AI-based auxiliary office application system and method for power grid dispatching.

Abstract

This invention discloses a method and system for auxiliary office applications in power grid dispatching based on AI technology, belonging to the field of artificial intelligence. It includes a data preparation and preprocessing module, a model building module, a model integration module, and a model training and optimization module. The system acquires data from the OMS system, collects and organizes historical and labeled data, and performs preprocessing operations; loads the BERT model, defines the DPCNN architecture, and prepares the parameters and configurations required for model training; connects the output of the BERT model with the input of the DPCNN to construct a complete BERT-DPCNN model; and performs model training, parameter updates, performance monitoring, and testing and evaluation. This invention meets the complex needs of various business processes in the dispatching domain, improves the overall efficiency of power grid dispatching operations, further promotes the digital transformation of enterprises and the construction of a digital power grid; and links and coordinates multiple existing production management systems in the dispatching domain to supplement integrated dispatching construction and intelligent dispatching error prevention technologies.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an AI-based auxiliary office application system and method for power grid control. Background Technology

[0002] During the power grid dispatching and operation shift, dispatchers not only need to conduct real-time and uninterrupted dispatching, monitoring and control of the power grid operation, but also need to spend a lot of time and energy to complete various auxiliary business processes, such as: drafting dispatching operation tickets based on professional operational policy opinions on power grid operation mode, logging defects, faults and other abnormal situations that occur during dispatching and operation shifts, compiling real-time risk warning notices for power grid operation and corresponding dispatching and handling plans, etc.

[0003] Specifically, the following three problems exist:

[0004] First, the flexible and ever-changing nature of dispatch operation tickets hinders efficiency due to manual ticketing. The power grid operates in various modes, resulting in different dispatch operation tickets. Although there are historical typical ticketing functions, manual modification is still required each time to adapt to the current power grid operation mode. Furthermore, dispatchers must verify information across multiple production management systems according to the "three-way comparison" requirement when filling out dispatch operation tickets to ensure logical correctness. During periods of concentrated maintenance work, the time spent by dispatchers filling out and reviewing operation tickets is lengthy, impacting operational efficiency. Additionally, errors and non-standard practices in manually filling out operation tickets are also frequent.

[0005] Secondly, the sheer volume of daily operational records makes manual recording a heavy burden. These records primarily include shift handover logs, power supply reports, equipment defect records, and accident and anomaly handling records. While the dispatch and operation management system has corresponding information management modules for various daily operational information, the level of automation is low. On-duty dispatchers still need to switch between multiple systems to collect and integrate information when recording various daily operational data, a complex and tedious process. Furthermore, manual recording often results in incomplete or missing entries, posing potential safety hazards to the power grid.

[0006] Third, the format and standardization requirements for official documents are complex, and document processing is excessively time-consuming and labor-intensive. During the operation of power grid dispatching, it is necessary to frequently send official documents to other specialties, departments, and users, such as dispatching operation information letters, new equipment start-up plans, and real-time power grid operation risk notices. These official documents have standardized terminology and format requirements. The large amount of document processing work increases the workload of dispatchers, distracts them, and brings risks to power grid dispatching operations.

[0007] Currently, existing power grid control and operation interaction projects still face the following challenges: Semantic analysis and knowledge graphs heavily rely on large amounts of high-quality data; maintaining and updating the knowledge base requires specialized knowledge and may present difficulties in handling and understanding ambiguities; Deep learning and neural networks require large amounts of labeled data for training, resulting in poor model interpretability and high computational resource requirements; Natural language processing has limited ability to understand and process domain-specific languages, and may encounter difficulties in handling polysemy and contextual relevance; Big data technology faces data privacy and security issues, making it difficult to guarantee data quality and consistency, requiring professional data management and processing skills; Artificial intelligence platforms face stability and scalability challenges, with complex resource allocation and management in multi-user environments, demanding high developer skills; Human-computer interaction technology faces user adaptability and acceptance issues, including the complexity of interaction design and reliability and stability issues in specific environments; Graphical operation and intelligent approval face challenges in the intuitiveness and usability of the user interface, potentially leading to limitations in complex operations and a strong reliance on graphical technology. Summary of the Invention

[0008] In view of the aforementioned existing problems, this invention can analyze and learn historical data, predict and judge auxiliary business needs, and efficiently generate written text to meet complex scheduling requirements. This method fully leverages the advantages of the BERT pre-trained model in text semantic understanding to generate high-quality text word vectors. Based on this, BERT is combined with an optimized DPCNN (Deep Convolutional Neural Network) to deeply optimize the text feature extraction model, not only replacing the original output layer but also finely adjusting the entire model structure. In particular, the Mish activation function is cleverly used instead of the traditional ReLU activation function in the DPCNN convolutional block design. This improvement allows the model to more sensitively capture key information in the text while effectively mitigating the gradient explosion problem, enhancing the model's stability and robustness. Furthermore, this paper introduces a dynamic learning rate mechanism to further optimize the model's training method, enabling the model to adaptively adjust the learning rate during training, thereby improving training efficiency while ensuring the model's final performance.

