Power grid intelligent operation and maintenance early warning method and system

Abstract

The embodiment of the invention provides a power grid intelligent operation and maintenance early warning method and system, and the method comprises the steps: obtaining a historical first data set and a historical second data set of a power grid, and marking fault data in the historical first data set and the historical second data set based on a fault event and a fault scene to form a fault data set; performing data enhancement on the fault data set to construct a sample training set, and inputting the sample training set into a fault early warning model for training to form a trained fault early warning model; and collecting a first data set and a second data set of the power grid in real time and inputting the first data set and the second data set into the trained fault early warning model for fault early warning. According to the scheme, enhancement processing is carried out on the fault data subsets with small sample sizes in multiple modes, meanwhile, balance processing is carried out on the fault data subsets with different sample sizes, and the generalization ability of the model and the accuracy of fault early warning can be improved.

Description

Technical Field

[0001] This application relates to the field of power grid technology, specifically to a method and system for intelligent operation and maintenance early warning of power grids. Background Technology

[0002] With the rapid development of society and the economy, electricity, as a fundamental energy source, is crucial for ensuring the stability and reliability of its supply, which is essential for guaranteeing the normal operation of industrial production, residential life, and all sectors of society. As a key infrastructure for power transmission and distribution, the power grid is constantly expanding in scale and becoming increasingly complex in structure. Modern power grids encompass a large number of power generation devices, transmission lines, substations, and distribution equipment, distributed across a vast geographical area in a complex and variable operating environment. In the actual operation of the power grid, faults are unavoidable, not only causing power supply interruptions and affecting users' normal electricity consumption, but also potentially triggering safety accidents, resulting in huge economic losses and social impacts. Therefore, timely and accurate detection and early warning of potential faults in power grid equipment are of great significance for ensuring the safe and stable operation of the power grid.

[0003] Traditional power grid operation and maintenance early warning methods mainly rely on single electrical data. For some complex fault modes, they are often difficult to judge accurately, resulting in low accuracy and reliability of early warning. In addition, for fault scenarios in the power grid that have a low probability of occurrence but serious harm, such as multiple faults caused by lightning strikes in a specific area of ​​the power grid or rare equipment composite insulation defect faults, it is difficult to accumulate enough historical data for model training because these faults occur very rarely. This causes the early warning model to be prone to overfitting in small sample scenarios and fail to accurately generalize to actual fault scenarios. Summary of the Invention

[0004] This application aims to provide a method and system for intelligent operation and maintenance early warning of power grids, which can improve the accuracy of power grid fault early warning.

[0005] The technical solution of this application is implemented as follows: In a first aspect, embodiments of this application provide a smart power grid operation and maintenance early warning method, the method comprising: acquiring a historical first dataset and a historical second dataset of the power grid, wherein the first dataset includes electrical data, environmental data and status data, and the second dataset includes image data and video data; A fault dataset is formed by labeling fault data in the first historical dataset and the second historical dataset based on fault events and fault scenarios. The fault dataset is augmented to construct a sample training set, and the sample training set is input into the fault early warning model for training to form a trained fault early warning model. The first and second datasets of the power grid are collected in real time and input into the trained fault early warning model for fault early warning.

[0006] In one specific implementation, the electrical data includes voltage data, current data, power data, frequency data, and harmonic data; the environmental data includes temperature data, humidity data, wind direction data, wind speed data, lightning data, rain and snow data, and pollution data; the status data includes equipment switch status, equipment operating status, and equipment alarm status; the image data includes equipment appearance images, infrared thermal imaging images, and on-site environmental images; and the video data includes equipment operation videos, fault process videos, and maintenance operation videos.

[0007] In one specific implementation, a fault dataset is formed by labeling fault data in the first historical dataset and the second historical dataset based on fault events and fault scenarios, including: Based on the recorded time points of fault events and fault scenarios, a time window is determined according to preset time expansion rules; Multiple fault data are obtained by matching the first historical dataset and the second historical dataset according to the time window; Multiple fault data are labeled with fault content stamps and fault timestamps according to a preset labeling format to form a fault dataset; wherein the fault content stamps include fault type, fault location, fault device identifier, and fault severity.

[0008] In one specific implementation, the fault dataset is augmented to construct a sample training set, including: The fault dataset is divided into multiple fault data subsets based on the fault type. Based on the data volume, the multiple fault data subsets are divided into small sample size fault data subsets, medium sample size fault data subsets, and large sample size fault data subsets; The small sample size fault data subset is augmented to form an augmented small sample size fault data subset. The enhanced small-sample-size fault data subset, the medium-sample-size fault data subset, and the large-sample-size fault data subset are merged to form a sample training set.

[0009] In one specific implementation, data augmentation is performed on the small sample size fault data subset to form an augmented small sample size fault data subset, including: The electrical data in the small sample size fault data subset is linearly scaled and time-series shifted to obtain an enhanced electrical data subset; The environmental data in the small sample size fault data subset is interpolated to obtain an enhanced environmental data subset. An enhanced state data subset is obtained by combining state data and state time-series changes from the state data subset with the small sample size of the fault data subset. The image data in the small sample size fault data subset is geometrically transformed, color adjusted, and occluded to obtain an enhanced image data subset; An enhanced video data subset is obtained by performing video stitching and video frame operations on the video data in the small sample size fault data subset; wherein, the video frame operations include video frame insertion and video frame repetition; The enhanced electrical data subset, the enhanced environmental data subset, the enhanced status data subset, the enhanced image data subset, and the enhanced video data subset are combined to form the enhanced small sample size fault data subset.

[0010] In one specific implementation, the following steps are included before constructing the sample training set: The enhanced small-sample-size fault data subset and the large-sample-size fault data subset are balanced based on the fault type; wherein the balancing process includes oversampling the enhanced small-sample-size fault data subset and undersampling the large-sample-size fault data subset. The medium-sized fault data subset, the balanced enhanced small-sized fault data subset, and the large-sized fault data subset are combined to form a sample training set. The sample training set is divided into a training set and a validation set according to a preset ratio.

[0011] In one specific implementation, the sample training set is input into the fault early warning model for training, including: The training set is input into the fault warning model to obtain the first fault warning result; The loss value is calculated using the cross-entropy loss function based on the first fault warning result, and the gradient value of the fault warning model is calculated using the backpropagation algorithm based on the loss value. The parameters of the fault warning model are updated using the stochastic gradient descent algorithm based on the gradient values. The validation set is input into the updated fault warning model to obtain the second fault warning result. The evaluation result is calculated based on the accuracy and recall. The parameters of the updated fault warning model are adjusted according to the evaluation result to obtain the final fault warning model.

