CN-122020392-A - Power network fault early warning method based on large model

CN122020392ACN 122020392 ACN122020392 ACN 122020392ACN-122020392-A

Abstract

The application relates to the technical field of power system automation, in particular to a power network fault early warning method based on a large model, which comprises the steps of using a lightweight encoder and a linear projection technology, the high-frequency power sensor data is directly mapped into a time sequence soft prompt vector which can be understood by a large model, so that deep alignment of continuous signals and discrete semantics is realized, and information loss and efficiency bottleneck caused by numerical texting are avoided. And then, the operation and maintenance knowledge base is directionally searched by utilizing the soft prompt vector, a cross-modal affinity matrix is introduced to construct a semantic gating mechanism, the searched key knowledge segments are dynamically screened and enhanced according to the real-time waveform characteristics, and redundant text noise irrelevant to the current working condition is automatically restrained. Finally, the composite sequence formed by the instruction, the enhanced knowledge and the soft prompt is subjected to deep attention interaction and reasoning through the large model, and a control signal containing fault classification and treatment suggestions is output, so that real-time accurate early warning which takes high-frequency signal acuity and expert knowledge logicality into consideration is realized.

Inventors

AN ZHIYUAN
YANG RUNHUA
LIU HUIFANG
ZHANG NINGNING
WANG CHUNYING
WU LIJIE
DONG JIAOJIAO
Quan Yizhan
FENG HAO
JIN JING
WANG HAOTIAN

Assignees

国网河南省电力公司信息通信分公司

Dates

Publication Date: 20260512
Application Date: 20260209

Claims (9)

1. The utility model provides a power network fault early warning method based on a large model, which is characterized by comprising the following steps: S1, carrying out standardized cleaning on an acquired original sensor data stream to obtain a normalized time sequence tensor, and carrying out vectorization processing on an operation and maintenance log text by utilizing a pre-training text embedding model to obtain a knowledge base vector index; S2, carrying out time sequence feature coding and soft prompt mapping on the normalized time sequence tensor to obtain a time sequence soft prompt embedded vector; S3, pooling the sequence soft prompt embedded vector to obtain a query vector, and performing neighbor search in a knowledge base vector index based on the query vector to search a search text set matched with the association; S4, converting the search text set into text embedded vectors, and performing sequential logic splicing on the text embedded vectors and sequential soft prompt embedded vectors and preset task instruction embedded vectors to obtain a composite input embedded sequence; s5, inputting the composite input embedded sequence into a pre-trained large language model to conduct large model reasoning so as to obtain fault classification probability distribution; and S6, generating a final early warning control signal containing fault types and treatment suggestions in response to the maximum probability value in the fault classification probability distribution exceeding a preset threshold.
2. The large model based power network fault pre-warning method of claim 1, wherein the raw sensor data stream comprises bus voltage, line current, active/reactive power, frequency and breaker on-off signals.
3. The large model-based power network fault early warning method according to claim 1, characterized in that step S1 comprises: Analyzing the original sensor data stream to identify a sensor node sampling time stamp, and resampling the original sensor data stream to obtain a sensor data matrix; Performing Z-Score normalization on the sensor data matrix to obtain a normalized time sequence tensor; washing and blocking unstructured texts of the collected operation and maintenance log texts to obtain text fragment sequences; and performing feature mapping coding on the text fragment sequence by using the pre-training text embedding model to obtain a knowledge base vector index.
4. The large model-based power network fault early warning method according to claim 1, characterized in that step S2 comprises: carrying out local time sequence pattern extraction based on 1D-CNN on the normalized time sequence tensor to obtain local feature mapping; Injecting position codes into the local feature mapping, and modeling global dependency relationship by utilizing a multi-head self-attention mechanism of a transducer encoder layer to obtain a global time sequence hiding state; And performing cross-modal characteristic linear projection on the global time sequence hidden state to obtain a time sequence soft prompt embedded vector.
5. The large model-based power network fault early warning method according to claim 1, characterized in that step S3 comprises: carrying out time sequence state aggregation on the time sequence soft prompt embedded vector to obtain a global state query vector; Vector space similarity calculation and neighbor sequencing are carried out on the global state query vector and each knowledge item vector in the knowledge base vector index so as to obtain a Top-K matching record list; And carrying out domain knowledge text content mapping extraction on the Top-K matching record list to obtain a retrieval text set.
6. The large model-based power network fault early warning method according to claim 1, characterized in that step S4 comprises: Vectorizing encoding is carried out on a preset system instruction text and a preset search text set to obtain an instruction embedded vector and a knowledge embedded vector; Performing splicing processing along the length dimension of the sequence on the instruction embedded vector, the knowledge embedded vector and the time sequence soft prompt embedded vector to obtain an unmasked mixed sequence; the unmasked mixed sequence is subjected to position coding injection to obtain a composite input embedded sequence.
7. The large model-based power network fault early warning method according to claim 1, characterized in that step S5 includes: Performing characteristic interaction based on a self-attention mechanism on the composite input embedded sequence to obtain deep context characteristic expression; performing linear transformation and nonlinear activation processing on a feedforward neural network for deep context feature expression input position perception to obtain an end hidden state vector of a convergence sequence reasoning conclusion; And performing dimension mapping and probability normalization on the tail end hidden state vector by using the trained linear classification head to obtain fault classification probability distribution.
8. The large model-based power network fault early warning method according to claim 1, characterized in that step S6 includes: Carrying out probability maximum likelihood estimation and risk threshold judgment on the fault classification probability distribution to obtain a risk triggering state tuple; responding to the risk triggering state tuple to indicate that the risk exists, and performing semantic mapping and interpretation text generation processing by using a fault metadata dictionary and a treatment suggestion template to obtain original message structure data; And carrying out serialization encapsulation and verification processing on the original message structure data based on a communication standard protocol of the power industry so as to obtain a final early warning control signal.
9. The large model-based power network fault early warning method according to claim 6, wherein the performing a splicing process along a sequence length dimension on the instruction embedded vector, the knowledge embedded vector, and the time sequence soft hint embedded vector to obtain the unmasked mixed sequence includes: Performing cross-modal homomorphic projection and affinity matrix calculation on the knowledge embedded vector and the time sequence soft prompt embedded vector to obtain a cross-modal affinity matrix; Performing signal-driven text semantic gating and feature re-weighting on the cross-modal affinity matrix and the knowledge embedding vector to obtain a gating enhanced knowledge vector; and performing self-adaptive topology splicing and mixed sequence construction on the gating enhancement knowledge vector, the instruction embedded vector and the time sequence soft prompt embedded vector to obtain an unmasked mixed sequence.