[0009] To address the aforementioned technical challenges, an AI-based auxiliary office application system for power grid control operations is proposed, comprising a data preparation and preprocessing module, a model building module, a model integration module, and a model training and optimization module.

[0010] The data preparation and preprocessing module acquires data from the system, organizes the collected data, cleans, standardizes, and segments the data to improve the model's ability to understand text and the effectiveness of subsequent processing.

[0011] The model building module loads the pre-trained BERT model to obtain the contextual information of the text and generate word vectors, and designs DPCNN convolutional blocks to extract deep text features, thereby expanding the feature extraction capabilities of the BERT model.

[0012] The model integration module connects the last hidden state output by BERT with the input of DPCNN to form a coherent data flow. The convolutional blocks and pooling layers of DPCNN process the data in sequence, extract deep features, and map the features onto classification labels through fully connected layers for text classification.

[0013] The model training and optimization module uses word vectors obtained from BERT training to perform forward and backward propagation, dynamically adjusts the learning rate and optimizer parameters, improves the training strategy by monitoring the loss function and accuracy, and uses the softmax function to output the probability of each category, providing a basis for decision-making.

[0014] As a preferred embodiment of the AI-based power grid control business auxiliary office application system of the present invention, the data preparation and preprocessing module includes acquiring three types of data through the system, including maintenance application data, business rule data, and annotation data.

[0015] The maintenance request data includes maintenance request data obtained from the OMS system, including operation and maintenance policy approval opinions, power outage scope, and key information on affected equipment. The maintenance request data is then input into the pre-trained BERT model in the model building module for training and testing.

[0016] The business rule data includes collecting and organizing historical typical operation ticket data, new equipment commissioning and startup plans and risk warning notices, providing text samples, and inputting the business logic and rules of power grid control and operation into the pre-trained BERT model of the model building module for learning;

[0017] The labeled data includes collected and organized scheduling procedures and power grid-related labels, and the data is input into the pre-trained BERT model of the model building module for supervised learning;

[0018] The three types of data acquired will be preprocessed, including text cleaning, standardization, and word segmentation.

[0019] The text cleaning includes removing irrelevant characters and punctuation marks; the standardization includes unifying text format and units; and the word segmentation includes splitting the text into words or phrases to enable the model to understand the text semantics.

[0020] As a preferred embodiment of the AI-based power grid control business auxiliary office application system described in this invention, the model building module includes loading a BERT model from a pre-trained model library. The model defines 12 Transformer layers, 768 hidden units, 12 self-attention headers, and 110 million parameters.

[0021] The original text is segmented using a Tokenizer, and the segmentation results are mapped to a predefined embedding matrix to obtain the corresponding embedding vectors, which are then input into BERT's Transformer encoder for further processing through a self-attention mechanism.

[0022] Define the convolutional blocks of the DPCNN model, each containing 250 fixed-size convolutional kernels. Apply the DPCNN architecture to the output of the BERT model, using the last hidden state of the BERT model as the input to the DPCNN.

[0023] As a preferred embodiment of the AI-based power grid control auxiliary office application system described in this invention, the self-attention mechanism further includes generating word representations containing contextual information through the self-attention mechanism:

[0024]

[0025] Where Q, K, and V are matrices composed of input word vectors, d k It is the dimension of the key vector, used to scale the dot product result to prevent gradient vanishing or exploding.

[0026] As a preferred embodiment of the AI-based auxiliary office application system for power grid control operations described in this invention, the model combination module includes text classification and classification prediction.

[0027] The text classification involves concatenating the output of the BERT model with the input of the DPCNN, and passing the last hidden state of the BERT model to the first convolutional block of the DPCNN.

[0028] In the DPCNN architecture, convolutional blocks and pooling layers process the input data sequentially to extract deep text features. Each convolutional block contains two convolutional layers and one pooling layer. The convolutional layers use equal-length convolutions, and the pooling layers use 1 / 2 max pooling.

[0029] The output of DPCNN is mapped to classification labels through fully connected layers to complete the text classification task.

[0030] The classification prediction involves a fully connected layer converting the feature vectors extracted by DPCNN into classification probabilities for text classification prediction.

