Operation and maintenance method, device and equipment of low-voltage power distribution system and storage medium

Abstract

The invention discloses an operation and maintenance method, device and equipment of a low-voltage power distribution system and a storage medium, and relates to the technical field of operation and maintenance of power systems. The method comprises the following steps: acquiring a natural language input by a user based on interaction equipment, determining an operation and maintenance entity and a query intention of the natural language as statement features, and determining original time sequence data in the low-voltage power distribution system according to the query features; determining data feature information corresponding to the original time sequence data and a data exception result; determining a diagnosis result corresponding to the natural language based on the data exception result and the query feature; based on a preset language model, operation and maintenance suggestions corresponding to the natural language are determined according to the data feature information, and the diagnosis result and the operation and maintenance suggestions are fed back to the interaction device so that the interaction device can visually display the diagnosis result and the operation and maintenance suggestions. Automatic operation and maintenance are achieved, manual intervention is reduced, and operation and maintenance efficiency is improved; and meanwhile, more visual and efficient man-machine interaction is realized through the interaction equipment, and the accuracy of operation and maintenance suggestions is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system operation and maintenance, and in particular to an operation and maintenance method, device, equipment and storage medium for a low-voltage power distribution system. Background Art

[0002] As the scale of electrical equipment within building systems continues to expand, the operation and maintenance of low-voltage distribution systems is becoming increasingly complex. Traditional O&M methods rely primarily on manual inspections and empirical judgment, but manual inspections are inefficient. For example, faults are diagnosed by regularly inspecting the equipment's appearance and recording operating parameters. However, manual inspections have the following drawbacks: they are inefficient, with a single inspection taking hours to days and requiring significant human resources; they are prone to missing potential issues, such as minor voltage fluctuations or early signs of faults; and they struggle to process massive amounts of operational data. Operations and maintenance personnel often make decisions based on experience, lacking data support, resulting in low diagnostic accuracy.

[0003] While automation technologies have advanced in recent years, such as fixed-rule monitoring systems or simple data collection tools, these technologies still have limitations. They lack natural language interaction capabilities, requiring users to operate through complex interfaces or specialized commands, making them difficult to meet the needs of non-professional users. Furthermore, they cannot dynamically analyze complex faults based on user queries, and diagnostic results lack interpretability. Furthermore, existing technologies recommend generating multi-dependent templates, making them difficult to adapt to diverse O&M scenarios. Therefore, achieving automated and intelligent O&M of low-voltage distribution systems based on user needs has become a pressing technical challenge. Summary of the Invention

[0004] The present invention provides an operation and maintenance method, device, equipment and storage medium for a low-voltage power distribution system, so as to solve the problem in the prior art that the low-voltage power distribution system cannot be automatically operated and maintained according to user needs.

[0005] According to one aspect of the present invention, a method for operating and maintaining a low-voltage power distribution system is provided, wherein the method comprises:

[0006] Acquire natural language input by a user based on an interactive device, determine the operation and maintenance entity and query intent of the natural language as sentence features, and determine original time series data in the low-voltage power distribution system based on the query features;

[0007] Determining data feature information and data anomaly results corresponding to the original time series data;

[0008] Determining a diagnosis result corresponding to the natural language based on the data abnormality result and the query feature;

[0009] Based on a preset language model, the operation and maintenance suggestions corresponding to the natural language are determined according to the data feature information, and the diagnosis results and the operation and maintenance suggestions are fed back to the interactive device so that the interactive device can visually display the diagnosis results and the operation and maintenance suggestions.

[0010] According to another aspect of the present invention, an operation and maintenance device for a low-voltage power distribution system is provided, wherein the device comprises:

[0011] A data acquisition module is used to acquire natural language input by a user based on an interactive device, determine the operation and maintenance entity and query intent of the natural language as sentence features, and determine the original time series data in the low-voltage power distribution system based on the query features;

[0012] A result determination module is used to determine the data feature information and data anomaly results corresponding to the original time series data;

[0013] A diagnosis result determination module, configured to determine a diagnosis result corresponding to the natural language based on the data abnormality result and the query feature;

[0014] An operation and maintenance suggestion determination module is used to determine the operation and maintenance suggestion corresponding to the natural language according to the data feature information based on a preset language model, and to feed back the diagnosis result and the operation and maintenance suggestion to the interactive device so that the interactive device can visually display the diagnosis result and the operation and maintenance suggestion.

[0015] According to another aspect of the present invention, an electronic device is provided, comprising:

[0016] at least one processor; and

[0017] a memory communicatively connected to the at least one processor; wherein,

[0018] The memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the operation and maintenance method of a low-voltage power distribution system described in any embodiment of the present invention.

[0019] According to another aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement an operation and maintenance method of a low-voltage power distribution system according to any embodiment of the present invention when executed.

[0020] The technical solution of the embodiment of the present invention realizes interaction with the user by obtaining the natural language input by the user based on the interactive device, determining the operation and maintenance entity and query intention of the natural language as sentence features; determining the original time series data in the low-voltage power distribution system according to the query features, determining the data feature information and data abnormality results corresponding to the original time series data, determining the diagnosis result corresponding to the natural language based on the data abnormality results and the query features, determining the operation and maintenance suggestions corresponding to the natural language according to the data feature information based on the preset language model, and feeding back the diagnosis result and operation and maintenance suggestions to the interactive device, so that the interactive device can visually display the diagnosis result and operation and maintenance suggestions, realize automated and intelligent operation and maintenance, reduce manual intervention, and improve operation and maintenance efficiency; at the same time, more intuitive and efficient human-computer interaction is realized through the interactive device, and the accuracy of operation and maintenance suggestions is improved.

[0021] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present invention, nor is it intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 This is a flow chart of an operation and maintenance method for a low-voltage power distribution system provided according to the first embodiment of the present invention;

[0024] Figure 2 This is a flow chart of an operation and maintenance method for a low-voltage power distribution system provided according to a second embodiment of the present invention;

[0025] Figure 3 This is a flow chart of an operation and maintenance method for a low-voltage power distribution system provided according to a third embodiment of the present invention;

[0026] Figure 4 This is a network topology diagram of a data collection module provided according to the third embodiment of the present invention;

[0027] Figure 5 This is a flowchart of a data transmission process provided according to Embodiment 3 of the present invention;

[0028] Figure 6 This is a flowchart of training a large language model according to the third embodiment of the present invention;

[0029] Figure 7This is a workflow diagram of an active interaction module provided according to the third embodiment of the present invention;

[0030] Figure 8 This is a workflow diagram of an intelligent decision-making module provided according to the third embodiment of the present invention;

[0031] Figure 9 is a timing diagram of an operation and maintenance method for a low-voltage power distribution system provided according to a third embodiment of the present invention;

[0032] Figure 10 2 is a schematic structural diagram of an operation and maintenance device for a low-voltage power distribution system provided according to a fourth embodiment of the present invention;

[0033] Figure 11 It is a structural diagram of an electronic device for implementing an operation and maintenance method for a low-voltage power distribution system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0035] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0036] Example 1

[0037] Figure 1 This is a flow chart of a method for operating and maintaining a low-voltage power distribution system according to the first embodiment of the present invention. This embodiment is applicable to operating and maintaining a low-voltage power distribution system. The method can be executed by an operation and maintenance device for a low-voltage power distribution system. The low-voltage power distribution system device can be implemented in the form of hardware and / or software. The operation and maintenance device for the low-voltage power distribution system can be configured in an electronic device. Figure 1As shown, the method includes:

[0038] S110: Obtain natural language input by the user based on the interactive device, determine the operation and maintenance entity and query intention of the natural language as sentence features, and determine the original time series data in the low-voltage power distribution system based on the query features.

[0039] The interactive device can be a device that uses natural language processing technology to enable intelligent dialogue with operation and maintenance personnel. It can be a smart terminal that supports natural language processing, such as a tablet equipped with a microphone and touch screen, or a voice interaction device. It can receive user input via voice or text and intelligently interact with operation and maintenance personnel based on multiple rounds of dialogue. In actual operation, the interactive device can recognize the natural language input of the user and extract characteristic information from the natural language as sentence features. Generally speaking, the natural language input of the user can be multi-round, and the interactive device can directly interact with the user in a corresponding dialogue. The interactive device can determine the user's required operation and maintenance entity and query intent. Sentence features refer to the characteristic information contained in natural language, such as operation and maintenance entities and query intent. Sentence features can be extracted based on natural language processing technology. Operation and maintenance entities refer to various specific objects or concepts involved in the operation and maintenance management process. Operation and maintenance entities can include at least the device name, fault type, or time range, such as the device name is "distribution cabinet No. 1" and the fault type is "overload". Query intent refers to the information content that the user wishes to obtain when performing an information query. For example, query intent may include but is not limited to fault diagnosis and device status query. Raw time series data refers to operational data of maintenance entities within a period of time extracted based on query features. This period can include the past hour, past two hours, or past 24 hours. For example, raw time series data may include, but is not limited to, information such as voltage, current, and temperature. During actual operation, operational data of maintenance entities can be collected and stored in real time using pre-set sensors, allowing for timely extraction of raw time series data when query features are captured.

[0040] In an embodiment, natural language input by a user can be received through an interactive device, and operation and maintenance entities and query intents in the natural language can be extracted through natural language processing technology. The operation data of the operation and maintenance entities in the low-voltage power distribution system over a period of time can be extracted as original time series data according to the operation and maintenance entities and query intents. In the actual operation process, a pre-trained deep learning intent classification model can be used, a bidirectional long-short term memory network (Bi-LSTM) combined with an attention mechanism can be used to extract the query intent contained in the natural language, and a conditional random field (CRF) combined with a neural network sequence labeling model can be used to extract the operation and maintenance entities contained in the natural language, and the query intent and operation and maintenance entities are used as sentence features.