[0012] In one specific implementation, the fault warning model includes a multimodal feature extraction layer, a multimodal feature fusion layer, and a classification prediction layer; wherein the multimodal feature extraction layer includes a one-dimensional convolutional sub-layer, a two-dimensional convolutional sub-layer, and a three-dimensional convolutional sub-layer; the multimodal feature fusion layer includes a feature vector concatenation sub-layer, an attention weight sub-layer, and a weighted fusion sub-layer; and the classification prediction layer includes a fully connected sub-layer, an activation function sub-layer, and an output sub-layer.

[0013] Secondly, embodiments of this application provide a smart grid operation and maintenance early warning system, comprising: The acquisition unit is used to acquire a historical first dataset and a historical second dataset of the power grid, wherein the first dataset includes electrical data, environmental data and status data, and the second dataset includes image data and video data; Annotation unit is used to annotate fault data in the first historical dataset and the second historical dataset based on fault events and fault scenarios to form a fault dataset; The training unit is used to perform data augmentation on the fault dataset to construct a sample training set, and input the sample training set into the fault early warning model for training to form a trained fault early warning model. The early warning unit is used to collect the first and second datasets of the power grid in real time and input them into the trained fault early warning model for fault early warning.

[0014] Thirdly, embodiments of this application provide a smart grid operation and maintenance early warning device, which includes: a processor and a memory; wherein, The memory is used to store computer programs; The processor is configured to call and run the computer program from the memory to perform the method as described in the first aspect.

[0015] The beneficial effects of this application are: This application proposes a power grid intelligent operation and maintenance early warning method. 1. By integrating electrical data, environmental data, status data, and image / video data, a spatiotemporal-physical-visual multi-dimensional feature matrix is ​​formed, overcoming the limitations of a single data source and achieving comprehensive capture of fault features, thus improving the comprehensiveness of early warning. 2. Based on the time point of the fault event and fault scenario, the time window is extended to accurately match the associated data before and after the fault occurrence, avoiding the omission of potential precursor features. 3. Data augmentation is performed on the fault dataset, especially for small sample size fault data subsets, using various augmentation methods such as linear scaling and time series shifting of electrical data, and geometric transformation of image data, increasing data diversity. Simultaneously, balancing processing is performed on fault data subsets of different sample sizes to avoid model bias caused by data imbalance, further improving the model's generalization ability and the accuracy of fault early warning. 4. The fault early warning model includes a multimodal feature extraction layer, a multimodal feature fusion layer, and a classification prediction layer, which can effectively extract and fuse features from different modal data. The classification prediction layer accurately judges the operating status of the power grid, improving the accuracy of fault early warning. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the technical solutions of this application. Obviously, the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0017] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0018] Figure 1 An optional flowchart illustrating a smart power grid operation and maintenance early warning method provided in this application embodiment. Figure 1 ; Figure 2 An optional flowchart illustrating a smart power grid operation and maintenance early warning method provided in this application embodiment. Figure 2 ; Figure 3 An optional flowchart illustrating a smart power grid operation and maintenance early warning method provided in this application embodiment. Figure 3 ; Figure 4 This application provides a schematic diagram of the structure of a smart power grid operation and maintenance early warning system. Figure 5This is a schematic diagram of the structure of a smart power grid operation and maintenance early warning device provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the specific technical solutions of this application will be further described in detail below with reference to the accompanying drawings of the embodiments of this application. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] In the following description, references to "some embodiments," "this embodiment," "this application embodiment," and examples, etc., describe a subset of all possible embodiments. However, it is understood that "some embodiments" may be the same subset or different subset of all possible embodiments and may be combined with each other without conflict.

[0022] If the application documents contain similar descriptions such as "first / second", the following explanation shall be added: In the following description, the terms "first / second / third" are used only to distinguish similar objects and do not represent a specific order of objects. It is understood that "first / second / third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0023] This application provides a method for intelligent operation and maintenance early warning of power grids. Figure 1 An optional flowchart illustrating a smart power grid operation and maintenance early warning method provided in this application embodiment. Figure 1 , will combine Figure 1 The steps shown are explained.

[0024] S101. Obtain the first historical dataset and the second historical dataset of the power grid, wherein the first dataset includes electrical data, environmental data and status data, and the second dataset includes image data and video data.

[0025] In one embodiment of this application, the electrical data includes voltage data, current data, power data, frequency data, and harmonic data; the environmental data includes temperature data, humidity data, wind direction data, wind speed data, lightning data, rain and snow data, and pollution data; the status data includes equipment switch status, equipment operating status, and equipment alarm status; the image data includes equipment appearance images, infrared thermal imaging images, and on-site environmental images; and the video data includes equipment operation videos, fault process videos, and maintenance operation videos.