[0031] As a preferred embodiment of the AI-based power grid control auxiliary office application system described in this invention, the model training and optimization module includes: using BERT to convert each word in the text into a word vector at the input layer, concatenating them into a vector matrix, inputting the vector matrix into the BERT model, and performing forward propagation to calculate the loss;

[0032] The gradient is calculated using the backpropagation algorithm, and the parameters of the BERT model are updated. The output of the BERT model is used as the input of the DPCNN network and fed into the DPCNN network. During the training process, the BertAdam optimizer and learning rate strategy are used to adjust the training strategy.

[0033] The DPCNN network is defined as consisting of 7 convolutional blocks, each containing two convolutional layers and one pooling layer.

[0034] When generating features during the convolution process, matrix multiplication is performed using convolution kernels of equal length, and nonlinear transformation is performed using the Mish activation function.

[0035] The transformed vector is fed into the Transformer encoding layer to extract vector features, and then dynamically adjusted using the BertAdam algorithm.

[0036] The output layer uses the softmax function to assign a probability value to each classification result, representing the likelihood of each classification.

[0037] The optimal model is tested using a test dataset to evaluate its performance in real-world application scenarios.

[0038] Another objective of this invention is to provide an AI-based auxiliary office application method for power grid dispatching. This invention improves work efficiency by reducing manual intervention and rapidly processing power grid dispatching documents and applications through automated processing and intelligent analysis. Simultaneously, the method aims to enhance decision support, helping dispatchers analyze historical data more accurately and make rational decisions, thereby reducing risks and losses. Through advanced AI models (such as BERT and DPCNN), this method enhances text processing capabilities, enabling intelligent question-and-answer functionality, allowing staff to quickly obtain the information they need. Furthermore, it supports data-driven business optimization and risk management, helping to identify and assess potential risks, ensuring the safe and stable operation of the power grid, and thus promoting the intelligent transformation and development of the power industry.

[0039] As a preferred embodiment of the AI-based auxiliary office application method for power grid control operations described in this invention, it is characterized by comprising:

[0040] Data is acquired from the OMS system, historical and labeled data are collected and organized, and preprocessed; the BERT model is loaded, the DPCNN architecture is defined, and the parameters and configurations required for model training are prepared; the output of the BERT model is connected to the input of the DPCNN to build a complete BERT-DPCNN model; the BERT-DPCNN model is trained and its parameters are updated, its performance is monitored, and it is tested and evaluated.

[0041] As a preferred embodiment of the AI-based power grid control auxiliary office application method described in this invention, the training of the BERT-DPCNN model includes: using BERT to process each word in the text into a word vector at the input layer through the Transformer encoder in the semantic extraction layer; concatenating each word vector to obtain a vector matrix X, as shown in the following formula:

[0042] X 1:n =x1⊕x2⊕……⊕x n

[0043] Where n is the number of words in the sequence, and each word is encoded into a vector x1, x2, ..., xn. n , ⊕ indicates a splicing operation;

[0044] The BERT model calculates the loss through forward propagation, calculates the gradient through backpropagation, and updates the model parameters.

[0045] The CLS vector output by the BERT model and the vector of each token extracted from the question and answer sentences are used as input to the DPCNN network to fine-tune the BERT model. The BertAdam optimizer and learning rate strategy are adopted, and the training strategy is adjusted by monitoring the model's loss function and accuracy.

[0046] The DPCNN network is defined as consisting of 7 convolutional blocks, each containing two convolutional layers and one pooling layer. The convolution process generates features using convolutions of equal length, where h is the size of the convolutional kernel W, and the feature S... i Represented as:

[0047]

[0048] Among them, S i It is the value at position j in the output feature map, W i X is the weight at position i in the convolution kernel. j+i-1 It is the element in the input data corresponding to the current position of the convolution kernel, k is the size of the convolution kernel, b is the bias term, and f is the non-linear transformation function;

[0049] After obtaining the result through matrix multiplication, an activation function is applied:

[0050] f(x) = xtanh(ln(1+e) x ))

[0051] The model obtains high-level features from the data, and the transformed vectors are fed into the Transformer encoding layer to extract vector features. The BertAdam algorithm is used to dynamically adjust the learning rate, enabling the model to converge quickly in the early stages of training. The weights are adjusted in the later stages of training. The output layer uses the softmax function to assign a probability value to each classification result, representing the likelihood of each classification.

[0052]

[0053] Where yi is the output of each node, and n is the number of output nodes, i.e. the number of categories.

[0054] The optimal model is tested using a test dataset to evaluate its performance in real-world application scenarios.