[0041] Alternatively, the natural language can be matched with pre-set intent words and operation and maintenance entity words to extract the query intent and operation and maintenance entity in the natural language. Generally speaking, when it is determined that the operation and maintenance entity or the query intent is not detected, a dialogue interaction can be conducted with the user to generate problem tracking information so that the user can supplement the operation and maintenance entity and / or query intent according to the problem tracking information. After the query intent and operation and maintenance entity are determined, the operation and maintenance entity operation data associated with the operation and maintenance entity and the query intent can be extracted as the original time series data. In one embodiment, when the natural language input by the user is "check whether the current of distribution cabinet No. 1 is normal", it can be parsed and determined that the operation and maintenance entity is "distribution cabinet No. 1, current" and the query intent is "status query". According to the operation and maintenance entity and the query intent, the operation and maintenance entity operation data within a preset time period (such as the past 24 hours) is extracted from the preset sensors (such as voltage sensors, current sensors) of the low-voltage power distribution system as the original time series data. The original time series data includes voltage (unit: volt), current (unit: ampere) and temperature (unit: Celsius), and the sampling frequency is 1 time per minute.

[0042] S120: Determine data feature information and data anomaly results corresponding to the original time series data.

[0043] Data feature information refers to feature data determined based on the original time series data. In actual operations, the token sequence corresponding to the original time series data can be used as data feature information. Data anomaly results indicate whether there are anomalies in the original time series data. Generally speaking, data anomaly results can include both normal data and abnormal data.

[0044] In an embodiment, the original time series data can be subjected to data preprocessing such as data source cleaning and standardization to obtain the target time series data, so as to facilitate data fusion of the target time series data, and then the data anomaly results of the target sequence data are determined according to the preset operation data interval threshold. Generally speaking, missing values, outliers and duplicate values ​​in the original time series data can be processed. For missing values, deletion, interpolation and other methods can be used to process them; for duplicate values, deletion operations can be performed directly; for outliers, statistical methods (such as Z-score) or machine learning methods can be used for detection and processing. In actual operation, when the target sequence data includes indicators such as current, voltage and temperature, each indicator has a normal data interval threshold. When a certain target sequence data exceeds the normal data interval threshold, it can be determined that there is an anomaly in the target sequence data, and the data anomaly result is determined to be data anomaly; when all the target sequence data are within the normal data interval threshold, the data anomaly result can be determined to be data normal. At the same time, the target sequence data can be split according to time, the operation and maintenance entity data corresponding to each time is determined, and the data categories contained in each operation and maintenance entity operation data and the data corresponding to each data category are determined, and each data is discretized to obtain data features. For example, data categories can be divided into several main categories: measurement values, status, events, and time. The data corresponding to each data category can be discretized to obtain data features. For example, measurement values ​​can be mapped to "very low," "low," "normal," "high," "very high," etc.; status can be mapped to "running," "stopped," "standby," "fault," etc.; events can be mapped to "start," "shutdown," "overload," "short circuit," etc.; time can remain in numerical form but incorporate Chinese units such as "hour," "minute," and "second." The data features are then mapped into a fixed-dimensional vector space to obtain a token sequence of the operation and maintenance entity's operational data corresponding to each time. The token sequence is then concatenated in chronological order as the data feature information.

[0045] In practical applications, since the target sequence data comes from different sources, a graph neural network (GNN) can be used to establish relationships between the data and map the data from different data sources into a unified feature space. Specifically, the different data sources (such as current, voltage, and temperature) in the target sequence data at the same time can be used as nodes in the graph. Nodes at adjacent times are connected to form directed edges. The adjacency matrix of the graph is generated based on the constructed edges. The token vector corresponding to each node is determined through the graph neural network. The token vectors corresponding to all nodes are arranged in a certain order to obtain a token sequence as the data feature information.

[0046] S130: Determine a diagnosis result corresponding to the natural language based on the data anomaly result and the query feature.

[0047] The diagnostic result refers to the diagnostic status of the operation and maintenance entity corresponding to the natural language. In the actual operation process, the diagnostic status may include normal or abnormal. In one embodiment, the diagnostic result may also include specific diagnostic status, such as the operating status or fault type of the operation and maintenance entity.

[0048] In an embodiment, when it is determined that the data anomaly result is normal, the diagnosis result is determined to be that the operation and maintenance entity in the query feature is normal; when it is determined that the data anomaly result is data anomaly, the abnormal data in the target sequence data is extracted, and the diagnosis result corresponding to the natural language is determined according to the abnormal data and the operation and maintenance entity. In the actual operation process, the data associated with the abnormal data and the operation and maintenance entity can be extracted from the knowledge base to form a diagnosis result; or, a preset knowledge graph matching can be used to determine the diagnosis result corresponding to the natural language. In the preset knowledge graph, entity types such as operation and maintenance entities, fault conditions and fault causes can be constructed, and the relationships between entities such as belonging to, causing, etc. can be defined, and the entities and relationships extracted from the text data in the knowledge base are converted into a preset knowledge graph.

[0049] In actual operations, the abnormal data corresponding to the data anomaly result can be determined. The associated paths can be searched in the pre-set knowledge graph based on the operation and maintenance entities and the abnormal data, and the corresponding diagnostic results can be inferred in natural language. Alternatively, methods such as Bayesian network probabilistic reasoning can be used to determine the probability of each fault, and the fault with the highest probability can be used as the diagnostic result. Alternatively, a fault tree model can be pre-established and used to infer the diagnostic results based on the abnormal data and operation and maintenance entities.

[0050] S140. Based on the preset language model, determine the operation and maintenance suggestions corresponding to the natural language according to the data feature information, and feed back the diagnosis results and operation and maintenance suggestions to the interactive device so that the interactive device can visually display the diagnosis results and operation and maintenance suggestions. Among them, the preset language model is a large language model based on the Transformer architecture, which is used to predict the most likely token sequence as the operation and maintenance suggestions through the data feature information. In the actual operation process, the preset language model adopts a three-stage language model training strategy: the first stage uses general text data (such as engineering literature) to perform masked language model (MLM) tasks and next sentence prediction tasks (NSP) tasks; the second stage introduces power system domain corpus, distribution system fault chain prediction (FCP) tasks and equipment state inference (ESI) tasks; the third stage fine-tunes and optimizes the operation and maintenance suggestions based on the actual distribution system historical operation and maintenance data. The suggestion generation uses the beam search algorithm to generate a token sequence path to control the relevance and completeness of the generated suggestions.

[0051] In an embodiment, the data feature information can be input into a preset language model, and the preset language model can generate corresponding operation and maintenance suggestions according to the data feature information, and the diagnosis results and operation and maintenance suggestions can be fed back to the interactive device. In the actual operation process, the preset language model can use the beam search algorithm to generate a layer token sequence path according to the data feature information, first output the probability distribution of the first token, select the k tokens with the highest probability as the initial candidate set, and for each candidate sequence in the initial candidate set, generate all possible next tokens, calculate the probability of the new sequence (current sequence probability * conditional probability of the next token), and select the k sequences with the highest probability from all newly generated sequences. These k sequences become the new candidate set until a complete operation and maintenance suggestion is generated or the maximum length is reached. In one embodiment, the number of operation and maintenance suggestions can be multiple, which can be set in advance according to user needs.

[0052] An embodiment of the present invention realizes interaction with the user by obtaining the natural language input by the user based on the interactive device, determining the operation and maintenance entity and query intention of the natural language as sentence features; determining the original time series data in the low-voltage power distribution system according to the query features, determining the data feature information and data abnormality results corresponding to the original time series data, determining the diagnosis result corresponding to the natural language based on the data abnormality results and the query features, determining the operation and maintenance suggestions corresponding to the natural language according to the data feature information based on a preset language model, and feeding back the diagnosis result and operation and maintenance suggestions to the interactive device, so that the interactive device can visually display the diagnosis result and operation and maintenance suggestions, realize automated and intelligent operation and maintenance, reduce manual intervention, and improve operation and maintenance efficiency; at the same time, more intuitive and efficient human-computer interaction is realized through the interactive device, and the accuracy of operation and maintenance suggestions is improved.

[0053] In one embodiment, the training based on the preset language model includes:

[0054] Obtain preset general text data, and input the preset general text data as the first training set into the preset language model, so that the preset language model performs the first stage pre-training task according to the first training set data; wherein, the preset language model is a large language model based on the Transformer architecture, and the first stage pre-training task of the preset language model includes a masked language model task and a next sentence prediction task; obtain power system domain text data, and input the power system domain text data as the second training set into the preset language model, so that the preset language model performs the second stage pre-training task according to the second training set; wherein, the second stage pre-training task includes a masked language model task, a distribution system fault chain prediction task, and an equipment status reasoning task; extract the historical operation and maintenance records of the low-voltage distribution system, and input the historical operation and maintenance records as the third training set into the preset language model, and use a supervised fine-tuning strategy to adjust the hyperparameters of the preset language model until the accuracy of the output result of the preset language model is greater than the preset probability, thereby completing the training of the preset language model.

[0055] Preset general text data refers to a large-scale, pre-set set of general text data containing knowledge from multiple domains. For example, this data may include literature from related disciplines such as physics, electrical engineering, and materials science; technical manuals and specifications from equipment manufacturers; and national standards and industry specifications from the State Grid Corporation of China and various power technologies. Power system domain text data may include specialized literature, operation and maintenance manuals, and historical fault reports. Hyperparameters, such as the learning rate and number of iterations, require manual configuration prior to training the pre-set general text data. The pre-set language model is trained in three stages: the first stage uses general text data to perform masked language modeling (MLM) and next sentence prediction (NSP) tasks; the second stage uses power system domain text to perform distribution system fault chain prediction (FCP) and equipment state inference (ESI) tasks; and the third stage uses historical operation and maintenance records of low-voltage distribution systems for supervised fine-tuning to optimize output performance. This method uses progressive knowledge infusion, combining power system domain-specific tasks with low-voltage distribution system operation and maintenance data, to optimize the model's ability to understand and generate complex operation and maintenance scenarios.