Description

Power network fault early warning method based on large model Technical Field The application relates to the technical field of power system automation, in particular to a power network fault early warning method based on a large model. Background With the rapid development of a novel power system, the topology structure of a power grid is more and more complex, and the randomness and fluctuation brought by new energy access enable the fault form to be diversified and hidden. In order to ensure safe and stable operation of the power grid, the use of artificial intelligence technology to realize accurate early warning of early faults has become industry consensus. In recent years, a large language model provides a new technical path for processing complex fault diagnosis and auxiliary decision-making problems in the electric power field by virtue of strong knowledge storage, logic reasoning and generalization capability, and particularly shows great potential in the aspect of fusing unstructured operation and maintenance knowledge and structured monitoring data. However, the application of large models, which are mainly good at processing discrete symbol text, directly to continuous, high-frequency power time series data analysis faces the fundamental challenge of non-uniformity of heterogeneous data encoding space in practical applications. In an attempt to solve this problem, the prior art generally employs a numerical textonym method, i.e., a method of mechanically converting a large number of sensor sample values into a string sequence (e.g., converting voltage values into text characters) input model. The method has the defects that on one hand, the continuous mathematical characteristics among numerical values are cut off, so that high-frequency transient information (such as voltage distortion or harmonic oscillation in microsecond level) is seriously lost in the conversion process, and on the other hand, the context window limit of a large model is rapidly exhausted when the high-frequency data are converted into text Token one by one, so that the reasoning cost is high and the efficiency is extremely low. In addition, when an external knowledge base is introduced by utilizing a search enhancement generation technology, the existing scheme is often used for simply and linearly splicing the searched historical fault text and the real-time data. Due to the lack of an active attention guiding mechanism based on real-time signal characteristics, redundant text noise which is mixed in the search result and is not matched with the current working condition cannot be effectively restrained. The simple modal stacking can not realize deep semantic interaction, is extremely easy to mislead the attention of a model, so that the model can generate illusion misjudgment against physical facts, and the severe requirements of a power system on high sensitivity, high real-time performance and high reliability of fault early warning are difficult to meet. Accordingly, an optimized large model-based power network fault warning scheme is desired. Disclosure of Invention The present application has been made to solve the above-mentioned technical problems. The embodiment of the application provides a power network fault early warning method based on a large model, which comprises the following steps: S1, carrying out standardized cleaning on an acquired original sensor data stream to obtain a normalized time sequence tensor, and carrying out vectorization processing on an operation and maintenance log text by utilizing a pre-training text embedding model to obtain a knowledge base vector index; S2, carrying out time sequence feature coding and soft prompt mapping on the normalized time sequence tensor to obtain a time sequence soft prompt embedded vector; S3, pooling the sequence soft prompt embedded vector to obtain a query vector, and performing neighbor search in a knowledge base vector index based on the query vector to search a search text set matched with the association; S4, converting the search text set into text embedded vectors, and performing sequential logic splicing on the text embedded vectors and sequential soft prompt embedded vectors and preset task instruction embedded vectors to obtain a composite input embedded sequence; s5, inputting the composite input embedded sequence into a pre-trained large language model to conduct large model reasoning so as to obtain fault classification probability distribution; and S6, generating a final early warning control signal containing fault types and treatment suggestions in response to the maximum probability value in the fault classification probability distribution exceeding a preset threshold. Compared with the prior art, the application provides a power network fault early warning method based on a large model. The method comprises the steps of constructing a time sequence feature encoder, directly mapping and projecting a high-frequency continuous original s