[0055] A computer device includes a memory and a processor, the memory storing a computer program, characterized in that the processor executes the computer program to implement the steps of the AI-based power grid control business auxiliary office application method.

[0056] A computer-readable storage medium storing a computer program thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the AI-based power grid control business auxiliary office application method.

[0057] The beneficial effects of this invention are as follows: This invention introduces generative AI technology into the field of power dispatching. Previously, similar power grid dispatching and operation interaction projects only deployed traditional analytical AI, which analyzed and learned from existing static data to make predictions and judgments. However, traditional analytical AI has weak capabilities in understanding and processing natural language, resulting in many inconveniences and limitations when used for analysis; furthermore, its ability to generate content and make decisions is relatively limited, making it unable to adapt to the complex and ever-changing needs of power grid operation. Generative AI technology, on the other hand, possesses powerful natural language understanding, content generation, and innovation capabilities, which can meet the complex needs of various business processes in the dispatching domain, greatly improving the work efficiency of the entire power grid dispatching and operation business, and further promoting the digital transformation of enterprises and the construction of digital power grids.

[0058] This project integrates and coordinates multiple existing production management systems across the power grid dispatch domain. Previous similar projects focused only on improving or optimizing specific business processes or existing production management systems for power grid control and operation. These systems remained relatively independent, resulting in significant information asymmetry and impacting work efficiency. This project, by introducing a generative AI large-scale model interface, coordinates and integrates multiple existing production management systems across the dispatch domain. This enables intelligent integration and processing of various dispatch operation information, significantly reducing the paperwork burden on dispatchers and allowing them to focus more on specialized tasks such as power grid operation and accident handling.

[0059] This project complements the integrated control and dispatching system and intelligent dispatching error prevention technology. Previous projects related to integrated control and dispatching and intelligent dispatching error prevention focused primarily on the technical error prevention and system construction of core power grid control and dispatching operations, such as dispatching orders and equipment operation. However, less attention was paid to potential risks of misoperation and misdispatch in other auxiliary control and dispatching processes. For example, errors in dispatching operation tickets, omissions during manual review, or incomplete dispatching logs leading to missing important information could result in misdispatch. This project, by deploying a generative AI system, intelligently generates various dispatching and dispatching information and records, effectively avoiding potential power grid operational risks and hazards such as misdispatch and misoperation caused by errors in dispatching and dispatching information. It also further reduces the workload of dispatchers, allowing them to focus more on power grid control and dispatching operations, thus further complementing the integrated control and dispatching system and intelligent dispatching error prevention technology. Attached Figure Description

[0060] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0061] Figure 1 A schematic diagram of a module of an AI-based auxiliary office application system for power grid control operations, provided as an embodiment of the present invention.

[0062] Figure 2 The BERT_DPCNN overall model is provided as an embodiment of the auxiliary office application method for power grid control based on AI technology.

[0063] Figure 3 The flowchart illustrates the training and optimization process of the BERT_DPCNN model for an AI-based auxiliary office application method for power grid control, as provided in one embodiment of the present invention. Detailed Implementation

[0064] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0065] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0066] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it an embodiment that is mutually exclusive, either alone or selectively, with other embodiments.

[0067] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0068] In the description of the present invention, it should be noted that the terms "upper, lower, inner, and outer" and other references to orientations or positional relationships are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first, second, or third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0069] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0070] Example 1, referring to Figure 1-Figure 2This is the first embodiment of the present invention. This embodiment provides an auxiliary office application system for power grid control business based on AI technology, including a data preparation and preprocessing module 10, a model building module 20, a model combining module 30, and a model training and optimization module 40.

[0071] The data preparation and preprocessing module 10 acquires data from the system, organizes the collected data, cleans, standardizes, and segments the data to improve the model's ability to understand text and the effectiveness of subsequent processing.

[0072] The model building module 20 loads the pre-trained BERT model to obtain the context information of the text and generate word vectors, and designs DPCNN convolutional blocks to extract deep text features, thereby expanding the feature extraction capability of the BERT model.

[0073] The model, combined with module 30, connects the last hidden state output by BERT with the input of DPCNN to form a coherent data flow. The convolutional blocks and pooling layers of DPCNN process the data in sequence, extract deep features, and map the features onto classification labels through fully connected layers for text classification.

[0074] The model training and optimization module 40 uses word vectors obtained from BERT training to perform forward and backward propagation, dynamically adjusts the learning rate and optimizer parameters, improves the training strategy by monitoring the loss function and accuracy, and uses the softmax function to output the probability of each category, providing a basis for decision-making.