[0056] The masked language model task randomly replaces some words in the input text with masking tokens (such as [MASK]), allowing the pre-set language model to predict the masked words based on contextual information. Through extensive training, the model learns semantic associations and grammatical rules between words, enhancing its understanding of text. The next sentence prediction task is a pre-training task that predicts whether two input sentences are consecutive or have a causal relationship, used to learn logical connections between sentences. Given two sentences, the pre-set language model determines whether the next sentence is the successor of the first. This helps the pre-set language model learn the coherence, logical relationships, and discourse structure between sentences, improving its understanding of the overall semantics of the text. The distribution system fault chain prediction task is a domain-specific task that predicts possible fault propagation paths based on initial fault conditions and generates a sequence of fault events. Based on power system domain text data such as distribution system operating data, equipment status information, and historical fault records, it can predict possible fault chains—a series of consecutive fault events and their sequence. This allows for early detection of potential fault risks, enabling preventive maintenance measures to reduce outage duration and the impact of faults on power supply reliability. Equipment State Inference (ESI): This domain-specific task infers the operating status of unmonitored devices based on partial monitoring data, using an attention mechanism to model inter-device relationships. This task infers the current operating status of a device (e.g., normal, abnormal, faulty, etc.) based on various monitoring data (e.g., temperature, vibration, current, voltage, etc.) and its operating history.

[0057] In one embodiment, pre-set general text data can be input as the first training set into a pre-set language model, causing the pre-set language model to perform masked language modeling and next sentence prediction tasks, enabling the pre-set language model to learn general language features and basic power system semantics. Specifically, the masked language model (MLM) task involves randomly masking 15% of the tokens in the input text, replacing 50% with domain-specific tokens, such as "transformer," "circuit breaker," and "voltage fluctuation," to enhance the model's semantic understanding of power-related vocabulary. Masked token prediction is based on a contextual bidirectional attention mechanism, using a cross-entropy loss function. The next sentence prediction (NSP) task predicts the next sentence and the causal relationship between sentences, such as the cause and effect of a fault. After a pre-set number of iterations, the first phase of training is completed, and the model parameters are initially equipped with general language understanding capabilities and basic power system semantics. In one embodiment, 50% of the training data samples are positive (real next sentences) and 50% are negative (randomly sampled sentences). A cosine annealing learning rate schedule is used, with a warm-up period of 10% of the total training steps. Exemplarily, the number of iterations is a preset value N1.

[0058] Power system domain text data is identified and used as a second training set to feed a pre-set language model. The pre-set language model then performs a second phase of pre-training based on this second training set, enhancing the model's understanding of power system domain-specific knowledge and focusing on its ability to reason about distribution system faults and device states. Specifically, the MLM task can increase the coverage rate to 20%, focusing on masking key device names, parameter values, and fault descriptions. By increasing the coverage rate and the proportion of domain-specific tokens, the model's ability to model specialized terminology is enhanced. The Fault Chain Prediction (FCP) task for distribution systems predicts possible fault propagation paths given initial fault conditions. A sequential path generation algorithm is used to generate fault chain sequences. For example, given an initial fault condition (e.g., "circuit breaker tripped"), a possible fault propagation path (e.g., "circuit breaker tripped → bus voltage abnormality → load power outage") is predicted. In one embodiment, a sequence path generation algorithm is employed, specifically: encoding the initial fault state as an input vector and using a Transformer decoder to generate a subsequent fault sequence; predicting the probability distribution of the next fault event at each step, selecting the k events with the highest probability (k = 5) as a candidate set, and retaining the sequence path with the highest cumulative probability based on a beam search algorithm until the sequence length reaches a preset value or the probability falls below a threshold. An equipment state inference (ESI) task is then performed, inferring the possible states of unmonitored equipment (e.g., "normal," "overheated," or "faulty") based on given partial monitoring data (e.g., voltage, current, and temperature). A state inference model based on an attention mechanism is employed, with a preset number of iterations, N2, to complete the second phase of training, resulting in model parameters that incorporate power system domain knowledge and fault inference capabilities. Specifically, the device state inference task performed using an attention-based state inference model includes: encoding monitoring data into feature vectors and inputting them into a multi-head self-attention layer; aggregating relevant device information through attention weights to predict the target device state; using a categorical cross-entropy loss function combined with contrastive learning loss to enhance state differentiation capabilities. Training parameters include a learning rate of 1e-5, a batch size of 64, iterations N2 of 500,000 steps, and a maximum norm of gradient clipping of 1.0. In one embodiment, the preset values ​​N1 and N2 can be user-defined values ​​based on training requirements, and the preset values ​​N1 and N2 can be the same or different.

[0059] Then, fault reports, expert diagnostic opinions, operation logs and other information are extracted as historical operation and maintenance records of the low-voltage distribution system, and the historical operation and maintenance records are input into the preset language model as the third training set. The hyperparameters of the preset language model are adjusted based on the output results of the preset language model and the loss function until the accuracy of the output results of the preset language model is greater than the preset probability. The training of the preset language model is completed, so that the preset language model is adapted to the high-precision language model of the low-voltage distribution system operation and maintenance scenario, which can generate accurate operation and maintenance suggestions and diagnostic results, thereby improving the accuracy and practicality of the operation and maintenance suggestions. In actual operation, a supervised fine-tuning strategy can be adopted. The fine-tuning includes three steps: the first step is to fine-tune only the output layer, with a learning rate of 1e-4 and 50 training rounds; the second step is to fine-tune the last three Transformer layers, with a learning rate of 5e-5 and 100 training rounds; the third step is to fine-tune all layers, with a learning rate of 1e-5 and 150 training rounds; a label-smoothed cross-entropy loss function with a smoothing factor of 0.1 is used, combined with contrastive learning loss for optimization, and the validation set evaluation indicator is the weighted F1 score. The early stopping strategy is triggered when the performance of the validation set does not improve after 10 consecutive rounds. Among them, contrastive learning loss is an auxiliary loss function that enhances the model's ability to distinguish similar operation and maintenance scenarios by maximizing the similarity of positive samples and minimizing the similarity of negative samples. In this embodiment, the preset language model is trained in three stages using three training sets, from general language knowledge to power system domain knowledge, and then to specific low-voltage distribution system operation and maintenance scenarios, gradually improving the model's professionalism and adaptability. Through a progressive knowledge injection method, the preset language model can better adapt to the complexity of the distribution system. At the same time, domain-specific task design, fault chain prediction and equipment status inference tasks are designed for distribution system operation and maintenance scenarios, which significantly improves the model's ability to understand fault propagation and equipment status; and, through the combination of step-by-step fine-tuning, label smoothing, comparative learning and early stopping mechanism, ensures that the model converges quickly under limited data while avoiding overfitting.

[0060] Example 2

[0061] Figure 2 This is a flow chart of an operation and maintenance method for a low-voltage power distribution system according to the second embodiment of the present invention. This embodiment is based on the above embodiment and is further optimized and expanded, and can be combined with various optional technical solutions in the above embodiment. Figure 2 As shown, the method includes:

[0062] S201. Receive natural language input by a user on an interactive device, and input the natural language into a preset intent classification model to determine the query intent corresponding to the natural language.

[0063] The preset intent classification model is a deep learning model based on a bidirectional long short-term memory network combined with an attention mechanism. The preset intent classification model can be used to identify and analyze query intent in natural language. In one embodiment, the algorithm of the preset intent classification model is y = softmax (W h + b) ， Where h is the hidden state of the Bi-LSTM, W is the weight matrix, b is the bias vector, and y is the probability distribution of intent. In an embodiment, natural language input from a user on an interactive device can be received, and the query intent corresponding to the natural language input can be determined according to a preset intent classification model. In one embodiment, query intents can include fault diagnosis, equipment status query, and predictive maintenance.

[0064] In one embodiment, after receiving the natural language input by the user on the interactive device, the method further includes:

[0065] When it is determined that any one of the operation and maintenance entity and the query intention does not exist, problem tracking information is generated based on the interactive device, so that the user can supplement the operation and maintenance entity and / or the query intention according to the problem tracking information.

[0066] Among them, issue tracking information can be understood as follow-up issues generated by interactive entities.

[0067] In an embodiment, when it is determined that no operation and maintenance entity or query intent has been obtained, issue tracking information can be generated based on the operation and maintenance entity or query intent, so that the user can supplement the operation and maintenance entity and / or query intent based on the issue tracking information, thereby achieving more intuitive and efficient human-computer interaction.

[0068] S202: Input the natural language into a preset sequence annotation model to determine the operation and maintenance entity corresponding to the natural language, and use the operation and maintenance entity and the query intent as sentence features.

[0069] The preset sequence labeling model is a deep learning model that combines conditional random fields with neural networks. In one embodiment, the preset sequence labeling model can use a sequence labeling model that combines conditional random fields (CRF) with neural networks to determine the operation and maintenance entity corresponding to the natural language. The potential function of the CRF layer is ψ(y i ,y i-1 ,x)=exp(W y ·f(y i ,y i-1 ,x)); where y i is the current label, y i-1 is the previous label, x is the input sequence, f is the feature function, W yis a weight vector. In an embodiment, natural language can be input into a preset sequence annotation model to identify the operation and maintenance entity corresponding to the natural language. In one embodiment, the operation and maintenance entity is an entity type of the power distribution system, which may include equipment name, time range, and fault type. After determining the operation and maintenance entity and query intent, the operation and maintenance entity and query intent are used as sentence features.

