A deep learning-based intelligent startup scheme generation system and method

Abstract

The application discloses a kind of based on deep learning's starting scheme intelligent generation system and method, its word segmentation module is used to utilize historical starting step and starting condition text to form starting operation word segmentation text library;Power grid equipment model separation module separates the power grid equipment model generated by EMS system, obtains equipment name library and power grid topology model, forms the topology model in the operation range;Topology feature extraction module extracts the topology feature in starting range according to the topology model in the operation range;Corpus generation module forms starting step and starting condition text corpus according to equipment name library and starting operation word segmentation text library;Deep learning training module is trained according to starting step sentence and starting condition sentence and finally forms the starting step model with sequence and the starting condition model without sequence, and generates starting scheme for equipment starting, improve the efficiency and accuracy of power grid equipment starting scheme compilation.

Description

Technical Field

[0001] This application relates to the field of information processing technology, and in particular to a deep learning-based intelligent generation system and method for startup schemes. Background Technology

[0002] With the rapid development of my country's economy, the demand for electricity is also growing rapidly. Every year, a large number of infrastructure and technological upgrading projects are put into operation. In power grid projects, a start-up plan must be prepared before the newly added equipment is put into use to verify whether the newly added equipment can be put into safe operation.

[0003] However, in the current technology, the preparation and review of equipment start-up plans can only be done manually by staff. This method is not only inefficient, but also prone to problems such as flaws in the correctness and rationality of the plan due to staff errors. Summary of the Invention

[0004] The purpose of this invention is to propose an intelligent startup scheme generation system and method based on deep learning, which can improve the efficiency and accuracy of startup scheme writing for power grid equipment.

[0005] To achieve this objective, the intelligent startup scheme generation system based on deep learning designed in this invention includes a word segmentation module, a power grid equipment model separation module, a topology feature extraction module, a corpus generation module, and a deep learning training module. The word segmentation module is used to form a word segmentation text library for startup operations by utilizing the historical startup steps and startup condition text of power grid equipment. The power grid equipment model separation module is used to separate the equipment name library and the power grid topology model from the power grid equipment model generated by the Energy Management System (EMS), and form a topology model within the operating range based on the power grid topology model and the start-up range limitation. The topology feature extraction module is used to extract topology features within the activation range based on the topology model within the operation range using a graph attention network (GAT). The corpus generation module is used to form a text corpus of startup steps and startup conditions based on the obtained device name library and startup operation word segmentation text library; The deep learning training module is used to train a sequential deep learning model on the start-up step statements based on the obtained topological features, start-up steps and start-up condition text corpus within the start-up range, to obtain a sequential start-up step model, and to train a non-sequential deep learning model on the start-up condition statements to obtain a non-sequential start-up condition model.

[0006] The beneficial effects of this invention are: This invention includes key processes such as word segmentation, power grid equipment model separation, topology feature extraction, corpus generation, and deep learning training. Through targeted feature processing of historical startup steps, startup texts, and power grid equipment models, it ultimately obtains a sequential startup step model and a non-sequential startup condition model. By generating these startup step and condition models, this invention achieves mechanization and intelligentization in the writing of startup plans for new power grid equipment, avoiding the inefficiencies, low accuracy, and insufficient rationality associated with manual compilation and verification of startup plans, thus improving the efficiency and accuracy of writing startup plans for new power grid equipment. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the structure of the present invention; Figure 2 This is a flowchart of the present invention; Figure 3 This is a mind map of the present invention; Figure 4 This is a flowchart illustrating the process of obtaining network topology features based on a graph attention neural network according to the present invention. Figure 5 This is a flowchart illustrating the process of constructing a deep learning model based on ERNIE according to the present invention. Detailed Implementation

[0008] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: like Figures 1-5 The intelligent generation system for startup schemes based on deep learning shown includes a word segmentation module, a power grid equipment model separation module, a topology feature extraction module, a corpus generation module, and a deep learning training module; The word segmentation module is used to form a word segmentation text library for startup operations by utilizing the historical startup steps and startup condition text of power grid equipment. The power grid equipment model separation module is used to separate the equipment name library and the power grid topology model from the power grid equipment model generated by the EMS system, and form a topology model within the operating range based on the power grid topology model and the start-up range restriction. The topology feature extraction module is used to extract topology features within the activation range based on the topology model within the operation range and using a graph attention neural network. The corpus generation module is used to form a text corpus of startup steps and startup conditions based on the obtained device name library and startup operation word segmentation text library; The deep learning training module is used to train a sequential deep learning model on the start-up step statements based on the obtained topological features, start-up steps and start-up condition text corpus within the start-up range, to obtain a sequential start-up step model, and to train a non-sequential deep learning model on the start-up condition statements to obtain a non-sequential start-up condition model.