[0026] Specifically, voltage and current data can be obtained through voltage and current transformers. Based on the rated voltage and current of the equipment, transformers with appropriate range and accuracy are selected. For example, for high-voltage transmission lines, voltage transformers with an accuracy of 0.2 class and a range selected according to the line voltage level are used, as well as current transformers with an accuracy of 0.2S class and a range matching the line current are used. For low-voltage power distribution equipment, transformers with an accuracy of 0.5 class and a suitable range are selected. Power data can be obtained through power measuring instruments. Power data includes active power, reactive power, apparent power, etc. Multifunctional power meters can simultaneously measure parameters such as active power, reactive power, and apparent power. Frequency data can be obtained through a frequency meter to measure the frequency of the power grid. The accuracy should reach ±0.01Hz, and it should be able to respond quickly to changes in frequency. Harmonic data can be obtained using a harmonic analyzer to analyze the harmonic components in electrical signals and measure parameters such as the content and distortion rate of each harmonic. The analyzer's measurement range should cover the 2nd to 50th harmonics, and its accuracy should meet the relevant standard requirements. Temperature data can be obtained through temperature sensors. The temperature data includes equipment temperature and ambient temperature. The measurement range of the temperature sensor is determined according to the environment of the power grid, generally -40℃ to +85℃, with an accuracy of ±0.1℃. In areas with dense equipment such as substations and power plants, temperature sensors should be reasonably arranged to monitor the temperature of equipment and the environment. Humidity data can be obtained through a humidity sensor. A capacitive humidity sensor is used, with a measurement range of 0%-100%RH and an accuracy of ±2%RH. It is used in conjunction with a temperature sensor and installed near the equipment to monitor the ambient humidity in real time. Wind direction and wind speed data can be obtained through wind direction and wind speed sensors. Ultrasonic wind direction and wind speed sensors can be installed in outdoor power transmission line towers, substations and other places to accurately measure wind direction and wind speed. The wind direction measurement accuracy is ±3° and the wind speed measurement range is 0-60m / s with an accuracy of ±0.1m / s. Lightning data can be obtained through lightning monitoring equipment or by using a lightning location system. By rationally deploying lightning detection stations within the power grid area, information such as the location, intensity, and time of lightning occurrence can be monitored in real time. The detection radius of the detection station should be determined based on the coverage area of ​​the power grid and the terrain. Rain and snow data can be obtained through rain and snow sensors. Infrared rain and snow sensors can be selected to accurately determine whether it is raining or snowing and to measure the amount of rainfall or snowfall. The sensors should be waterproof and dustproof and should be installed in an open outdoor area. Pollution data can be obtained through pollution sensors. For potential pollution of power grid equipment, gas pollution sensors (such as those for detecting the concentration of gases such as sulfur dioxide and nitrogen oxides) and particulate pollution sensors (such as those for measuring the concentration of particulate matter such as PM2.5 and PM10) can be installed. The measurement range and accuracy of the sensors should meet the requirements of environmental monitoring. The switching status of the equipment can be obtained through the equipment switching status monitoring device. By installing auxiliary contacts at the switching contacts of the equipment or using non-contact sensors (such as Hall sensors), the switching status (closed or open) of the equipment can be monitored in real time. The monitoring device should have status indication and communication functions and be able to upload the switching status information to the data acquisition system in a timely manner. The operating status of equipment can be obtained through equipment operating status monitoring devices. Different monitoring methods are used depending on the type and operating characteristics of the equipment. For rotating equipment (such as generators and motors), vibration sensors can be installed to monitor parameters such as vibration frequency and amplitude to determine whether the equipment is operating normally. For equipment such as transformers, the operating status of the equipment can be assessed by monitoring parameters such as oil temperature, oil level, and gas content. The monitoring device should have data storage and communication functions to upload the operating status data of the equipment in real time. For equipment alarm status, it can be obtained through equipment alarm status acquisition device, which can be connected with equipment protection devices, monitoring systems, etc. to collect equipment alarm information in real time. Alarm information includes fault alarms such as overcurrent, overvoltage, undervoltage, overload, and short circuit, as well as equipment abnormal operation alarms. The acquisition device should be able to accurately record the time, type and equipment identification information of the alarm. For images of the equipment's appearance, high-definition cameras can be used. High-definition cameras should be installed in key parts of the equipment (such as transformers, circuit breakers, disconnect switches, etc.) to capture images of the equipment's appearance. The resolution of the cameras should be no less than 1080P, and they should have functions such as autofocus and night vision to clearly capture the details of the equipment's appearance. The number of cameras and the shooting angles should be reasonably determined according to the importance of the equipment and the monitoring requirements. Infrared thermal imaging images can be acquired using an infrared thermal imager. By installing an infrared thermal imager, the temperature distribution of the equipment can be monitored in real time. The resolution of the thermal imager should be no less than 320×24 pixels, and the temperature measurement range is determined according to the operating temperature of the equipment, generally -20℃ to +500℃, with an accuracy of ±2℃. Infrared thermal imagers can detect overheating faults in equipment in a timely manner, allowing for early warning and handling. For on-site environmental images, they can be obtained through on-site environmental cameras. On-site environmental cameras can be installed around substations, power plants, and other places to capture on-site environmental images and monitor for abnormal situations such as foreign object intrusion, fire, and flood. The installation location of the camera should be able to cover the entire on-site environment and have a wide-angle shooting function. Video recording devices can be used to record equipment operation videos, fault process videos, and maintenance operation videos. These devices can be installed in areas prone to faults (such as fault-prone sections of transmission lines and important equipment areas of substations). The video recording devices should have functions such as high-definition recording, long-term storage, and remote playback. The recording resolution should be no less than 1080P, and the storage capacity can be configured according to actual needs.

[0027] In one embodiment of the present invention, different sampling frequencies can be set according to the type and importance of the equipment. For important equipment such as generators and transformers, the sampling frequency of voltage data, current data, and power data can be set to 10 times per second. For general power distribution equipment, the sampling frequency can be set to once per minute. The sampling frequency of frequency data and harmonic data is determined according to the monitoring requirements and can be set to once per second. Synchronous sampling technology is adopted to ensure that electrical data are collected at the same time to ensure the accuracy and consistency of the data. The timestamp of each sampled data can be recorded with millisecond accuracy.

[0028] In one embodiment of the present invention, environmental data can be collected continuously to monitor changes in environmental parameters in real time and to periodically calibrate environmental sensors to ensure the accuracy of measurement data.

[0029] In one embodiment of the present invention, the on / off status and operating status of the device are monitored in real time. When the status changes, data acquisition is immediately triggered to record information such as the time, type and device identification of the status change. For the alarm status of the device, alarm information is collected in a timely manner when an alarm signal is received.

[0030] In one embodiment of the present invention, image data is captured at set intervals (e.g., every minute). Simultaneously, when an abnormal device status or a change in the on-site environment is detected, image capture is triggered to record the abnormal situation in a timely manner.

[0031] In one embodiment of the present invention, the video recording device records video according to a preset recording mode (e.g., continuous recording, event-triggered recording), and the recorded video data can use video compression technology to reduce the bandwidth requirements of the video stream.

[0032] In one embodiment of the present invention, a first historical dataset and a second historical dataset of the power grid over a period of 5 to 10 years can be obtained to obtain diverse data sources for subsequent fault datasets.

[0033] S102. Based on the fault events and fault scenarios, fault data is labeled in the first historical dataset and the second historical dataset to form a fault dataset, such as... Figure 2 As shown, Figure 2 An optional flowchart illustrating a smart power grid operation and maintenance early warning method provided in this application embodiment. Figure 2 Specifically, it includes the following steps: S1021. Based on the recorded time points of fault events and fault scenarios, determine the time window according to the preset time expansion rules; In one embodiment of the present invention, a fault event refers to an abnormal condition that occurs during the operation of the power grid, deviating from the normal operating parameters and state of the power grid, which may lead to adverse consequences such as equipment damage and power outages. This may include short-circuit fault events: a direct short circuit between conductive parts at different potentials in the power grid, resulting in a sharp increase in current; overload fault events: the current carried by equipment or lines in the power grid exceeds its rated capacity, but has not yet reached the level of a short circuit; grounding fault events: an abnormal electrical connection occurs between conductors in the power grid and the ground; and equipment aging fault events: the performance of equipment in the power grid deteriorates due to long-term operation, environmental factors, etc. The decline gradually leads to a failure. Failure scenarios refer to various aspects such as the specific environment, operating mode, equipment status, and meteorological conditions at the time of the failure event. These can include failure scenarios under severe weather conditions: in severe weather conditions such as heavy rain, lightning, strong winds, and snow, power grid equipment is easily affected by external factors and may fail; failure scenarios after equipment maintenance: after power grid equipment has completed maintenance and been put back into operation, failures may occur due to negligence during maintenance or improper equipment installation; failure scenarios after power grid restructuring: when the power grid undergoes structural adjustments, such as the commissioning of new lines, the decommissioning of old lines, or the capacity expansion of transformers, failure scenarios may occur. Subsequently, changes in the power grid's operation mode and power flow distribution may trigger new faults; Multiple fault concurrency scenarios: During power grid operation, multiple faults may occur simultaneously or sequentially within a short period. These faults influence and are interconnected, making fault diagnosis and handling more difficult; Based on the recorded information of fault events and scenarios, the recorded time points of the faults are obtained, and further screening is performed to identify key time points closely related to the occurrence, development, and termination of the fault; for example, the fault initiation time (the time when electrical parameters first become abnormal), the fault development turning point time (the time when the current suddenly increases), and the fault termination time (the time when the equipment returns to normal operation or the protection device...). The time range includes factors such as the reset time; for different fault events and scenarios, corresponding time expansion factors are determined, including forward expansion factors and backward expansion factors. The forward expansion factor represents the time range to be expanded before the fault recording time point, and the backward expansion factor represents the time range to be expanded after the fault recording time point. For example, for short-circuit fault time, the forward expansion factor can be set to 0.05 seconds before the fault occurs, and the backward expansion factor can be set to 0.2 seconds after the fault occurs; for equipment aging fault events, the forward expansion factor can be set to 1 hour before the fault becomes obvious, and the backward expansion factor can be set to 0.2 seconds after the fault is repaired.Considering the complexity of actual fault situations, a dynamic adjustment mechanism for the time expansion factor is established for a 5-hour timeframe. This mechanism adjusts the time expansion factor in real time based on factors such as the severity, scope of impact, and environmental conditions of the fault event and scenario. For example, when a fault causes a large-scale power outage, the backward expansion factor is appropriately increased to include more comprehensive fault recovery process data; when a fault occurs under severe weather conditions, the forward expansion factor is appropriately increased to capture the impact of potential environmental factors on the fault. Based on the recorded time point and the preset time expansion factor, the start and end times of the time window are calculated. For example, if the fault recording time point is T, the forward expansion factor is ΔT1, and the backward expansion factor is ΔT2, then the start time of the time window is T-ΔT1, and the end time is T+ΔT2.