[0070] S203: extracting operation data of the operation and maintenance entities in the low-voltage power distribution system within a preset time period collected by preset sensors as original time series data according to the operation and maintenance entities and the query intent.

[0071] In an embodiment, the operation and maintenance entity operating data associated with the operation and maintenance entity and the query intent can be determined, and the operation and maintenance entity operating data within a preset time period can be extracted from the low-voltage power distribution system collected by preset sensors, and the operation and maintenance entity operating data within the preset time period can be used as the original time series data. In actual operation, the preset time period can range from 1 hour to 24 hours, which is not limited and can be selected according to user needs. The operation and maintenance entity operating data can include but is not limited to voltage, current, and temperature data.

[0072] S204 , preprocessing the original time series data to obtain target sequence data, and determining data anomaly results of the target sequence data according to a preset running data interval threshold.

[0073] Among them, the data anomaly results include data anomaly and data normal. The preset running data interval threshold refers to the preset interval information for detecting the target sequence data. There is a corresponding preset running data interval threshold for each type of preset running data. In an embodiment, the original time series data can be subjected to data preprocessing such as data source cleaning and standardization to obtain the target sequence data. It is then determined whether the target sequence data are all within the preset running data interval threshold. If so, the data anomaly result of the target sequence data is determined to be data normal; if not, the data anomaly result of the target sequence data is determined to be data anomaly. In one embodiment, the isolation forest algorithm can also be used to detect data anomaly results.

[0074] S205: Determine the data categories included in the operation and maintenance entity operation data corresponding to the same time in the target sequence data and the data corresponding to each data category, and discretize each data according to the preset interval of the data category to obtain data features.

[0075] Among them, data categories can include information such as measurement values, status, time, and events. In an embodiment, the operation and maintenance entity operation data in the target sequence data can be divided according to time. That is, the operation and maintenance entity operation data corresponding to the same time are grouped together, and the data categories contained in the operation and maintenance entity operation data in each group and the data corresponding to each data category are determined, and the data is discretized according to a preset interval. For example, for measurement values: the continuous value is specifically mapped to a predefined interval, and the corresponding Chinese vocabulary is selected as the data feature according to the interval, such as the voltage value 0.95 is mapped to "normal"; for status and events: predefined Chinese vocabulary can be used as data features, such as: the operation status "1" is mapped to "running"; for time: the number can be retained, but the Chinese unit is added at the end, for example: 13:45:30 is mapped to "13 hours 45 minutes 30 seconds"; for abnormal values ​​or missing values, special tokens such as "unknown" or "abnormal" can be used.

[0076] S206: Map the data features into a vector space of fixed dimension to obtain a token sequence of the operation data of the operation and maintenance entity corresponding to each time, and splice the token sequence in a preset time sequence as data feature information.

[0077] In an embodiment, the data features can be mapped into a vector space of fixed dimension, the token sequence of the operation data of the operation and maintenance entity corresponding to each time is determined, and the token sequence is arranged and spliced ​​in chronological order to obtain a new token sequence as data feature information.

[0078] S207: When it is determined that the data abnormality result is normal, the diagnosis result may be determined that the operation and maintenance entity in the query feature is normal.

[0079] In an embodiment, when the data abnormality result is determined to be normal data, it can be considered that there is no abnormality. At this time, the diagnosis result can be determined to be that the operation and maintenance entity in the query feature is normal.

[0080] S208. When the data anomaly result is determined to be a data anomaly, the abnormal data in the target sequence data is extracted, and the diagnosis result corresponding to the natural language is determined by matching the abnormal data and the operation and maintenance entity in the preset knowledge graph.

[0081] Among them, the preset knowledge graph can be a knowledge graph constructed in advance according to the knowledge base information. In the construction of the preset knowledge graph, you can first build entity types, such as operation and maintenance entities, fault conditions and fault causes, define the relationship between entities, such as belongs to, causes, etc., and extract entities and relationships from the text data in the knowledge base and convert them into the preset knowledge graph.

[0082] In an embodiment, when a data anomaly result is determined to be a data anomaly, the abnormal data in the target sequence data can be extracted, and nodes associated with the abnormal data and the operation and maintenance entity can be matched in a preset knowledge graph, and the diagnosis result can be inferred based on the nodes. In actual operation, methods such as Bayesian network probabilistic reasoning can be used to determine the probability of each fault, and the fault with the highest probability is used as the diagnosis result.

[0083] S209: Input the data feature information into a preset language model, determine the probability distribution of the first prediction sequence corresponding to the data feature information based on the preset language model, and extract a preset number of first prediction sequences in the probability distribution as candidate sets in descending order of prediction probability.

[0084] In an embodiment, data feature information can be input into a preset language model. A beam search algorithm is then used within the preset language model to predict the probability distribution of a first prediction sequence based on the data feature information. The first of a preset number of prediction sequences with the highest probability is selected as the candidate set. Generally, the number of predictions can be set based on user needs, and the preset number is the same as the number recommended by the operator. In one embodiment, the first prediction sequence can be a token sequence.

[0085] S210 , generating a next prediction sequence according to the candidate set based on a preset language model, and combining the next prediction sequence and each first prediction sequence in the candidate set into a sequence path.

[0086] In an embodiment, the beam search algorithm may be used to predict the next prediction sequence according to the candidate set using the preset language model, and determine that the first prediction sequence in each candidate set and each next prediction sequence form a sequence path.

[0087] S211. Determine the probability product of the predicted probability of the next predicted sequence and the predicted probability of the first predicted sequence in each sequence path based on the preset language model, extract a preset number of sequence paths as candidate sets in descending order of the probability products, and use the sequence paths in the candidate set as operation and maintenance recommendations until the preset conditions are met.

[0088] The preset conditions include that the probability product is less than a preset probability product or the sequence path reaches a preset length.

[0089] In this embodiment, a preset language model can be used to continuously determine the probability product of the next predicted sequence in each sequence path and the predicted probability of the first predicted sequence. A preset number of sequence paths with the highest probabilities are selected as a candidate set until the probability product is less than the preset probability product or the sequence path reaches a preset length. The sequence paths in the candidate set are then used as operation and maintenance recommendations. For example, for generating operation and maintenance recommendations for a substation, the preset number k is set to 2. The initial candidate set is determined as follows: "Inspect" (probability: 0.6) and "Replace" (probability: 0.4). After the first expansion, the following paths are selected: "Check the transformer" (probability: 0.6*0.5 = 0.30); "Check the wiring" (probability: 0.6*0.3 = 0.18); "Replace the fuse" (probability: 0.4*0.4 = 0.16); and "Replace the switch" (probability: 0.4*0.3 = 0.12). The two sequential paths, "Check the transformer" (probability: 0.30) and "Check the wiring" (probability: 0.18), are selected as candidate sets. This process continues until the pre-defined conditions are met. For example, the O&M recommendations might be: "Check the transformer oil level and add insulating oil" (the highest probability path) and "Check the wiring insulation condition and perform maintenance" (the second highest probability path).

[0090] S212: Feedback the diagnosis results and operation and maintenance suggestions to the interactive device, so that the interactive device can visually display the diagnosis results and operation and maintenance suggestions.

[0091] In one embodiment, after the diagnosis results and operation and maintenance suggestions are fed back to the interactive device, the data processing strategy may also be fed back to the preset language model to optimize the preset language model.

[0092] In an embodiment of the present invention, by receiving natural language input by a user on an interactive device, the natural language is input into a preset intent classification model to determine the query intent corresponding to the natural language, the natural language is input into a preset sequence annotation model to determine the operation and maintenance entity corresponding to the natural language, and the operation and maintenance entity and the query intent are used as sentence features to achieve accurate acquisition of the query intent and the operation and maintenance entity; by extracting the operation and maintenance entity operation data within a preset time in a low-voltage power distribution system collected by a preset sensor as original time series data according to the operation and maintenance entity and the query intent, the original time series data is subjected to data preprocessing to obtain target sequence data, and the data anomaly results of the target sequence data are determined according to a preset operation data interval threshold, thereby achieving automatic determination of the data anomaly results; by determining the data categories contained in the operation and maintenance entity operation data corresponding to the same time in the target sequence data and the data corresponding to each data category, each data is discretized according to the preset interval of the data category to obtain data features, and the data features are mapped to a fixed-dimensional vector space to obtain the data features corresponding to each time The token sequence of the operation and maintenance entity operation data is spliced ​​in a preset time sequence as data feature information, the diagnosis result corresponding to the data feature information is determined, and the accuracy and reliability of the diagnosis result are improved through the preset knowledge graph; by inputting the data feature information into a preset language model, the probability distribution of the first prediction sequence corresponding to the data feature information is determined based on the preset language model, and a preset number of first prediction sequences in the probability distribution are extracted as candidate sets in descending order of prediction probability, and the next prediction sequence is generated according to the candidate set based on the preset language model, and the next prediction sequence and each first prediction sequence in the candidate set are combined into a sequence path, and the probability product of the prediction probability of the next prediction sequence and the prediction probability of the first prediction sequence in each sequence path is determined based on the preset language model, and a preset number of sequence paths are extracted as candidate sets in descending order of probability products until the preset conditions are met, and the sequence paths in the candidate set are used as operation and maintenance suggestions to improve operation and maintenance efficiency and accuracy.