[0075] The overall framework of the BERT-DPCNN model is as follows: Figure 2 As shown, the short-circuit connection plays a crucial role. It is responsible for adding the tensor before pooling to the tensor of the current layer, and then continuously feeding this fused result into the recurring structure of the model for downsampling. This process not only helps the model to mine text features more deeply, but also enhances the model's ability to process complex text information through feature fusion, providing strong technical support for intelligent text processing in the field of power grid dispatching and operation. Its implementation principle is as follows:

[0076] Furthermore, the data preparation and preprocessing module 10 acquires three types of data through the system, including maintenance application data, business rule data, and annotation data;

[0077] The maintenance request data, including operational policy approvals, outage scope, and affected equipment, will be retrieved from the OMS system. This data will be used for model training and testing.

[0078] We collected and organized historical typical operation ticket data, new equipment commissioning plans, and risk warning notices. This data provides rich text samples, which helps the model learn the business logic and rules of power grid control and operation.

[0079] We will collect and organize relevant dispatching procedures and power grid-related annotations. This annotated data will be used for supervised learning of the model to improve its predictive accuracy.

[0080] The three types of data acquired will be preprocessed, including text cleaning, standardization, and word segmentation.

[0081] The text cleaning includes removing irrelevant characters and punctuation marks.

[0082] The standardization includes unifying text formats and units.

[0083] The word segmentation involves breaking down text into words or phrases, enabling the model to understand the semantics of the text.

[0084] It should be noted that the model building module 20 first prepares the BERT model by loading it from a pre-trained model library. The model definition includes 12 Transformer layers, 768 hidden units, and 12 self-attributed layers. - The tion header contains 110 million parameters.

[0085] BERT employs a self-attention mechanism to capture the contextual semantic information of sentences. Before being input into BERT, the original text is segmented using a tokenizer, and these segments are mapped to a predefined embedding matrix to obtain corresponding embedding vectors. These embedding vectors are then fed into BERT's Transformer encoder, where they are further processed through the self-attention mechanism to generate word representations that include contextual information. The self-attention calculation formula is as follows:

[0086]

[0087] Where Q, K, and V are matrices composed of input word vectors, d k It is the dimension of the key vector, used to scale the dot product result to prevent gradient vanishing or exploding.

[0088] Next, we define the DPCNN architecture:

[0089] Define convolutional blocks for the DPCNN model, each containing 250 fixed-size convolutional kernels for extracting text features. The size and number of convolutional kernels can be adjusted according to actual needs.

[0090] The DPCNN architecture is applied to the output of the BERT model, using the last hidden state of the BERT model as the input to DPCNN. This allows DPCNN to further extract deep features from the contextual word representations generated by BERT.

[0091] It should also be noted that the model integration module 30 includes text classification and classification prediction;

[0092] The output of the BERT model is concatenated with the input of the DPCNN, passing the last hidden state of the BERT model to the first convolutional block of the DPCNN.

[0093] In the DPCNN architecture, convolutional blocks and pooling layers process the input data sequentially to extract deeper text features. Each convolutional block contains two convolutional layers and one pooling layer. The convolutional layers use equal-length convolutions, and the pooling layer uses 1 / 2 max pooling.

[0094] The output of DPCNN is mapped onto classification labels through fully connected layers to complete the text classification task.

[0095] The fully connected layer converts the feature vectors extracted by DPCNN into classification probabilities, thereby enabling text classification prediction.

[0096] Furthermore, the model training and optimization module 40 uses BERT to convert each word in the text into a word vector at the input layer, and then concatenates them into a vector matrix. The vector matrix is ​​then input into the BERT model for forward propagation to calculate the loss.

[0097] The gradient is calculated using the backpropagation algorithm, and the parameters of the BERT model are updated. This step optimizes the performance of the BERT model, enabling it to better extract semantic features from the text.

[0098] The output of the BERT model is used as the input to the DPCNN network, and both are fed into the DPCNN network. During training, the BertAdam optimizer and learning rate strategy are used to monitor the model's loss function and accuracy, and the training strategy is adjusted accordingly.

[0099] The DPCNN network is defined as consisting of 7 convolutional blocks, each containing two convolutional layers and one pooling layer. Within each block, the two convolutional layers use equal-length convolutions with a kernel size of 250. The pooling layer uses half-pooling with size=3 and stride=2.

[0100] When generating features during the convolution process, matrix multiplication is performed using convolution kernels of equal length, and a non-linear transformation is applied using the Mish activation function. Compared to the traditional ReLU activation function, the Mish activation function can more accurately capture key information in the text while effectively mitigating the gradient explosion problem.