[0009] In the above technical solution, the word segmentation module uses historical startup steps and startup condition text to form a startup operation word segmentation text library, including: Based on historical startup steps and startup condition text, text cleanup is performed to remove device name information and obtain startup operation text statements. Based on a neural network word segmentation model, the startup operation text statements are decomposed and separated to form a startup operation word segmentation text library.

[0010] In the above technical solution, the word segmentation module extracts startup steps and startup condition text from the historical startup schemes of new power grid equipment based on the Word2vec model in a neural network. It then cleans the text using the Natural Language Toolkit (NLTK) in the Python library, removing redundant information such as equipment names. Finally, based on a recurrent neural network word segmentation model, it decomposes and separates the startup operation text to obtain startup operation texts such as impact, phase determination, phase verification, load test, switch opening, and switch closing, forming a startup operation word segmentation text library. This startup operation word segmentation text library includes numerous startup texts, which can be parsed object-oriented and, combined with equipment ID information, vectorized to form a startup step and startup condition text library.

[0011] In the above technical solution, the power grid equipment model separation module separates the equipment name library and the power grid topology model based on the power grid equipment model generated by the EMS system, including: The power grid equipment model generated by the EMS system is transformed into a mathematical model described by nodes or branches. This node or branch-described mathematical model is then used to perform topology analysis on the power grid equipment model, separating the equipment name library and the power grid topology model. The resulting equipment name library can be parsed using text vectorization to obtain equipment IDs and startup operation information. The resulting power grid topology model can be used for power grid connection logic analysis and can generate a power grid topology model within the startup operation range based on the startup range.

[0012] In the above technical solution, the power grid equipment model separation module forms a topology model within the operating range based on the power grid topology model and in conjunction with the start-up range limitations, including: Based on the aforementioned power grid topology model, the power grid connection logic is analyzed, and an adjacency matrix is ​​generated in the connection logic table using the starting range as the retrieval condition, ultimately forming a topology model within the operational range. The resulting topology model within the operational range can be parsed by a graph neural network, forming topological features including the connection relationships between nodes in the network and the direction of power flow in the power grid, which are used as input for training the model.

[0013] In the above technical solution, the topology feature extraction module extracts topology features within the starting range based on the obtained topology model within the operating range, using a graph attention neural network algorithm, including: Taking the topology model within the operational range as the analysis object, the graph attention neural network algorithm is used to mine the information between nodes in the network, capture the correlation between nodes and edges in the topology, and extract the network topology features within the starting range. Graph attention neural network algorithms extract network topology features through node feature transformation, inter-node attention coefficient calculation, and information aggregation. The first step involves feature transformation of nodes and calculation of inter-node attention coefficients. α ic : ; In the formula, α ic For nodes i and nodes c Attention coefficient between; x i , x c x d They are nodes i ,node c and nodes d The initial attribute information vector; W∈R n1×n2 To map the node feature vectors from n2 dimensions to n1 dimensions; D(i) For nodes i The set of neighboring nodes; || represents the concatenation operation; It is a single-layer feedforward neural network; Leaky Re LU Function Compute Node i and its neighboring nodes c Attention weights between them; After traversing all neighbor nodes of a vertex in the graph attention neural network based on the attention coefficients, the information of each node is weighted and aggregated to form a new attribute vector for the target node. Then, all vertices in the graph attention neural network are traversed to form topological features within the starting range. These topological features within the starting range include information such as the power grid connection logic and power flow direction, and can be used as input for training deep learning models.

[0014] In the above technical solution, the corpus generation module is based on a vectorization model and uses the obtained device name library and startup operation word segmentation text library to form a startup step and startup condition text corpus, including: Based on the vectorization model, the startup operation word segmentation text library and device name library are vectorized. The device ID and startup operation phrase are split and recombined to transform the fragmented operation text and device name into semantic text that can express startup steps or startup conditions, and finally form a startup step and startup condition text corpus.