[0034] S1022. Based on the time window, match the first historical dataset and the second historical dataset to obtain multiple fault data; In one embodiment of the present invention, electrical data, environmental data, and status data that conform to the time range are retrieved from a first historical dataset according to the time window; similarly, image data and video data that conform to the time range are retrieved from a second historical dataset according to the time window. For image data, matching can be performed by comparing the image capture time with the time window range; for video data, video editing techniques can be used to extract video segments within the time window, or further filtering can be performed based on device identifiers, scene features, etc. in the images and videos. Considering that there may be slight time differences in data from different data sources, a time alignment algorithm is used to align the data from different data sources according to the time axis to ensure that the matched data are synchronized in time. For example, for electrical data and image data, the closest time point can be found for matching based on the timestamp of the electrical data and the capture time of the image; at the same time, spatial correlation of data is considered to ensure that the devices in the images and videos correspond to the devices in the electrical data and status data.

[0035] S1023. Label multiple fault data with fault content stamps and fault timestamps according to a preset labeling format to form a fault dataset; wherein the fault content stamps include fault type, fault location, fault equipment identifier and fault severity.

[0036] In one embodiment of the present invention, the fault content stamp may include information such as fault type, fault location, fault equipment identifier, fault severity, and fault cause. The specific annotation format can be as follows: For fault type, different fault types can be classified according to fault events and fault scenarios. Fault types can include short-circuit faults, overload faults, grounding faults, equipment aging faults, lightning strike faults, icing faults, etc., and can be represented using a unified coding system, for example, 01 represents a short-circuit fault, 02 represents an overload fault, etc. The fault location can be accurate to the specific equipment installation location or line segment, using Geographic Information System (GIS) coordinates for positioning. The fault equipment identifier can use the equipment's unique number. The fault severity can be divided into four levels: minor, moderate, severe, and critical, and represented by corresponding numerical values. The fault cause can be classified according to the fault characteristics and analysis results, such as equipment aging, external force damage, improper operation, etc. The fault timestamp should include the accurate time of the fault occurrence, accurate to the millisecond level, using a standard time format, such as ISO 8601 format, to ensure the standardization and readability of the timestamp. Simultaneously, the time zone information of the fault timestamp can be recorded to avoid time confusion caused by time zone differences.

[0037] In one embodiment of the present invention, an automated annotation tool can be developed to automatically annotate the selected fault data according to a preset annotation format. The automated annotation tool can automatically fill in fault content stamps and fault timestamps. For example, it can automatically determine the fault type based on fault characteristics in electrical data, determine the fault location based on equipment identification and GIS coordinates, and determine the fault severity based on the duration and development of the fault. At the same time, it can automatically record the annotation time and annotation personnel information for easy traceability and management. Since automated annotation may have certain errors, it can be manually corrected and supplemented. For example, for some complex fault scenarios, automated annotation may not be able to accurately determine the fault cause, requiring manual analysis and annotation by annotation personnel. At the same time, annotation personnel can also supplement some additional information based on the actual situation, such as the fault handling process and the scope of fault impact. Furthermore, an annotation quality inspection mechanism can be established to conduct quality checks on the annotated fault data. A sampling inspection method can be used to randomly select a certain proportion of the annotated data for manual review to ensure the accuracy and completeness of the annotation information. Finally, the annotated fault data is integrated to generate a fault dataset.

[0038] In one embodiment of the present invention, a suitable storage method can be selected to store the fault dataset; a database storage method, such as MySQL, Oracle, MongoDB, Redis, etc., can be used to facilitate data management and query; or a file system storage method can be used to store the fault dataset in a specific folder.

[0039] S103. Perform data augmentation on the fault dataset to construct a sample training set, and input the sample training set into the fault early warning model for training to form a trained fault early warning model.

[0040] Furthermore, augmenting the fault dataset to construct a sample training set may include the following steps: S1031. Divide the fault dataset into multiple fault data subsets based on the fault type; For example, based on different fault types, such as short circuit faults, overload faults, equipment aging faults, lightning strike faults, icing faults, pollution flashover faults, misoperation faults, and construction damage faults, an independent fault data subset can be created for each fault type. Wavelet transform or STFT time-frequency analysis can be used to classify the fault data, and the classification can assist in manual verification.

[0041] S1032. Based on the data volume, the multiple fault data subsets are divided into small sample size fault data subsets, medium sample size fault data subsets, and large sample size fault data subsets. For example, the number of samples in each fault subset is calculated, and a sample division threshold is set. The threshold for small samples is set to <1000 samples, the threshold for medium samples is set to 1000-10000 samples, and the threshold for large samples is set to >10000 samples. Finally, the fault data subset is divided into a small sample size fault data subset, a medium sample size fault data subset, and a large sample size fault data subset.

[0042] S1033. The small sample size fault data subset is augmented to form an augmented small sample size fault data subset. For example, selecting corresponding data augmentation methods for different data types within a small subset of fault data can also correlate fault data with specific physical processes or system models. By adjusting model parameters and simulating system responses under different fault conditions, corresponding simulated fault data samples can be generated. When performing data augmentation, it is necessary to ensure that the augmented data still retains the fault characteristics and semantic information of the original data, avoiding the introduction of unreasonable or erroneous fault modes. At the same time, the intensity and scope of data augmentation should be reasonably controlled to prevent excessive augmentation from causing data distortion or model overfitting. Finally, the new samples generated through data augmentation methods are merged with the original small subset of fault data to form an augmented small subset of fault data, increasing the diversity and quantity of data and providing richer and more comprehensive data support for subsequent tasks such as fault diagnosis and model training.