[0093] Example 3

[0094] Figure 3 This is a flowchart of an operation and maintenance method for a low-voltage power distribution system provided according to the third embodiment of the present invention. This embodiment is based on the above embodiment, and the operation and maintenance system of the low-voltage power distribution system performs the operation and maintenance method of the low-voltage power distribution system. The distribution system data is used as the original data, the original data is used as the original time series data, and the pre-processed data is used as the target sequence data. For example, the operation and maintenance method of the low-voltage power distribution system is further explained. Figure 3 As shown in FIG, the operation and maintenance system of the low-voltage distribution system includes a data collection module, a data processing module, a large language model module, an active interaction module and an intelligent decision-making module.

[0095] In one embodiment, the data collection module collects real-time operation data from various nodes of the low-voltage power distribution system. Its input data: raw data (i.e., real-time operation data) from various sensors (such as voltage sensors, current sensors, temperature sensors, etc.). Generated data: sensor reading data marked with timestamps. Output to: data processing module. In one embodiment, Figure 4 This is a data collection module network topology diagram provided by the third embodiment of the present invention. Figure 4 As shown, the distribution transformer area includes distribution transformer 1, distribution transformer 2, distribution transformer 3, edge computing node 1 and central server, wherein the sensor data of distribution transformer 1, distribution transformer 2 and distribution transformer 3 are transmitted to edge computing node 1, and the data is transmitted to the central server via industrial Ethernet for transmission to the database. The distribution cabinet area 1 includes distribution cabinet 1, distribution cabinet 2, distribution cabinet 3 and NB-IoT gateway 1. Distribution cabinet 1, distribution cabinet 2 and distribution cabinet 3 send NB-IoT to NB-IoT gateway 1. NB-IoT gateway 1 transmits information to the central server via industrial Ethernet for transmission to the database. The distribution cabinet area 2 includes distribution cabinet 4, distribution cabinet 5, distribution cabinet 6 and NB-IoT gateway 2. Distribution cabinet 4, distribution cabinet 5 and distribution cabinet 6 send NB-IoT to NB-IoT gateway 2. NB-IoT gateway 2 transmits information to the central server via industrial Ethernet for transmission to the database. In one embodiment, Figure 5 FIG. 1 is a flow chart of a data transmission process according to the third embodiment of the present invention. Figure 5 As shown, raw data can be collected through the sensor network, converted into digital data by the data collection unit in the data collection module, cached by the edge data cache, transmitted through the data transmission network, processed by the data cleaning unit in the data collection module to obtain cleaned data, processed by the standardization unit to obtain standardized data, extracted features of the standardized data by the feature extraction unit, compressed by the data compression unit for data storage, and stored in the large language model module and the intelligent decision module. Specifically, the data transmission module can deploy various sensors at key nodes of the low-voltage power distribution system through the sensor network unit. These sensors include voltage sensors, current sensors, temperature sensors, humidity sensors, etc. Using intelligent sensor technology, the accuracy and reliability of data collection are improved.

[0096] The sensor data is read through the data acquisition unit and preliminarily digitized. This unit designs a multi-channel high-speed data acquisition circuit that supports synchronous sampling. Automatic adjustment of the programmable gain amplifier (PGA) and analog-to-digital converter (ADC) is achieved. A field-programmable gate array (FPGA) is used to achieve high-speed data processing and local storage. The data transmission unit can use a hybrid wired and wireless network to transmit data to the data processing module. Among them, wired transmission can use industrial Ethernet and power line carrier communication (PLC). Wireless transmission can use a multi-protocol compatible wireless communication module that supports WiFi, ZigBee, NB-IoT, etc.

[0097] The data processing module can pre-process the collected raw data in preparation for subsequent analysis. Its input data is the raw sensor data from the data collection module and the data processing strategy updates from the intelligent decision-making module. It generates data: cleaned, standardized, and feature-extracted data. The pre-processed data is output to the large language model module, and the anomaly detection results are sent to the intelligent decision-making module. The collected raw data is pre-processed, including data cleaning, standardization, feature extraction, and data compression steps. This module is responsible for pre-processing the collected raw data in preparation for subsequent analysis. Its main functions include data cleaning, data standardization, feature extraction, and data compression. In one embodiment, the large language model module can process and analyze distribution system data to generate operation and maintenance recommendations and decision-making recommendations. Its input data comes from the pre-processed distribution system data from the data processing module. Output: Operation and maintenance recommendations, including operation and maintenance recommendations, fault diagnosis results, and predictive maintenance recommendations. The results are output to: the active interaction module and the intelligent decision-making module. This module uses a multi-layer Transformer encoder and decoder. Each layer contains a multi-head self-attention mechanism and a feedforward neural network. Through pre-training and fine-tuning, the model is adapted to the specific application scenarios of low-voltage distribution system operation and maintenance.

[0098] The multi-layer Transformer encoder and decoder units utilize a large language model (LLM) based on the Transformer architecture as the core technology for processing and analyzing distribution system data. The model architecture comprises multiple layers of Transformer encoders and decoders, each composed of the following key components: a multi-head self-attention mechanism, a feedforward neural network unit, a pre-training unit, a fine-tuning mechanism, an input processing unit, and an output parsing unit.

[0099] Among them, the multi-head self-attention mechanism: uses the scaled dot product attention algorithm to calculate the attention weight. Formula: Where Q, K, and V represent query, key, and value matrices, respectively. k The dimension of the key. A feedforward neural network unit consists of two linear transformations and a ReLU activation function. Formula: FFN(x) = max(0,xW1+b1)W2+b2; where W1 and W2 are weight matrices, and b1 and b2 are bias vectors.

[0100] Pre-training unit, the pre-training of this module is divided into two stages: the first stage: using large-scale general text data for pre-training to learn the basic characteristics of the language. The masked language model (MLM) and next sentence prediction (NSP) tasks are used. The second stage: using professional literature and operation and maintenance manuals in the power system field for domain adaptation training. The same MLM and NSP tasks are used, but domain-specific data are used. This unit is characterized by including the following steps: In one embodiment, Figure 6 This is a flowchart of a large language model training method according to the third embodiment of the present invention. Figure 6 As shown: To achieve high accuracy and domain adaptability for low-voltage distribution system operation and maintenance recommendations, this paper proposes a three-stage training method for a preset language model based on the Transformer architecture. This method differs from the training process of general language models such as Bidirectional Encoder Representations from Transformers (BERT) and Generative Pretrained Transformer (GPT). Through progressive knowledge injection, this method combines specific tasks in the power system domain with low-voltage distribution system operation and maintenance data to optimize the model's understanding and generation capabilities for complex operation and maintenance scenarios.

[0101] Technical details of the three-stage training: (1) First-stage pre-training: The goal is to enable the preset language model to learn general language features and basic power system semantic associations. The input data uses large-scale general text data containing multi-domain knowledge for preliminary pre-training. The general text data includes: literature in related disciplines such as physics, electrical engineering, and materials science; technical manuals and specifications of equipment manufacturers; national standards and industry specifications of the State Grid and various power technologies, with a total data volume of about 1 billion words. The training task can adopt an improved masked language model (MLM) and next sentence prediction (NSP) task, which includes: The improved MLM task includes randomly masking 15% of the words in the input text, of which 50% are replaced with words in specific fields, such as "transformer", "circuit breaker", "voltage fluctuation", etc., to enhance the model's semantic understanding of the vocabulary in the power field. The prediction of masked words is based on the contextual bidirectional attention mechanism, and the loss function is the cross-entropy loss. The improved NSP task not only predicts the next sentence, but also predicts the causal relationship between sentences, such as the cause and result of the fault. Specifically: Given two sentences, predict whether the second sentence is the next sentence of the first, and further predict the causal relationship between the sentences (e.g., fault cause and consequence). The training data consists of 50% positive samples (the actual next sentence) and 50% negative samples (randomly selected sentences). A cosine annealing learning rate schedule is used, with a warm-up period of 10% of the total training steps. Output: Model parameters demonstrating preliminary general language understanding capabilities and basic semantic associations for power systems.

[0102] (2) Second stage pre-training: Objective: To enhance the model's understanding of specific knowledge in the power system domain, focusing on the ability to reason about distribution system faults and equipment status. Input data: Professional literature, operation and maintenance manuals, and historical fault reports in the power system field, including: professional textbooks and papers on distribution system design, operation, and maintenance; historical distribution system fault case analysis reports; fault diagnosis manuals and maintenance guides for various distribution equipment; and safety regulations and emergency plans for the power industry, with a total data volume of approximately 100 million tokens. Training tasks: Distribution system-specific MLM tasks, with a coverage rate increased to 20%, focusing on masking parameters such as equipment names (such as "distribution cabinet" and "busbar"), parameter values ​​(such as "220V" and "50Hz"), and fault descriptions (such as "short circuit" and "overload"). By improving the coverage rate and the proportion of domain-specific tokens, the model's ability to model professional terminology is enhanced. ; Distribution system fault chain prediction (FCP) task, predicting the possible fault propagation paths under given initial fault conditions. Specifically, based on the initial fault conditions (such as "circuit breaker tripping"), possible fault propagation paths (such as "circuit breaker tripping → bus voltage abnormality → load power off") can be predicted. A sequence path generation algorithm is used, and the specific steps are as follows: 1. Encode the initial fault state as an input vector, and use the Transformer decoder to generate the subsequent fault sequence; 2. Predict the probability distribution of the next fault event at each step, and select the k events with the highest probability (k = 5) as the candidate set; 3. Based on the beam search algorithm, retain the sequence path with the highest cumulative probability until the sequence length reaches the preset value or the probability is lower than the threshold. Equipment State Inference (ESI) task: Based on partial monitoring data (such as voltage, current, temperature), infer the operating status of unmonitored equipment (such as "normal", "overheated", "faulty"). A state inference model based on an attention mechanism is employed. Specifically, the following steps are taken: 1. Monitoring data is encoded as feature vectors and fed into a multi-head self-attention layer. 2. Attention weights are used to aggregate relevant device information and predict the target device state. 3. The loss function is categorical cross entropy combined with contrastive learning loss to enhance state differentiation. Training parameters include a learning rate of 1e-5, a batch size of 64, iterations N2 = 500,000 steps, and a maximum norm of gradient clipping of 1.0. Output: Model parameters that incorporate power system domain knowledge and fault inference capabilities.