[0015] In the above technical solution, the corpus generation module, based on the equipment name database and the startup operation word segmentation text database, performs vectorized recombination on the two, limits the equipment operation items, matches the equipment ID with its corresponding startup operation, and generates fixed operation statements, such as: "The 220kV bus differential protection of a certain transformer has been reset and deactivated"; "Open a certain switch of the 220kV bus tie"; "A certain line is experiencing an impact, phase verification"; "Activate the PCS-923A overcurrent protection of a certain switch of the 220kV bus tie", etc., ultimately forming a startup step and startup condition text corpus. The formed startup step and startup condition text corpus contains all the necessary information such as equipment name and operation item in the operation statements, and can be used as input for training deep learning models.

[0016] In the above technical solution, the deep learning training module trains a sequential deep learning model on the startup step statements based on the obtained topological features, startup steps, and startup condition text corpus within the startup range, to obtain a sequential startup step model, including: Based on the Enhanced Representation through Knowledge Integration (ERNIE) model, this method takes a corpus of textual descriptions of startup steps and conditions, along with network topology features within the startup scope, as input. It then combines this with the inherent operational logic between startup step statements to form a sequential startup step model. This generated startup step model enables the mechanization and intelligentization of startup plan development for new power grid equipment, avoiding the inefficiencies, low accuracy, and insufficient rationality associated with manual development and verification of startup plans. This improves the efficiency and accuracy of startup plan development for new power grid equipment.

[0017] In the above technical solution, the deep learning training module trains the start-condition statements using a non-sequential deep learning model based on the obtained topological features, start-up steps, and start-condition text corpus within the start-up range, to obtain a non-sequential start-condition model, including: Based on the ERNIE model, this study uses a corpus of startup steps and conditions texts, along with network topology features within the startup scope, as input. By combining this with the unordered nature of startup condition statements, a non-sequential startup condition model is formed. The generated startup condition model enables the mechanization and intelligentization of startup scheme writing for new power grid equipment, avoiding the inefficiencies, low accuracy, and insufficient rationality associated with manual startup scheme writing and verification. This improves the efficiency and accuracy of startup scheme writing for new power grid equipment.

[0018] During model training, the network is divided into four layers based on the ERNIE model framework: an encoding layer, a knowledge integration layer, a multi-task learning layer, and a fine-tuning layer. In the encoding layer, a Transformer architecture is used, stacking multiple encoders to encode the input initiation steps / conditions into a context-aware representation. To obtain a word representation enhanced with richer semantic information, this application employs a multi-stage knowledge masking strategy during data training in the encoding layer, integrating phrase and entity-level knowledge into the language representation. Training is divided into three learning stages based on the lexical, syntactic, and semantic levels of the algorithm. The first stage uses basic-level masking. During training, through hyperparameter settings, 15% of the basic language units are randomly masked, including all Chinese characters in the sentence, and other basic units in the sentence are used as input to train the encoder to predict the masked units. The second stage uses phrase-level masking. Basic language units are also used as training input, but this stage randomly selects several phrases in the sentence, masking and predicting all basic units within the same phrase. The third stage is entity-level masking. Named entities include meaningful units such as station names and equipment operation items. In the entity masking phase, the named entities in the sentence are analyzed first, then all slots within those entities are masked and predicted. Training ends when the model achieves 95% accuracy in predicting entity units, summarizing the fixed power grid operation procedures and forming a small-scale expert rule base before moving to the next layer. In the knowledge integration layer, the ERNIE model integrates external domain-specific knowledge learned through pre-training on large amounts of data and referencing diverse prior knowledge, including the aforementioned small-scale expert rule base, into the training data. This allows the model to learn power grid operation procedures, equipment startup operation procedures, and basic knowledge, enhancing its semantic expressive ability and improving the accuracy of the generated text. The multi-task learning layer reassembles and concatenates the equipment IDs, operation statements, and other words learned from the previous two layers to form the output statement, such as "a line experiences an impact once," and calculates the cross-entropy loss function that minimizes the language model's prediction. L CE This is used to characterize the model training effect. The expression for the cross-entropy loss function is: ; In the formula,L CE It is cross-entropy loss; n The length of the sequence. V It refers to the size of the vocabulary (text database); y ij It is a one-hot vector representing the actual next word. In this vector, the first... j An element being 1 indicates that the actual next word is the 1st word in the vocabulary. j One word, and all other elements are 0; P ij The model predicts the next word as the first word in the vocabulary. j The probability of each word.