[0043] S1034. The enhanced small-sample-size fault data subset, the medium-sample-size fault data subset, and the large-sample-size fault data subset are merged to form a sample training set.

[0044] For example, the data from the enhanced small-sample-size fault data subset, the medium-sample-size fault data subset, and the large-sample-size fault data subset can be concatenated in sequence to form a sample training set.

[0045] In one embodiment of the present invention, before forming the sample training set, the method further includes balancing the enhanced small-sample-size fault data subset and the large-sample-size fault data subset based on the fault type; wherein the balancing process includes oversampling the enhanced small-sample-size fault data subset and undersampling the large-sample-size fault data subset; merging the medium-sample-size fault data subset, the balanced enhanced small-sample-size fault data subset, and the large-sample-size fault data subset to form the sample training set; and dividing the sample training set into a training set and a validation set according to a preset ratio.

[0046] For example, the number of samples for each fault type in the enhanced small-sample-size fault data subset and the large-sample-size fault data subset is statistically analyzed to form a mapping table of fault type and sample number. For example, fault types A, B, C, and D are statistically analyzed, and fault type A has 10,000 samples, fault type B has 80 samples, fault type C has 30 samples, and fault type D has 20,000 samples. The fault type with the fewest samples is used as the benchmark, or a reasonable benchmark sample number is set according to actual needs and experience. In this embodiment, the number of samples for fault type C with the fewest samples (30) is used as the benchmark sample number for illustration. After balancing, the sample number for each fault type is a multiple of the benchmark sample number, and the multiple can be 10 to 1000. Oversampling is performed on the enhanced small-sample-size fault data subset, which can be done using random oversampling or the SMOTE (Synthetic Minority Over-sampling Technique) algorithm. Random oversampling increases the number of minority class samples by randomly copying them; SMOTE... The algorithm analyzes the feature space of the samples and synthesizes new samples from adjacent samples, which can better preserve the distribution characteristics of the data. The following process uses the SMOTE algorithm as an example to illustrate: Calculate the k nearest neighbor samples of fault type C (the value of k can be set according to the characteristics of the data and experience, and is generally around 5). Randomly select one sample from the k nearest neighbor samples, and then randomly select a point on the line connecting the two samples as the new synthesized sample; Repeat the above steps until a sufficient number of new samples are generated, such that the number of samples for fault type C reaches a multiple (e.g., 10 times) of the baseline number of samples.

[0047] For example, undersampling a large subset of fault data can be performed using random undersampling, the Tomek Links algorithm, or the Edited Nearest Neighbors (ENN) algorithm. Random undersampling randomly deletes samples from the majority class; the Tomek Links algorithm deletes boundary samples from the majority class by finding boundary sample pairs between different classes; and the ENN algorithm determines whether to delete a sample based on its nearest neighbor class. The following steps use random undersampling as an example: For samples of fault type D, a random undersampling method is used to reduce the number of samples and determine the number of samples to be deleted, which is the original number of samples minus the target number of samples (e.g., 20000 - 30 * 300 = 11000).

[0048] 11,000 samples of this fault type are randomly selected and deleted, so that the number of remaining samples is a multiple of the baseline sample number (e.g., 300 times).

[0049] Finally, the training set and validation set are divided. The ratio of the training set and validation set is set according to the actual needs and model training requirements. It can be set to 7:3, 8:2, etc. Taking 7:3 as an example, for each fault type in the balanced sample training set, it is divided into training set and validation set according to the preset ratio. For each fault type, 70% of the total number of samples is used as the number of training set samples and 30% is used as the number of validation set samples.

[0050] Furthermore, data augmentation is performed on the small sample size fault data subset to form an augmented small sample size fault data subset; this may include the following steps: S10331. Linearly scale and time-series shift the electrical data in the small sample size fault data subset to obtain an enhanced electrical data subset; For example, linear scaling of electrical data can be achieved by scaling the electrical data proportionally (e.g., ±10%), and the scaled electrical data must meet the rated parameters of the power grid equipment; time-series shifting of electrical data can be achieved by shifting the electrical data on the time axis (e.g., ±5 sampling points) to simulate the randomness of fault triggering time.

[0051] S10332. Perform data interpolation on the environmental data in the small sample size fault data subset to obtain an enhanced environmental data subset; For example, cubic spline interpolation is used for continuous data such as temperature and humidity to generate high-resolution time series, and the number of interpolation points is dynamically adjusted according to the sampling rate of the original data (e.g., reducing interpolation for high-frequency sampling data); Gaussian noise (σ=0.5% of the range) can also be added to simulate sensor error; and impulse noise can be injected into lightning data, rain and snow data, and pollution data.

[0052] For example, in a certain substation, temperature and humidity data are collected every 10 minutes, totaling 144 data points per day. The cubic spline interpolation method is used for the temperature and humidity data, inserting 5 interpolation points between every two original data points, thereby improving the time resolution of the data to one data point every 2 minutes.

[0053] For example, some locations are randomly selected in the lightning data sequence, and the data values ​​at these locations are set to a large random number (much larger than the normal lightning data range) to simulate the strong interference of lightning. Normal lightning data is between 0 and 10, while the data at the randomly selected locations is set to 100 or -50, etc.

[0054] S10333: The state data in the small sample size fault data subset is used to obtain an enhanced state data subset based on state combination and state time series changes; For example, state combination can be a combination of device switching state (on / off) and operating state (normal / abnormal) (e.g., "closed + abnormal"), using Markov chain to simulate state transition probability; state time sequence change can be a time warp of the state sequence (e.g., accelerating the fault evolution process), inserting random state jumps (e.g., suddenly changing from normal to alarm).

[0055] For example, a power device has two state types: switch state (on / off) and operating state (normal / abnormal). Combining these two states yields four different state combinations: "on + normal", "on + abnormal", "on + normal", and "on + abnormal". Using a Markov chain to simulate the state transition probability, for example, if the current state is "on + normal", according to the Markov chain model, there is a certain probability that it will transition to the "on + abnormal" state, and there is also a certain probability that it will remain in the "on + normal" state.

[0056] For example, the original state sequence of a device from normal operation to failure is "normal - normal - normal - abnormal - severe abnormal". By time-distorting the state sequence, the failure evolution process is accelerated, and the state sequence is changed to "normal - abnormal - severe abnormal", which shortens the time from the occurrence of failure to its severity.