[0103] (3) The third stage of pre-training: Goal: to adapt the model to the operation and maintenance scenarios of specific low-voltage distribution systems and improve the accuracy and practicality of operation and maintenance recommendations. The third training set is the historical operation and maintenance records of the low-voltage distribution system, including fault reports, expert diagnostic opinions, operation logs, etc., with a total data volume of approximately 10 million words. Data preprocessing includes text segmentation, stop word removal, professional vocabulary annotation, and numerical data normalization. Model fine-tuning uses specific low-voltage distribution system operation and maintenance data to perform supervised fine-tuning (Supervised Fine-Tuning, SFT) on the model to adapt to specific application scenarios. Specifically, the first step is to fine-tune only the output layer, fix the Transformer encoder and decoder parameters, the learning rate is 1e-4, and the number of training rounds is 50; the second step is to fine-tune the last three Transformer layers, the learning rate is 5e-5, and the number of training rounds is 100; the third step is to fine-tune all layers, the learning rate is 1e-5, and the number of training rounds is 150. Loss Function: The primary loss function is a label-smoothed cross-entropy loss (with a smoothing factor of 0.1). Auxiliary loss functions include contrastive learning loss, which is used to enhance the model's ability to distinguish between similar O&M scenarios. Contrastive learning loss is an auxiliary loss function that maximizes the similarity of positive samples and minimizes the similarity of negative samples, thereby enhancing the model's ability to distinguish between similar O&M scenarios. Optimization Parameters: Batch size is 32, gradient accumulation updates parameters every 4 steps (equivalent to a batch size of 128), the optimizer is AdamW, the learning rate schedule is cosine annealing with a warm-up period (the warm-up period is 10% of the total number of steps), and the L2 regularization weight decay coefficient is 0.01. Early stopping strategy: The weighted F1 score on the validation set is used as the evaluation metric. Early stopping is triggered if there is no performance improvement on the validation set for 10 consecutive rounds. Data augmentation: Synonym replacement technology and distribution system simulation software are used to generate virtual O&M data to increase the diversity of the training set. Output: A high-precision language model adapted to low-voltage distribution system O&M scenarios, capable of generating accurate O&M recommendations and diagnostic results.

[0104] The significant differences from traditional model training methods are: 1. Progressive knowledge injection: Through three stages of training, the model's professionalism and adaptability are gradually improved, from general language knowledge to power system domain knowledge, and then to specific low-voltage distribution system operation and maintenance scenarios. 2. Domain-specific task design: The fault chain prediction (FCP) and equipment state inference (ESI) tasks are designed for distribution system operation and maintenance scenarios, significantly improving the model's understanding of fault propagation and equipment status. 3. Efficient optimization strategy: A combination of step-by-step fine-tuning, label smoothing, comparative learning, and early stopping mechanisms ensures rapid model convergence with limited data while avoiding overfitting.

[0105] In actual low-voltage distribution system operation and maintenance scenarios, the preset language model trained using this three-stage training method improved the accuracy of operation and maintenance recommendations by approximately 15% compared to the general BERT model and the untuned power domain model. The weighted F1 score of fault diagnosis increased by approximately 12%, and the average response time for generating recommendations was controlled within 1 second, meeting real-time operation and maintenance requirements. In one embodiment, the parameters shown in Table 1 can be used for fine-tuning:

[0106] Table 1 Fine-tuning parameters

[0107]

[0108]

[0109] The input processing unit converts various time series data of the power distribution system (such as voltage, current, temperature, etc.) into a token sequence that the model can understand. The main steps include: a) Data normalization: standardizing different types of data to the same scale. b) Time encoding: using sine and cosine functions to encode time information. c) Feature concatenation: concatenating different types of features in a predefined order. d) Word segmentation: using a specific word segmenter to convert the concatenated feature sequence into a token sequence (i.e., data feature information). The main data features of this solution include:

[0110] The data features are divided into several main categories: measurement value, status, event, and time, and Chinese vocabulary mapping rules are designed for each category: measurement value: use words such as "very low", "low", "normal", "high", and "very high"; status: such as "running", "stop", "standby", and "fault"; event: such as "start", "shutdown", "overload", and "short circuit"; time: retain the digital form, but add Chinese units such as "hour", "minute", and "second". a. For measurement values: Continuous values ​​are mapped to predefined intervals, and corresponding Chinese terms are selected based on the intervals. For example, a voltage value of 0.95 is mapped to "normal"; b. For states and events: Predefined Chinese terms are used directly, for example, operating state "1" is mapped to "running"; c. For time: Numeric values ​​are retained but Chinese units are appended. For example, 13:45:30 is mapped to "13:45:30"; d. Special value handling: For outliers or missing values, special tokens such as "unknown" or "abnormal" are used. The output parsing unit primarily parses the token sequence output by the model into specific operation and maintenance recommendations and decisions. Its main technical feature is the use of a beam search heuristic algorithm to decode the most likely token sequence from the model output.

[0111] The logic is as follows: Step 1: Initialization: The model outputs the probability distribution of the first token; the k tokens with the highest probabilities are selected as the initial candidate set. Step 2: Expansion: For each candidate sequence, all possible next tokens are generated and the probability of the new sequence is calculated (current sequence probability * next token conditional probability). Step 3: Selection: From all newly generated sequences, the k sequences with the highest probabilities are selected; these k sequences become the new candidate set. Step 4: Repeat steps 2 and 3 until a complete operation and maintenance recommendation is generated or the maximum length is reached.

[0112] In one embodiment, Figure 7 This is a workflow diagram of an active interaction module provided according to the third embodiment of the present invention. Figure 7 As shown in the figure, the active interaction module can realize intelligent dialogue with operation and maintenance personnel based on natural language processing technology. The input data is the natural language query of the operation and maintenance personnel and the operation and maintenance suggestions and diagnosis results from the large language model module. The generated data is the parsed query intent and operation and maintenance entities, and the generated dialogue response. The output is to: Intelligent decision-making module (query intent and operation and maintenance entities), direct interaction with users (dialogue response). This module is based on natural language processing technology to realize intelligent dialogue with operation and maintenance personnel, including modules such as intent recognition, entity extraction, dialogue management, and response generation. The active interaction module mainly includes: Intent interaction recognition unit: The function is to identify the query intent of the operation and maintenance personnel, such as distribution fault diagnosis, distribution equipment status query, etc. Implementation method: a) Adopting a deep learning-based intent classification model, using a bidirectional long short-term memory network (Bi-LSTM) combined with an attention mechanism. b) Using pre-trained word embedding. This solution uses Word2Vec as input features. c) Introducing a set of intent labels specific to the power industry, such as "fault diagnosis", "equipment status query", "predictive maintenance", etc. The intent classification model algorithm of this module is: y = softmax(W·h+b); where h is the hidden state of Bi-LSTM, W is the weight matrix, b is the bias vector, and y is the intent probability distribution.

[0113] Distribution system entity extraction unit: Extract key information of the distribution system from user interaction input, such as equipment name, time range, etc. Its implementation method: Use the sequence labeling model of conditional random field (CRF) combined with neural network. Design entity types specific to the distribution system, such as "equipment name", "fault type", "time range", etc. Use a combination of character-level and word-level features to improve the accuracy of the device in identifying distribution system entities. Its key algorithm: The potential function of the CRF layer is ψ(y i ,y i-1 ,x)=exp(W y ·f(y i ,y i-1 ,x)); where y iis the current label, y i-1 is the previous label, x is the input sequence, f is the feature function, W y is the weight vector.

[0114] The function of the dialogue management unit is to combine the user dialogue context, maintain the device dialogue state, and handle multiple rounds of dialogue. Its implementation method is as follows: a) Adopt a dialogue strategy model based on deep reinforcement supervised learning. b) Design a dialogue state representation specific to distribution system operation and maintenance, including user intention, extracted entities, dialogue history, etc. c) Define an action space for operation and maintenance scenarios, such as "request more information", "perform diagnosis", "provide suggestions", etc. Among them, the Q-value update strategy of Q-learning (DQN): Q(s t ,a t )←Q(s t ,a t )+α[r t +γmax a Q(s t+1 ,a)-Q(s t ,a t )]; where s t is the current state, a t is the selected action, r t is the reward, γ is the discount factor, and α is the learning rate.

[0115] The response generation unit generates a corresponding reply based on the intent identified by this module and the output of the entity unit extracted. This is achieved by: a) employing a Transformer-based sequence-to-sequence model; b) introducing a mechanism to inject distribution system expertise, such as using a knowledge graph to enhance the professionalism of responses; and c) designing response templates specific to operation and maintenance scenarios to improve the controllability of generated content.