[0101] The transformed vectors are fed into the Transformer encoding layer to extract vector features, and the BertAdam algorithm is used to dynamically adjust the learning rate. In the early stages of training, the learning rate is higher to allow the model to converge quickly; in the later stages of training, the learning rate is gradually reduced to allow the model to fine-tune the weights and avoid overfitting.

[0102] The output layer uses the softmax function to assign a probability value to each classification result, representing the likelihood of each classification. In this way, the model can output the probability distribution of the predicted results for user reference and decision-making.

[0103] The optimal model is tested using a test dataset to evaluate its performance in real-world applications. The test dataset should contain samples similar to but different from the training dataset to ensure the model's generalization ability.

[0104] The BERT_DPCNN model combines the advantages of BERT and DPCNN to improve the accuracy and efficiency of text classification. BERT, based on the Transformer architecture, enhances text representation capabilities by capturing dependencies between words through bidirectional context. DPCNN, a deep convolutional network specifically designed for text classification, effectively extracts long-distance text dependencies. In BERT_DPCNN, text is first encoded by BERT, fully utilizing its bidirectional modeling capabilities to obtain rich text embeddings. Subsequently, DPCNN performs deep feature extraction on these embeddings to further capture complex relationships within the text. Finally, a classification layer categorizes the processed features, outputting the probability of the text belonging to a category. This combination allows BERT_DPCNN to leverage both BERT's contextual understanding capabilities and DPCNN's advantages in deep feature extraction, significantly improving text comprehension. This project utilizes the BERT_DPCNN model, incorporating specialized terminology and corpus from the power grid dispatching domain, to implement auxiliary question-answering functionality in power grid dispatching, improving office efficiency.

[0105] Example 2 is the second embodiment of the present invention, which provides an auxiliary office application system for power grid control based on AI technology. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0106] In the experiment, the training data and environment set parameters were set according to Tables 1, 2 and 3. The epoch ranged from 1 to 10000. The training loss and validation loss of the model gradually decreased, while the training accuracy and validation accuracy gradually increased. This shows the optimization effect of the model during the training process.

[0107] Table 1: Experimental Data

[0108] Total number of data training episodes Verification episodes Test episodes 28,200 27,000 0.06 million 0.06 million

[0109] Table 2: Experimental Environment

[0110] Related systems Configuration operating system Windows 11 processor 12thGenIntel(R)Core(TM)i7-12700H2.30GHz Memory 40.0GB language Python 3.8.10 frame 2.4.0+ CPU

[0111] Table 3: Experimental Parameter Settings

[0112] parameter Configuration epoch 10000 Sentence length (fill in short sentences, cut long sentences) 32 Attention 128 Learning rate 5e-5 Number of convolution kernels Python 3.8.10 Hidden layer dimensions 768

[0113] Table 4: Optimized Training Results

[0114] epoch Training loss Training accuracy Verify loss Verification accuracy 1 2.31 10.12％ 2.4 8.01％ 2 2.12 19.89％ 2.2 13.1％ 3 1.28 25.3％ 2.05 22.36％ ... ... ... ... ... 9999 0.5 95.99％ 1.12 90.32％ 10000 0.49 96.01％ 1.1 91.03％

[0115] Example 3, the third embodiment of the present invention, differs from the previous two embodiments in that:

[0116] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present 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 described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0118] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0119] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0120] Example 4, refer to Figure 3 This is the fourth embodiment of the present invention, which provides a method for auxiliary office application of power grid control based on AI technology, including:

[0121] like Figure 3 As shown, the BERT model, DPCNN model, the connections between them, and the model training and optimization process are illustrated.

[0122] Data is acquired from the OMS system, historical data and labeled data are collected and organized, and preprocessing operations are performed.

[0123] Load the BERT model, define the DPCNN architecture, and prepare the parameters and configurations required for model training.

[0124] The output of the BERT model is concatenated with the input of the DPCNN to construct the complete BERT-DPCNN model.

[0125] Model training, parameter updates, performance monitoring, and testing and evaluation.

[0126] Furthermore, the model training and optimization module:

[0127] First, BERT is used to transform each word in the text into a word vector at the input layer. This process is handled by the Transformer encoder in the semantic extraction layer. Then, each word vector is concatenated to obtain a vector matrix X, as shown in the following formula:

[0128] X 1:n =x1⊕x2⊕……⊕x n

[0129] Where n is the number of words in the sequence, and each word is encoded into a vector x1, x2, ..., xn. n , ⊕ denotes a concatenation operation (in a matrix, this usually means stacking these vectors row by row).

[0130] The BERT model calculates the loss through forward propagation, calculates the gradient through backpropagation, and updates the model parameters.