[0057] S10334. Perform geometric transformations, color adjustments, and occlusions on the image data in the small sample size fault data subset to obtain an enhanced image data subset; For example, geometric changes in image data can be random rotation (±15), scaling (90%-110%), and translation (±10% pixels) to simulate camera angle changes, locally distorting the infrared thermal image to simulate uneven surface temperature of the device; color adjustment can be adjusted in brightness (±20%), contrast (±15%), and saturation (±10%) to simulate different lighting conditions; pseudo-color mapping transformation can be performed on the infrared image to enhance the visualization of fault features; occlusion simulation can be random occlusion of 10%-30% of the image area to simulate foreign object obstruction or sensor failure; raindrop and fog effects can be added (based on physically based rendering).

[0058] S10335. The video data in the small sample size fault data subset is spliced ​​and subjected to video frame operations to obtain an enhanced video data subset; wherein, the video frame operations include video frame insertion and video frame repetition. For example, video splicing can be done by splicing video clips of different fault scenarios along the timeline (e.g., overload followed by short circuit), using optical flow to ensure motion continuity at the splicing point; video frame operations can include frame insertion: inserting repeated frames at critical fault moments to emphasize fault characteristics; frame deletion: randomly deleting non-critical frames to simulate data loss; and frame rate variation: adjusting the playback speed (0.8x-1.2x) to simulate different sampling rates.

[0059] S10336. The enhanced electrical data subset, the enhanced environmental data subset, the enhanced status data subset, the enhanced image data subset, and the enhanced video data subset are combined to form the enhanced small sample size fault data subset.

[0060] For example, the enhanced electrical data subset, the enhanced environmental data subset, the enhanced state data subset, the enhanced image data subset, and the enhanced video data subset are aligned by their original timestamps.

[0061] In one embodiment of the present invention, the sample training set is input into the fault early warning model for training, such as... Figure 3 As shown, Figure 3 An optional flowchart illustrating a smart power grid operation and maintenance early warning method provided in this application embodiment. Figure 3 Specifically, it includes the following steps: S201. Input the training set into the fault warning model to obtain the first fault warning result; Before training begins, the parameters of the fault warning model are initialized and randomly assigned. The training set is divided into multiple batches, each containing a certain number of samples. The training set is then input into the fault warning model in batches to obtain the first fault warning result.

[0062] In one embodiment of the present invention, the fault warning model includes a multimodal feature extraction layer, a multimodal feature fusion layer, and a classification prediction layer; wherein the multimodal feature extraction layer includes a one-dimensional convolutional sub-layer, a two-dimensional convolutional sub-layer, and a three-dimensional convolutional sub-layer; the multimodal feature fusion layer includes a feature vector concatenation sub-layer, an attention weight sub-layer, and a weighted fusion sub-layer; and the classification prediction layer includes a fully connected sub-layer, an activation function sub-layer, and an output sub-layer.

[0063] In one embodiment of the present invention, a multimodal feature extraction layer is used to extract multimodal data features from the input training set, providing a foundation for subsequent fusion and classification; wherein the multimodal feature extraction layer includes a data preprocessing sublayer, a one-dimensional convolutional sublayer, a two-dimensional convolutional sublayer, and a three-dimensional convolutional sublayer; specifically including the following: The data preprocessing sublayer receives the training set and preprocesses the multimodal data in the training set to improve the training efficiency and performance of the model. It normalizes one-dimensional time-series data such as electrical and environmental data, using Min-Max normalization for electrical data, Z-Score normalization for environmental data, and one-hot encoding for state data. It also performs size adjustment and normalization on two-dimensional image data such as equipment appearance images and infrared thermal imaging images, uniformly adjusting the image data to the size required by the model using bilinear interpolation to normalize the image pixel values ​​to the [0,1] interval. Finally, it performs size adjustment and normalization on three-dimensional video data such as equipment operation videos and fault process videos, uniformly adjusting each frame of the video data to the size required by the model, for example, 224×224 pixels, using bilinear interpolation to normalize the pixel values ​​of each frame of the video data to the [0,1] interval.

[0064] One-dimensional convolutional sublayers are used to extract one-dimensional data features from electrical data, environmental data, and state data. Specifically, multi-scale convolutional kernels are designed to extract features from different types of data. For electrical data, the convolutional kernels include kernel 1, kernel 2, and kernel 3. Convolution kernel 1: Captures local details (e.g., current spikes, voltage drops). Kernel size = 3: Suitable for detecting local edge features Step size = 1: Preserve high temporal resolution without losing details. Convolution kernel 2: Extracts intermediate-scale patterns (periodic / harmonic / mode). Core size = 5: corresponds to a medium window Step size = 2: Moderate downsampling reduces computation while preserving periodic characteristics. Convolution kernel 3: Captures global trends (e.g., voltage changes). Core size = 7: corresponds to a longer window Step size = 4: Significant downsampling to highlight macro trends. Two-dimensional convolutional sublayers are used to extract features such as edges, textures, and shapes from preprocessed two-dimensional image data. Specifically, a two-dimensional convolutional kernel W∈R is used. p×q (p and q are the height and width of the convolution kernel, respectively) The convolution operation is performed by sliding on the image, and a two-dimensional feature tensor is extracted after convolution.

[0065] The 3D convolutional sublayer is used to simultaneously capture temporal and spatial features of the video from the preprocessed 3D video data, such as the motion trajectory of objects and action patterns, using a 3D convolutional kernel K∈R. r×p×q (r, p, and q are the sizes of the convolution kernel in the time, height, and width directions, respectively.) The convolution operation is performed by sliding the video to obtain a three-dimensional feature tensor.

[0066] In one embodiment of the present invention, a multimodal feature fusion layer is used to fuse features from different modalities obtained by the multimodal feature extraction layer, including a feature vector concatenation sublayer, an attention weight sublayer, and a weighted fusion sublayer; specifically, The feature vector concatenation sub-layer concatenates the feature vectors extracted from the one-dimensional, two-dimensional, and three-dimensional convolutional sub-layers in sequence to form a long feature vector for subsequent attention mechanism processing.

[0067] The attention weight sublayer assigns different weights to different features in the concatenated long feature vector, highlighting important features and suppressing unimportant features, enabling the model to focus more on information crucial for fault warning. The concatenated long feature vector is v. First, a fully connected layer maps v to a low-dimensional space, obtaining a hidden vector h. Then, the softmax function is used to calculate the attention weight for each feature. Assuming the dimension of the hidden vector h is d, the attention weight a for the i-th feature is... i The calculation is as follows: Where hi is the i-th element of the hidden vector h.

[0068] Output attention weight vector , where d is the dimension of the concatenated long feature vector.

[0069] The weighted fusion sublayer is used to perform a weighted summation of the attention weights and the concatenated long feature vector to obtain the fused feature vector. The concatenated long feature vector v and the attention weight vector a are multiplied by corresponding elements of the attention weight vector a and the long feature vector v and then summed to obtain the fused feature vector u, whose dimension is the same as that of the concatenated long feature vector.