[0116] In one embodiment, Figure 8 This is a workflow diagram of an intelligent decision-making module according to the third embodiment of the present invention. Figure 8As shown, the intelligent decision-making module can automatically generate diagnostic results based on the collected data and analysis results. Its input data: analysis results from the large language model module, user intentions and entities from the active interaction module, and anomaly detection results from the data processing module. Generated data: operation and maintenance decisions and strategies, including maintenance plans, resource allocation recommendations, etc. Output to: active interaction module (used to communicate decisions with users), data processing module (used to update data processing strategies). Based on the collected data and analysis results, the automatic generation of operation and maintenance decisions and strategies includes steps such as data fusion, anomaly detection, fault diagnosis and decision generation. In one embodiment, the intelligent decision-making module can directly call the large language model module to determine the diagnostic results. The intelligent decision-making module includes a data fusion unit, anomaly detection unit, fault diagnosis unit and decision generation unit. Among them, the function of the data fusion unit is to integrate data from different sources to form a comprehensive decision-making basis. Implementation method: a) Use multi-source heterogeneous data fusion technology to process data from large language models, active interaction and data processing modules. b) Use graph neural network (GNN) to establish the association relationship between data. c) Realize the collaborative processing of time series data and static data. Key algorithm:

[0117] Graph Convolutional Network (GCN) update rules: Among them, A is the adjacency matrix, D is the degree matrix, and H (l) is the node feature of the lth layer, W (l) is the weight matrix and σ is the activation function.

[0118] The function of the anomaly detection unit is to identify abnormal conditions in the distribution system based on the integrated distribution system data. Implementation method: a) Combining statistical methods and deep learning models, such as autoencoders and long short-term memory networks (LSTMs). b) Introducing domain knowledge to set anomaly thresholds and rules specific to the distribution system. c) Implementing multi-dimensional anomaly detection, including device-level, system-level, and regional-level anomalies. Key algorithm: Autoencoder reconstruction error calculation: E = ‖XX′‖ 2 ; Where X is the input data, X' is the reconstructed data, and E is the reconstruction error.

[0119] The fault diagnosis unit analyzes the causes and locates faults after detecting anomalies. This is achieved by: a) employing a knowledge graph-based reasoning mechanism, combined with expert rules and machine learning models; b) constructing a fault tree model to implement multi-level fault reasoning; and c) introducing causal reasoning techniques to analyze the root cause and propagation path of the fault. The key algorithm is Bayesian network probabilistic reasoning: Where Y is the cause of the failure and X is the observed symptom. The decision generation unit generates operations and maintenance decisions and strategies based on diagnostic results. Implementation methods: a) Utilize a multi-objective optimization algorithm to balance factors such as reliability, economic efficiency, and environmental impact. b) Introduce reinforcement learning technology to achieve long-term decision optimization. c) Integrate with expert systems to integrate domain knowledge and data-driven decision making.

[0120] In one embodiment, Figure 9 is a timing diagram of an operation and maintenance method of a low-voltage power distribution system provided according to the third embodiment of the present invention, such as Figure 9 As shown:

[0121] Step 1: The sensor sends raw data to the data collection module; Step 2: The data collection module transmits the raw data to the data processing module; Step 3: The data processing module cleans and standardizes the data; Step 4: The data processing module extracts features from the pre-processed data; Step 5: The data processing module sends the pre-processed data to the large language model module; Step 6: The data processing module sends the anomaly detection results to the intelligent decision-making module; Step 7: The large language model module deeply analyzes the data; Step 8: The large language model module sends the deep analysis results to the intelligent decision-making module; Step 9: The large language model module The block sends the operation and maintenance suggestions and diagnosis results to the active interaction module; step 10, the intelligent decision-making module generates operation and maintenance strategies and policies according to the anomaly detection results and analysis results; step 11, the intelligent decision-making module sends the decision results to the active interaction module; step 12, the operation and maintenance personnel send queries or instructions to the active interaction module; step 13, the active interaction module parses the intent and entity; step 14, the active interaction module sends the parsed user intent and entity to the intelligent decision-making module; step 15, the active interaction module returns a response or execution result to the operation and maintenance personnel; step 16, the intelligent decision-making module feeds back the updated data strategy to the data processing module.

[0122] This embodiment adopts a three-stage pre-training strategy, with each stage specifically designed for specific tasks. The FCP and ESI tasks are specifically designed for power distribution systems. This progressive knowledge injection approach better adapts to the complexity of distribution systems. Automated operations and maintenance reduce manual intervention and improve efficiency. Big data analysis and machine learning are used to predict potential faults, enabling proactive maintenance and providing data-driven decision-making advice to operators, improving decision accuracy. Natural language processing technology enables more intuitive and efficient human-computer interaction, reducing power outages caused by faults and lowering overall operation and maintenance costs.

[0123] Example 4

[0124] Figure 10 Schematic diagram of the structure of an operation and maintenance device for a low-voltage power distribution system according to the fourth embodiment of the present invention. Figure 10 As shown, the device includes: a data acquisition module 101, a result determination module 102, a diagnosis result determination module 103 and an operation and maintenance suggestion determination module 104.

[0125] The data acquisition module 101 is used to acquire the natural language input by the user based on the interactive device, determine the operation and maintenance entity and query intent of the natural language as sentence features, and determine the original time series data in the low-voltage power distribution system based on the query features;

[0126] Result determination module 102, used to determine data feature information and data anomaly results corresponding to the original time series data;

[0127] A diagnosis result determination module 103 is used to determine a diagnosis result corresponding to the natural language based on the data abnormality result and the query feature;

[0128] The operation and maintenance suggestion determination module 104 is used to determine the operation and maintenance suggestions corresponding to the natural language according to the data feature information based on the preset language model, and feed back the diagnosis results and operation and maintenance suggestions to the interactive device so that the interactive device can visually display the diagnosis results and operation and maintenance suggestions.

[0129] The technical solution of the embodiment of the present invention obtains the natural language input by the user based on the interactive device through the data acquisition module, determines the operation and maintenance entity and query intention of the natural language as sentence features, and realizes interaction with the user; determines the original time series data in the low-voltage power distribution system according to the query features, the result determination module determines the data feature information and data abnormality results corresponding to the original time series data, the diagnosis result determination module determines the diagnosis result corresponding to the natural language based on the data abnormality results and the query features, and the operation and maintenance suggestion determination module determines the operation and maintenance suggestions corresponding to the natural language according to the data feature information based on the preset language model, and feeds back the diagnosis results and operation and maintenance suggestions to the interactive device, so that the interactive device can visually display the diagnosis results and operation and maintenance suggestions, realize automated and intelligent operation and maintenance, reduce manual intervention, and improve operation and maintenance efficiency; at the same time, more intuitive and efficient human-computer interaction is realized through the interactive device, and the accuracy of operation and maintenance suggestions is improved.

[0130] In one embodiment, the data acquisition module 101 includes:

[0131] An intent determination unit is configured to receive natural language input from a user on an interactive device and input the natural language into a preset intent classification model to determine the query intent corresponding to the natural language; wherein the preset intent classification model is a deep learning model based on a bidirectional long short-term memory network combined with an attention mechanism;

[0132] An entity determination unit is configured to input natural language into a preset sequence annotation model to determine the operation and maintenance entity corresponding to the natural language, using the operation and maintenance entity and query intent as sentence features; wherein the preset sequence annotation model is a deep learning model that combines conditional random fields with neural networks;

[0133] The data acquisition unit is used to extract the operation data of the operation and maintenance entities in the low-voltage power distribution system within a preset time period collected by preset sensors as original time series data according to the operation and maintenance entities and query intentions.

[0134] In one embodiment, the operation and maintenance device of the low-voltage power distribution system further includes:

[0135] The question generation module is used to generate question tracking information based on the interactive device when it is determined that any one of the operation and maintenance entity and the query intention does not exist, so that the user can supplement the operation and maintenance entity and / or query intention according to the question tracking information.

[0136] In one embodiment, the result determination module 102 includes:

[0137] An abnormal result determination unit is used to perform data preprocessing on the original time series data to obtain target sequence data, and determine the data abnormal result of the target sequence data according to a preset running data interval threshold; wherein the data abnormal result includes data abnormality and data normality;

[0138] A data feature determination unit is used to determine the data categories and data corresponding to each data category contained in the operation and maintenance entity operation data corresponding to the same time in the target sequence data, and discretize each data according to the preset interval of the data category to obtain data features;

[0139] The feature information determination unit is used to map the data features into a vector space of fixed dimension to obtain a token sequence of the operation data of the operation and maintenance entity corresponding to each time, and to splice the token sequence in a preset time order as the data feature information.

[0140] In one embodiment, the diagnosis result determination module 103 includes:

[0141] A first result determination unit is configured to, when the data abnormality result is determined to be normal, determine that the diagnosis result is that the operation and maintenance entity in the query feature is normal;

[0142] The second result determination unit is used to extract the abnormal data in the target sequence data when the data anomaly result is determined to be a data anomaly, and determine the diagnostic result corresponding to the natural language according to the matching of the abnormal data and the operation and maintenance entity in the preset knowledge graph.

[0143] In one embodiment, the operation and maintenance device of the low-voltage power distribution system further includes:

[0144] A first training module is configured to obtain preset general text data and input the preset general text data as a first training set into a preset language model, so that the preset language model performs a first-stage pre-training task according to the first training set data; wherein the preset language model is a large language model based on the Transformer architecture, and the first-stage pre-training task of the preset language model includes a masked language model task and a next sentence prediction task;

[0145] A second training module is configured to obtain text data in the power system domain and input the text data in the power system domain into a preset language model as a second training set, so that the preset language model performs a second-stage pre-training task according to the second training set; wherein the second-stage pre-training task includes a masked language model task, a distribution system fault chain prediction task, and an equipment state reasoning task;

[0146] The third training module is used to extract the historical operation and maintenance records of the low-voltage distribution system, input the historical operation and maintenance records into the preset language model as the third training set, and adjust the hyperparameters of the preset language model until the accuracy of the output result of the preset language model is greater than the preset probability, thereby completing the training of the preset language model.