[0131] The output of the BERT model is used as input to the DPCNN network. In addition to the [CLS] vector from the BERT output, vector representations of each token in the question and answer sentences are extracted. These vectors are fed jointly into the DPCNN network for fine-tuning along with the BERT model using the BertAdam optimizer and learning rate policy. Metrics such as the model's loss function and accuracy are monitored, and the training strategy is adjusted accordingly.

[0132] The DPCNN network is defined as consisting of 7 convolutional blocks, each containing two convolutional layers and one pooling layer. Within each block, the two convolutional layers use equal-length convolutions with a kernel size of 250. The pooling layer uses half-pooling with size=3 and stride=2. This design compresses the sequence length to half its original length after each convolutional block, thus doubling the range of perceptible text segments.

[0133] The convolution process generates features using convolutions of equal length. h is the size of the convolution kernel W, and the feature S... i It can be expressed by the following formula:

[0134]

[0135] Among them, S i It is the value at position j in the output feature map, W i X is the weight at position i in the convolution kernel. j+i-1 is the element in the input data corresponding to the current position of the convolution kernel, k is the size of the convolution kernel, b is the bias term, and f is the nonlinear transformation function.

[0136] Its process involves first performing matrix multiplication, then applying an activation function to the result.

[0137] f(x) = xtanh(ln(1+e) x ))

[0138] A nonlinear transformation is performed, followed by a max pooling operation.

[0139] After the above steps, the model obtains high-level features of the data.

[0140] The transformed vector is then fed into the Transformer encoding layer to extract vector features. The BertAdam algorithm is used to dynamically adjust the learning rate, enabling the model to converge quickly in the early stages of training and finely adjust the weights in the later stages to avoid overfitting. Finally, the output layer assigns a probability value to each classification result using the softmax function, representing the likelihood of each classification. The function is as follows:

[0141]

[0142] Where yi is the output of each node, and n is the number of output nodes, i.e. the number of categories.

[0143] Finally, the optimal model is tested using a test dataset to evaluate its performance in real-world application scenarios.

[0144] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-based auxiliary office application system for power grid dispatching, characterized in that: It includes modules for data preparation and preprocessing, model building, model integration, and model training and optimization. The data preparation and preprocessing module acquires data from the system, organizes the collected data, cleans, standardizes, and segments the data to improve the model's ability to understand text and the effectiveness of subsequent processing. The model building module loads the pre-trained BERT model to obtain the contextual information of the text and generate word vectors, and designs DPCNN convolutional blocks to extract deep text features, thereby expanding the feature extraction capabilities of the BERT model. The model integration module connects the last hidden state output by BERT with the input of DPCNN to form a coherent data flow. The convolutional blocks and pooling layers of DPCNN process the data in sequence, extract deep features, and map the features onto classification labels through fully connected layers for text classification. The model training and optimization module uses word vectors obtained from BERT training to perform forward and backward propagation, dynamically adjusts the learning rate and optimizer parameters, improves the training strategy by monitoring the loss function and accuracy, and uses the softmax function to output the probability of each category, providing a basis for decision-making. The data preparation and preprocessing module includes acquiring three types of data through the system: maintenance application data, business rule data, and annotation data. The maintenance request data includes maintenance request data obtained from the OMS system, including operation and maintenance policy approval opinions, power outage scope, and key information on affected equipment. The maintenance request data is then input into the pre-trained BERT model in the model building module for training and testing. The business rule data includes collecting and organizing historical typical operation ticket data, new equipment commissioning and startup plans and risk warning notices, providing text samples, and inputting the business logic and rules of power grid control and operation into the pre-trained BERT model of the model building module for learning; The labeled data includes collected and organized scheduling procedures and power grid-related labels, and the data is input into the pre-trained BERT model of the model building module for supervised learning; The three types of data acquired will be preprocessed, including text cleaning, standardization, and word segmentation. The text cleaning includes removing irrelevant characters and punctuation marks; the standardization includes unifying text format and units; and the word segmentation includes splitting the text into words or phrases to enable the model to understand the text semantics. The model building module includes loading a BERT model from a pre-trained model library. The model defines 12 Transformer layers, 768 hidden units, 12 self-attention headers, and 110 million parameters. The original text is segmented using a Tokenizer, and the segmentation results are mapped to a predefined embedding matrix to obtain the corresponding embedding vectors, which are then input into BERT's Transformer encoder for further processing through a self-attention mechanism. Define the convolutional blocks of the DPCNN model, each containing 250 fixed-size convolutional kernels. Apply the DPCNN architecture to the output of the BERT model, using the last hidden state of the BERT model as the input to the DPCNN. The model integration module includes text classification and classification prediction; The text classification involves concatenating the output of the BERT model with the input of the DPCNN, and passing the last hidden state of the BERT model to the first convolutional block of the DPCNN. In the DPCNN architecture, convolutional blocks and pooling layers process the input data sequentially to extract deep text features. Each convolutional block contains two convolutional layers and one pooling layer. The convolutional layers use equal-length convolutions, and the pooling layers use 1 / 2 max pooling. The output of DPCNN is mapped to classification labels through fully connected layers to complete the text classification task. The classification prediction is performed by converting the feature vectors extracted by DPCNN into classification probabilities through a fully connected layer to perform text classification prediction. The model training and optimization module includes using BERT to convert each word in the text into a word vector at the input layer, concatenating them into a vector matrix, inputting the vector matrix into the BERT model, and performing forward propagation to calculate the loss. The gradient is calculated using the backpropagation algorithm, and the parameters of the BERT model are updated. The output of the BERT model is used as the input of the DPCNN network and fed into the DPCNN network. During the training process, the BertAdam optimizer and learning rate strategy are used to adjust the training strategy. The DPCNN network is defined as consisting of 7 convolutional blocks, each containing two convolutional layers and one pooling layer. When generating features during the convolution process, matrix multiplication is performed using convolution kernels of equal length, and nonlinear transformation is performed using the Mish activation function. The transformed vector is fed into the Transformer encoding layer to extract vector features, and then dynamically adjusted using the BertAdam algorithm. The output layer uses the softmax function to assign a probability value to each classification result, representing the likelihood of each classification. The optimal model is tested using a test dataset to evaluate its performance in real-world application scenarios.

