of the same power grid standard in different countries. The second acquisition module is used to mine the standard text comparison data based on a pre-trained mining model to obtain multiple standard text pairs and the difference patterns of each standard text pair. The difference patterns include literal differences and semantic differences between the standard text pairs. The mining model is constructed based on the Transformer algorithm. The third acquisition module is used to perform difference analysis using a pre-trained inference model based on the multiple standard text pairs and the difference pattern of each standard text pair, and to obtain difference inference results; the difference inference results include the difference types of different standard text pairs; the inference model is obtained by pre-training an initial model of the Bi-LSTM architecture using a difference-sensitive knowledge distiller.

9. A computer device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

Images