[0019] Training stops when the model's loss function on the training data no longer decreases within 500 iterations, and the model enters the fine-tuning layer. Located at the top of the model, the fine-tuning layer is used to fine-tune the model for downstream tasks to adapt it to specific task requirements. In the fine-tuning layer, the accuracy of the trained model is validated to adjust model hyperparameters or training data distribution. This training process is repeated until the model accuracy reaches 99%, resulting in the startup steps or startup condition model.

[0020] Furthermore, given that each statement in the startup step statement sequence contains operational logic such as the power grid equipment safety operation procedures and basic five-prevention systems, a deep learning model with a sequence is selected to train the startup step statements, ultimately forming a sequential startup step model. For the startup condition statement sequence, since there are no operational logic constraints between each statement, a deep learning model without a sequence is trained, ultimately forming a non-sequential startup condition model. The two models differ only in the operational logic judgment statements: "1" represents the sequential model, and "0" represents the non-sequential model; the rest of the structure is identical.

[0021] A deep learning-based intelligent generation method for startup schemes includes the following steps: S101. Use the historical startup steps and startup condition text of power grid equipment to form a startup operation word segmentation text library; S102. Based on the power grid equipment model generated by the EMS system, separate the equipment name library and the power grid topology model, and on the basis of the power grid topology model, combine the start-up range restrictions to form a topology model within the operating range; S103. Based on the topological model within the operating range, use a graph attention neural network to extract the topological features within the activation range. S104. Based on the obtained device name database and startup operation word segmentation text database, form a startup step and startup condition text corpus; S105. Based on the obtained topological features, startup steps, and startup condition text corpus within the startup range, a sequential deep learning model is used to train the startup step statements to obtain a sequential startup step model, and a non-sequential deep learning model is used to train the startup condition statements to obtain a non-sequential startup condition model.

[0022] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above.

[0023] The contents not described in detail in this specification are prior art known to those skilled in the art. Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0024] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0025] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0026] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0027] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

[0028] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A deep learning-based intelligent startup scheme generation system, characterized in that: It includes a word segmentation module, a power grid equipment model separation module, a topology feature extraction module, a corpus generation module, and a deep learning training module; The word segmentation module is used to form a word segmentation text library for startup operations by utilizing the historical startup steps and startup condition text of power grid equipment. The power grid equipment model separation module is used to separate the equipment name library and the power grid topology model from the power grid equipment model generated by the EMS system, and to form a topology model within the operating range based on the power grid topology model and the start-up range limitation. Separating the equipment name library and the power grid topology model includes converting the power grid equipment model generated by the EMS system into a mathematical model described by nodes or branches, and using the mathematical model described by nodes or branches to perform topology analysis on the power grid equipment model to separate the equipment name library and the power grid topology model. The topology feature extraction module is used to extract topology features within the activation range based on the topology model within the operation range and using a graph attention neural network. The corpus generation module is used to form a text corpus of startup steps and startup conditions based on the obtained device name library and startup operation word segmentation text library; The deep learning training module is used to train a sequential deep learning model on the start-up step statements based on the obtained topological features, start-up steps and start-up condition text corpus within the start-up range, to obtain a sequential start-up step model, and to train a non-sequential deep learning model on the start-up condition statements to obtain a non-sequential start-up condition model. The model training process is based on the ERNIE model framework, which divides the network into four layers: the encoding layer, the knowledge integration layer, the multi-task learning layer, and the fine-tuning layer. In the encoding layer, the Transformer architecture is used to stack multiple encoders together to encode the input initiation steps or initiation conditions text into context-aware representations. When training data in the encoding layer, a multi-stage knowledge masking strategy is adopted to integrate phrase and entity-level knowledge into the language representation. The training is divided into three learning stages according to the lexical, syntactic, and semantic levels of the algorithm.