2. The AI-based auxiliary office application system for power grid control as described in claim 1, characterized in that: The self-attention mechanism further includes generating word representations that contain contextual information: Where Q, K, and V are matrices composed of input word vectors, d k It is the dimension of the key vector, used to scale the dot product result to prevent gradient vanishing or exploding.

3. A method for using an AI-based auxiliary office application system for power grid control operations as described in any one of claims 1 to 2, characterized in that: Data is acquired from the OMS system, historical data and labeled data are collected and organized, and preprocessing operations are performed. Load the BERT model, define the DPCNN architecture, and prepare the parameters and configurations required for model training; The output of the BERT model is concatenated with the input of the DPCNN to construct the complete BERT-DPCNN model; Perform training and parameter updates, performance monitoring, and testing and evaluation of the BERT-DPCNN model.

4. The AI-based auxiliary office application method for power grid dispatching as described in claim 3, characterized in that: The training of the BERT-DPCNN model includes using BERT to process each word in the text into a word vector in the semantic extraction layer through the Transformer encoder, and concatenating each word vector to obtain a vector matrix X, as shown in the following formula: Where n is the number of words in the sequence, and each word is encoded into a vector x1, x2, ..., xn. n , Indicates a splicing operation; The BERT model calculates the loss through forward propagation, calculates the gradient through backpropagation, and updates the model parameters. The CLS vector output by the BERT model and the vector of each token extracted from the question and answer sentences are used as input to the DPCNN network to fine-tune the BERT model. The BertAdam optimizer and learning rate strategy are adopted, and the training strategy is adjusted by monitoring the model's loss function and accuracy. The DPCNN network is defined as consisting of 7 convolutional blocks, each containing two convolutional layers and one pooling layer. The convolution process generates features using convolutions of equal length, where h is the size of the convolutional kernel W, and the feature S... i Represented as: Among them, S i It is the value at position j in the output feature map, W i X is the weight at position i in the convolution kernel. j+i-1 It is the element in the input data corresponding to the current position of the convolution kernel, k is the size of the convolution kernel, b is the bias term, and f is the non-linear transformation function; After obtaining the result through matrix multiplication, an activation function is applied: f(x)=xtanh(ln(1+e x )) The model obtains high-level features from the data, and the transformed vectors are fed into the Transformer encoding layer to extract vector features. The BertAdam algorithm is used to dynamically adjust the learning rate, enabling the model to converge quickly in the early stages of training. The weights are adjusted in the later stages of training. The output layer uses the softmax function to assign a probability value to each classification result, representing the likelihood of each classification. Among them, y i For each node's output, n is the number of output nodes, which is the number of categories. The optimal model is tested using a test dataset to evaluate its performance in real-world application scenarios.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the AI-based auxiliary office application method for power grid control business as described in any one of claims 3 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the AI-based auxiliary office application method for power grid control business as described in any one of claims 3 to 4.

Images