[0070] The classification prediction layer is used to transform multimodal fusion features into interpretable fault type prediction results, including a fully connected sublayer, an activation function sublayer, and an output sublayer; specifically, The fully connected sublayer is used to further integrate and abstract the fused features, mapping high-dimensional features to a low-dimensional space related to the fault type. It receives the fused feature vector output from the weighted fusion sublayer, assuming its dimension is d, and learns the weight matrix W∈R. m×d (m is the output dimension) The input feature vector v∈Rd is transformed into a new feature representation h∈Rm, that is, h= W·v + b, where b is the bias term, and the output feature vector h is m in dimension, where m is usually less than or equal to d, providing a more compact feature representation for classification prediction.

[0071] The activation function sublayer is used to introduce nonlinearity into the model, enabling the model to learn complex, nonlinear fault feature relationships. It receives the feature vector h∈Rm output by the fully connected sublayer and introduces a nonlinear function, such as the ReLU function, to calculate ReLU(x)=max(0,x) for each element in h, mapping the input data to a nonlinear space. The output feature vector h'∈Rm after the activation function is processed remains unchanged in dimension, but the feature values ​​are mapped to a new nonlinear space.

[0072] The output sublayer maps the feature vector to the probability distribution of specific fault types, outputting the predicted probability of each fault type. It receives the feature vector h'∈Rm output by the activation function sublayer, and uses the Softmax function to convert the input feature vector into a probability distribution. The Softmax function converts each element hi' in h' into a probability value pi, and outputs a probability vector p with dimension N (N is the number of fault types). Each element represents the predicted probability of the corresponding fault type, and the sum of all probability values ​​is 1, which can be used to determine the fault type.

[0073] S202. Calculate the loss value using the cross-entropy loss function based on the first fault warning result, and calculate the gradient value of the fault warning model using the backpropagation algorithm based on the loss value; For example, to calculate the loss value: collect the first fault warning result (the probability distribution of the fault type predicted by the model) and the true fault labels of the training set samples. For example, the model may output a predicted probability vector of [0.1, 0.7, 0.2], while the true label is [0, 1, 0] (representing the second fault type); use the cross-entropy loss function, which measures the degree of difference between the model's prediction and the actual situation, to calculate the loss value based on the predicted probability and the true label. The smaller the loss value, the closer the prediction result is to the actual situation; starting from the output layer, calculate the gradient layer by layer forward. Taking the fully connected layer as an example, calculate the gradient of the loss function with respect to the weight matrix and the bias vector.

[0074] S203. The parameters of the fault warning model are updated using the stochastic gradient descent algorithm and based on the gradient values. For example, set an appropriate learning rate, such as 0.001. Based on the calculated gradient value and the learning rate, use the stochastic gradient descent algorithm to update the model parameters. For each parameter of the model, the update method is to subtract the learning rate multiplied by the gradient value from the current parameter value to obtain the updated parameter value. For example, if the current value of a certain weight is 0.5 and the gradient is 0.1, then the updated weight value is 0.5 - 0.001 × 0.1 = 0.499.

[0075] S204. Input the validation set into the updated fault warning model to obtain the second fault warning result. Calculate the evaluation result based on the accuracy and recall. Adjust the parameters of the updated fault warning model according to the evaluation result to obtain the final fault warning model.

[0076] For example, the validation set samples are input into the updated fault warning model to obtain the second fault warning result, which is the probability distribution of the fault type for each sample. The number of fault samples correctly predicted in the validation set is counted and divided by the total number of samples in the validation set to obtain the accuracy. For example, if there are 100 samples in the validation set and the model correctly predicts 85, the accuracy is 85%. For each fault type, its recall is calculated. Taking a certain fault type as an example, the recall = (number of fault samples correctly predicted as that type) / (total number of actual fault samples of that type in the validation set). For example, if there are 20 samples of a certain fault type in the validation set and the model correctly predicts 16, the recall of that fault type is 80%.

[0077] Analysis and evaluation results: If the accuracy is low, it indicates that the model has a lot of false positives. It may be necessary to adjust the decision threshold of the classification prediction layer or optimize the parameters of the multimodal feature fusion layer to enhance the model's ability to distinguish different fault features. If the recall is low, it indicates that the model has a lot of false negatives for certain fault types. It may be necessary to re-examine the parameters of the feature extraction layer and strengthen the ability to extract features for these fault types.

[0078] Adjust hyperparameters: Adjust hyperparameters such as learning rate and batch size based on the evaluation results; for example, if the model's performance on the validation set fluctuates greatly, try reducing the learning rate to make the model training more stable.

[0079] Repeated training and evaluation: The model with adjusted parameters is retrained, and the validation and evaluation process is repeated until the model achieves satisfactory performance indicators such as accuracy and recall on the validation set. Finally, this model is determined as the fault warning model.

[0080] S104. Collect the first and second datasets of the power grid in real time and input them into the trained fault early warning model for fault early warning.

[0081] Real-time data collection includes voltage, current, power, frequency, harmonics, temperature, humidity, wind direction, wind speed, lightning, rain / snow, pollution, equipment switch status, equipment operating status, equipment alarm status, equipment appearance images, infrared thermal imaging images, on-site environmental images, equipment operation videos, fault process videos, and maintenance operation videos. The collected data is then input into a trained fault early warning model to provide fault warnings.

[0082] Based on the power grid intelligent operation and maintenance early warning method of the above embodiments, this application also provides a power grid intelligent operation and maintenance early warning system, such as... Figure 4 As shown, Figure 4 This is a schematic diagram of the structure of a smart grid operation and maintenance early warning system provided in an embodiment of this application. The smart grid operation and maintenance early warning system 4 includes, wherein, Based on the power grid intelligent operation and maintenance early warning method of the above embodiments, this application also provides a power grid intelligent operation and maintenance early warning device, such as... Figure 5 As shown, Figure 5 This is a schematic diagram of the structure of a smart grid operation and maintenance early warning device provided in an embodiment of this application. The smart grid operation and maintenance early warning device 5 includes a processor 501 and a memory 502. The memory 502 is used to store computer programs; the processor 501 is used to call and run the computer programs from the memory to execute the smart grid operation and maintenance early warning method as described in the above embodiment.

[0083] In the embodiments of this application, the processor 501 described above can be at least one of the following: Application-Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), Controller, Microcontroller, and Microprocessor. It is understood that for different devices, the electronic device used to implement the above processor function can also be other types, and the embodiments of this application do not specifically limit it.

[0084] This application provides a computer-readable storage medium storing a computer program for implementing the power grid intelligent operation and maintenance early warning method as described in any of the above embodiments when executed by a processor.

[0085] For example, the program instructions corresponding to a smart grid operation and maintenance early warning method in this embodiment can be stored on storage media such as optical discs, hard disks, and USB flash drives. When the program instructions corresponding to a smart grid operation and maintenance early warning method in the storage media are read or executed by an electronic device, the smart grid operation and maintenance early warning method as described in any of the above embodiments can be implemented.