2. The intelligent generation system for startup schemes based on deep learning according to claim 1, characterized in that, The word segmentation module uses historical startup steps and startup condition text to form a startup operation word segmentation text library, including: Based on historical startup steps and startup condition text, text cleanup is performed to remove device name information and obtain startup operation text statements. Based on a neural network word segmentation model, the startup operation text statements are decomposed and separated to form a startup operation word segmentation text library.

3. The intelligent generation system for startup schemes based on deep learning according to claim 1, characterized in that, The power grid equipment model separation module, based on the power grid topology model and considering the start-up range limitations, forms a topology model within the operating range, including: Based on the power grid topology model, the power grid connection logic is analyzed, and an adjacency matrix is ​​generated in the connection logic table using the starting range as the retrieval condition, ultimately forming a topology model within the operating range.

4. The intelligent generation system for startup schemes based on deep learning according to claim 1, characterized in that, The topology feature extraction module, based on the obtained topology model within the operational range, uses a graph attention neural network algorithm to extract topology features within the activation range, including: Taking the topology model within the operational range as the analysis object, the graph attention neural network algorithm is used to mine the information between nodes in the network, capture the correlation between nodes and edges in the topology, and extract the network topology features within the starting range. Graph attention neural network algorithms extract network topology features through node feature transformation, inter-node attention coefficient calculation, and information aggregation. The first step involves feature transformation of nodes and calculation of inter-node attention coefficients. α ic : ； In the formula, α ic For nodes i and nodes c Attention coefficient between; x i , x c , x d They are nodes i ,node c and nodes d The initial attribute information vector; W∈R n1×n2 To transform the node feature vector from n2 Dimension mapping to n1 dimension; D(i) For nodes i The set of neighboring nodes; || represents the concatenation operation; It is a single-layer feedforward neural network; Leaky Re LU Function Compute Node i and its neighboring nodes c Attention weights between them; After traversing all neighboring nodes of a vertex in the graph attention neural network based on the attention coefficient, the information of each node is weighted and aggregated to form a new attribute vector for the target node. Then, all vertices in the graph attention neural network are traversed to form topological features within the starting range.

5. The intelligent generation system for startup schemes based on deep learning according to claim 1, characterized in that, The corpus generation module is based on a vectorization model and uses the obtained device name database and startup operation word segmentation text database to form a startup step and startup condition text corpus, including: Based on the vectorization model, the startup operation word segmentation text library and device name library are vectorized. The device ID and startup operation phrase are split and recombined to transform the fragmented operation text and device name into semantic text that can express startup steps or startup conditions, and finally form a startup step and startup condition text corpus.

6. The intelligent generation system for startup schemes based on deep learning according to claim 1, characterized in that, The deep learning training module trains a sequential deep learning model on the start-up step statements based on the obtained topological features, start-up steps, and start-up condition text corpus within the start-up range, resulting in a sequential start-up step model, including: Based on the ERNIE model, a sequential startup step model is formed by taking the startup step and startup condition text corpus, as well as the network topology features within the startup range as input, and combining the inherent operational logic between startup step statements.

7. The intelligent generation system for startup schemes based on deep learning according to claim 1, characterized in that, The deep learning training module trains a non-sequence deep learning model on the activating condition statements based on the obtained topological features, activating steps, and activating condition text corpus within the activation range, resulting in a non-sequence activating condition model, including: Based on the ERNIE model, a non-sequential start condition model is formed by taking the start step and start condition text corpus and the network topology features within the start range as input, and combining the unordered nature of the start condition statements.

8. A deep learning-based intelligent generation method for startup schemes utilizing the system of claim 1, comprising the following steps: S101. Use the historical startup steps and startup condition text of power grid equipment to form a startup operation word segmentation text library; S102. Based on the power grid equipment model generated by the EMS system, separate the equipment name library and the power grid topology model, and on the basis of the power grid topology model, combine the start-up range restrictions to form a topology model within the operating range; S103. Based on the topological model within the operating range, use a graph attention neural network to extract the topological features within the activation range. S104. Based on the obtained device name database and startup operation word segmentation text database, form a startup step and startup condition text corpus; S105. Based on the obtained topological features, startup steps, and startup condition text corpus within the startup range, a sequential deep learning model is used to train the startup step statements to obtain a sequential startup step model, and a non-sequential deep learning model is used to train the startup condition statements to obtain a non-sequential startup condition model.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in claim 8.

Images