[0147] In one embodiment, the operation and maintenance suggestion determination module 104 includes:

[0148] a candidate set determination unit, configured to input the data feature information into a preset language model, determine the probability distribution of the first prediction sequence corresponding to the data feature information based on the preset language model, and extract a preset number of first prediction sequences in the probability distribution in descending order of prediction probability as candidate sets;

[0149] a sequence path set determination unit, configured to generate a next prediction sequence according to the candidate set based on a preset language model, and to form a sequence path by combining the next prediction sequence and each first prediction sequence in the candidate set;

[0150] An operation and maintenance suggestion determination unit is used to determine the probability product of the predicted probability of the next predicted sequence and the predicted probability of the first predicted sequence in each sequence path based on a preset language model, extract a preset number of sequence paths as candidate sets in descending order of the probability products, and use the sequence paths in the candidate sets as operation and maintenance suggestions until a preset condition is met;

[0151] The preset conditions include that the probability product is less than a preset probability product or the sequence path reaches a preset length.

[0152] An operation and maintenance device for a low-voltage power distribution system provided by an embodiment of the present invention can execute an operation and maintenance method for a low-voltage power distribution system provided by any embodiment of the present invention, and has functional modules and beneficial effects corresponding to the execution method.

[0153] Example 5

[0154] Figure 11 1 is a schematic diagram of the structure of an electronic device for implementing a method for operating and maintaining a low-voltage power distribution system according to an embodiment of the present invention. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices (such as helmets, glasses, watches, etc.) and other similar computing devices. The components shown herein, their connections and relationships, and their functions are provided for example only and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0155] like Figure 11 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., which is communicatively connected to the at least one processor 11. The memory stores a computer program that can be executed by the at least one processor. The processor 11 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 12 or the computer program loaded from the storage unit 18 into the random access memory (RAM) 13. Various programs and data required for the operation of the electronic device 10 can also be stored in the RAM 13. The processor 11, ROM 12, and RAM 13 are connected to each other via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0156] Multiple components in the electronic device 10 are connected to the I / O interface 15, including an input unit 16, such as a keyboard, a mouse, etc.; an output unit 17, such as various types of displays, speakers, etc.; a storage unit 18, such as a magnetic disk, an optical disk, etc.; and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0157] The processor 11 can be any general-purpose and / or specialized processing component with processing and computing capabilities. Some examples of the processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various processors that run machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The processor 11 executes the various methods and processes described above, such as a method for operating and maintaining a low-voltage power distribution system.

[0158] In some embodiments, a method for operating and maintaining a low-voltage power distribution system may be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as a storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 10 via the ROM 12 and / or the communication unit 19. When the computer program is loaded into the RAM 13 and executed by the processor 11, one or more steps of the method for operating and maintaining a low-voltage power distribution system described above may be performed. Alternatively, in other embodiments, the processor 11 may be configured to execute a method for operating and maintaining a low-voltage power distribution system in any other appropriate manner (e.g., by means of firmware).

[0159] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system comprising at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0160] Computer programs for implementing the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when the computer program is executed by the processor, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The computer program may be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0161] In the context of the present invention, computer-readable storage media can be tangible media that can contain or store a computer program for use with an instruction execution system, device or equipment or used in combination with an instruction execution system, device or equipment. Computer-readable storage media can include but are not limited to electronic, magnetic, optical, electromagnetic, infrared or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. Alternatively, computer-readable storage media can be machine-readable signal media. More specific examples of machine-readable storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0162] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0163] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0164] A computing system may include clients and servers. The clients and servers are typically remote from each other and typically interact via a communication network. This client-server relationship arises through computer programs running on the respective computers, creating a client-server relationship. The server may be a cloud server, also known as a cloud computing server or cloud host. This server is a hosting product within the cloud computing service ecosystem that addresses the management difficulties and limited scalability of traditional physical hosting and VPS services.

[0165] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present invention can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solution of the present invention can be achieved. This is not limited herein.

[0166] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A method for operating and maintaining a low-voltage power distribution system, characterized in that: include: Acquire natural language input by a user based on an interactive device, determine the operation and maintenance entity and query intent of the natural language as sentence features, and determine original time series data in the low-voltage power distribution system based on the query features; Determining data feature information and data anomaly results corresponding to the original time series data; Determining a diagnosis result corresponding to the natural language based on the data abnormality result and the query feature; Based on a preset language model, the operation and maintenance suggestions corresponding to the natural language are determined according to the data feature information, and the diagnosis results and the operation and maintenance suggestions are fed back to the interactive device so that the interactive device can visually display the diagnosis results and the operation and maintenance suggestions.

2. The method according to claim 1, characterized in that The method of obtaining natural language input by a user based on an interactive device, determining an operation and maintenance entity and a query intent of the natural language as sentence features, and determining original time series data in a low-voltage power distribution system based on the query features includes: Receive natural language input from a user on an interactive device, and input the natural language into a preset intent classification model to determine the query intent corresponding to the natural language; wherein the preset intent classification model is a deep learning model based on a bidirectional long short-term memory network combined with an attention mechanism; Inputting the natural language into a preset sequence annotation model to determine the operation and maintenance entity corresponding to the natural language, and using the operation and maintenance entity and the query intent as sentence features; wherein the preset sequence annotation model is a deep learning model combining conditional random fields with neural networks; According to the operation and maintenance entity and the query intention, operation data of the operation and maintenance entity in the low-voltage power distribution system within a preset time period collected by a preset sensor is extracted as original time series data.

3. The method according to claim 2, characterized in that After receiving the natural language input by the user on the interactive device, the method further includes: When it is determined that any one of the operation and maintenance entity and the query intention does not exist, problem tracking information is generated based on the interactive device, so that the user can supplement the operation and maintenance entity and / or the query intention according to the problem tracking information.

4. The method according to claim 1, wherein Determining the data feature information and data anomaly results corresponding to the original time series data includes: Performing data preprocessing on the original time series data to obtain target sequence data, and determining data anomaly results of the target sequence data according to a preset running data interval threshold; wherein the data anomaly results include data anomaly and data normality; Determine the data categories included in the operation and maintenance entity operation data corresponding to the same time in the target sequence data and the data corresponding to each data category, and discretize each data according to a preset interval of the data category to obtain data features; The data features are mapped into a vector space of fixed dimension to obtain a token sequence of the operation data of the operation and maintenance entity corresponding to each time, and the token sequence is spliced ​​in a preset time order as the data feature information.

5. The method according to claim 4, characterized in that The determining of the diagnosis result corresponding to the natural language based on the data abnormality result and the query feature includes: When it is determined that the data abnormality result is normal data, the diagnosis result is determined to be that the operation and maintenance entity in the query feature is normal; When it is determined that the data anomaly result is a data anomaly, the abnormal data in the target sequence data is extracted, and the diagnosis result corresponding to the natural language is determined by matching the abnormal data and the operation and maintenance entity in the preset knowledge graph.

6. The method according to claim 1, characterized in that The training based on the preset language model includes: Obtaining preset general text data, and inputting the preset general text data into the preset language model as a first training set, so that the preset language model performs a first-stage pre-training task according to the first training set data; wherein the preset language model is a large language model based on the Transformer architecture, and the first-stage pre-training task of the preset language model includes a masked language model task and a next sentence prediction task; Acquire power system domain text data, and input the power system domain text data into the preset language model as a second training set, so that the preset language model performs a second-stage pre-training task according to the second training set; wherein the second-stage pre-training task includes a masked language model task, a distribution system fault chain prediction task, and an equipment state reasoning task; The historical operation and maintenance records of the low-voltage power distribution system are extracted, and the historical operation and maintenance records are input into the preset language model as a third training set. The hyperparameters of the preset language model are adjusted using a supervised fine-tuning strategy until the accuracy of the output result of the preset language model is greater than the preset probability, thereby completing the training of the preset language model.

7. The method according to claim 1, characterized in that The determining, based on the preset language model and according to the data feature information, the operation and maintenance suggestion corresponding to the natural language includes: Inputting the data feature information into a preset language model, determining a probability distribution of a first prediction sequence corresponding to the data feature information based on the preset language model, and extracting a preset number of first prediction sequences in the probability distribution as candidate sets in descending order of prediction probability; generating a next prediction sequence according to the candidate set based on the preset language model, and forming a sequence path by combining the next prediction sequence and each of the first prediction sequences in the candidate set; Determining, based on the preset language model, the probability product of the predicted probability of the next predicted sequence and the predicted probability of the first predicted sequence in each of the sequence paths, extracting a preset number of sequence paths as candidate sets in descending order of the probability products until a preset condition is met, and using the sequence paths in the candidate sets as operation and maintenance recommendations; The preset conditions include that the probability product is less than a preset probability product or the sequence path reaches a preset length.

8. An operation and maintenance device for a low-voltage power distribution system, characterized in that: include: A data acquisition module is used to acquire natural language input by a user based on an interactive device, determine the operation and maintenance entity and query intent of the natural language as sentence features, and determine the original time series data in the low-voltage power distribution system based on the query features; A result determination module is used to determine the data feature information and data anomaly results corresponding to the original time series data; A diagnosis result determination module, configured to determine a diagnosis result corresponding to the natural language based on the data abnormality result and the query feature; An operation and maintenance suggestion determination module is used to determine the operation and maintenance suggestion corresponding to the natural language according to the data feature information based on a preset language model, and to feed back the diagnosis result and the operation and maintenance suggestion to the interactive device so that the interactive device can visually display the diagnosis result and the operation and maintenance suggestion.

9. An electronic device, characterized in that: The electronic device comprises: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the operation and maintenance method of the low-voltage power distribution system according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the operation and maintenance method of the low-voltage power distribution system according to any one of claims 1 to 7 when executed.

Images