[0086] Furthermore, in the embodiments of this application, the functional modules can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional module.

[0087] If the integrated unit is implemented as a software functional module and is not sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this embodiment, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the method of this embodiment. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0088] It should be understood that the phrases "one embodiment," "an embodiment," or "some embodiments" mentioned throughout the specification mean that a specific feature, structure, or characteristic related to an embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment," "in one embodiment," or "in some embodiments" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. The descriptions of the various embodiments above tend to emphasize the differences between the various embodiments; their similarities or commonalities can be referred to mutually, and for the sake of brevity, they will not be repeated here.

[0089] The modules described above as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0090] In addition, each functional module in the various embodiments of this application can be integrated into one processing unit, or each module can be a separate unit, or two or more modules can be integrated into one unit; the integrated modules can be implemented in hardware or in the form of hardware plus software functional units.

[0091] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0092] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0093] The features disclosed in the several product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0094] The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0095] The above description is merely an embodiment of this application, but the protection scope of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for intelligent operation and maintenance early warning of power grids, characterized in that, The method includes: Acquire a first historical dataset and a second historical dataset of the power grid, wherein the first dataset includes electrical data, environmental data, and status data, and the second dataset includes image data and video data; A fault dataset is formed by labeling fault data in the first historical dataset and the second historical dataset based on fault events and fault scenarios. The fault dataset is augmented to construct a sample training set, and the sample training set is input into the fault early warning model for training to form a trained fault early warning model. The first and second datasets of the power grid are collected in real time and input into the trained fault early warning model for fault early warning.

2. The power grid intelligent operation and maintenance early warning method according to claim 1, characterized in that, The electrical data includes voltage data, current data, power data, frequency data, and harmonic data; the environmental data includes temperature data, humidity data, wind direction data, wind speed data, lightning data, rain and snow data, and pollution data; the status data includes equipment switch status, equipment operating status, and equipment alarm status; the image data includes equipment appearance images, infrared thermal imaging images, and on-site environmental images; and the video data includes equipment operation videos, fault process videos, and maintenance operation videos.

3. The power grid intelligent operation and maintenance early warning method according to claim 1, characterized in that, A fault dataset is formed by labeling fault data in the first historical dataset and the second historical dataset based on fault events and fault scenarios, including: Based on the recorded time points of fault events and fault scenarios, a time window is determined according to preset time expansion rules; Multiple fault data are obtained by matching the first historical dataset and the second historical dataset according to the time window; Multiple fault data are labeled with fault content stamps and fault timestamps according to a preset labeling format to form a fault dataset; wherein the fault content stamps include fault type, fault location, fault device identifier, and fault severity.

4. The power grid intelligent operation and maintenance early warning method according to claim 3, characterized in that, The fault dataset is augmented to construct a sample training set, including: The fault dataset is divided into multiple fault data subsets based on the fault type. Based on the data volume, the multiple fault data subsets are divided into small sample size fault data subsets, medium sample size fault data subsets, and large sample size fault data subsets; The small sample size fault data subset is augmented to form an augmented small sample size fault data subset. The enhanced small-sample-size fault data subset, the medium-sample-size fault data subset, and the large-sample-size fault data subset are merged to form a sample training set.

5. The power grid intelligent operation and maintenance early warning method according to claim 4, characterized in that, Augmenting the small sample size fault data subset to form an augmented small sample size fault data subset includes: The electrical data in the small sample size fault data subset is linearly scaled and time-series shifted to obtain an enhanced electrical data subset; The environmental data in the small sample size fault data subset is interpolated to obtain an enhanced environmental data subset. An enhanced state data subset is obtained by combining state data and state time-series changes from the state data subset with the small sample size of the fault data subset. The image data in the small sample size fault data subset is geometrically transformed, color adjusted, and occluded to obtain an enhanced image data subset; An enhanced video data subset is obtained by performing video stitching and video frame operations on the video data in the small sample size fault data subset; wherein, the video frame operations include video frame insertion and video frame repetition; The enhanced electrical data subset, the enhanced environmental data subset, the enhanced status data subset, the enhanced image data subset, and the enhanced video data subset are combined to form the enhanced small sample size fault data subset.

6. The power grid intelligent operation and maintenance early warning method according to claim 5, characterized in that, Before constructing the sample training set, the following are also included: The enhanced small-sample-size fault data subset and the large-sample-size fault data subset are balanced based on the fault type; wherein the balancing process includes oversampling the enhanced small-sample-size fault data subset and undersampling the large-sample-size fault data subset. The medium-sized fault data subset, the balanced enhanced small-sized fault data subset, and the large-sized fault data subset are combined to form a sample training set. The sample training set is divided into a training set and a validation set according to a preset ratio.

7. The power grid intelligent operation and maintenance early warning method according to claim 6, characterized in that, The sample training set is then input into the fault early warning model for training, including: The training set is input into the fault warning model to obtain the first fault warning result; The loss value is calculated using the cross-entropy loss function based on the first fault warning result, and the gradient value of the fault warning model is calculated using the backpropagation algorithm based on the loss value. The parameters of the fault warning model are updated using the stochastic gradient descent algorithm based on the gradient values. The validation set is input into the updated fault warning model to obtain the second fault warning result. The evaluation result is calculated based on the accuracy and recall. The parameters of the updated fault warning model are adjusted according to the evaluation result to obtain the final fault warning model.

8. The power grid intelligent operation and maintenance early warning method according to claim 7, characterized in that, The fault warning model includes a multimodal feature extraction layer, a multimodal feature fusion layer, and a classification prediction layer; wherein the multimodal feature extraction layer includes a one-dimensional convolutional sub-layer, a two-dimensional convolutional sub-layer, and a three-dimensional convolutional sub-layer; the multimodal feature fusion layer includes a feature vector concatenation sub-layer, an attention weight sub-layer, and a weighted fusion sub-layer; and the classification prediction layer includes a fully connected sub-layer, an activation function sub-layer, and an output sub-layer.

9. A smart power grid operation and maintenance early warning system, characterized in that, include: The acquisition unit is used to acquire a historical first dataset and a historical second dataset of the power grid, wherein the first dataset includes electrical data, environmental data and status data, and the second dataset includes image data and video data; An annotation unit is used to annotate fault data in the first historical dataset and the second historical dataset based on fault events and fault scenarios to form a fault dataset. The training unit is used to perform data augmentation on the fault dataset to construct a sample training set, and input the sample training set into the fault early warning model for training to form a trained fault early warning model. The early warning unit is used to collect the first and second datasets of the power grid in real time and input them into the trained fault early warning model for fault early warning.

10. A smart power grid operation and maintenance early warning device, characterized in that, include: Processor and memory, of which, The memory is used to store computer programs; The processor is configured to call and run the computer program from the memory to perform the method as described in any one of claims 1 to 7.

Images