A smart power distribution network fault early warning and remote diagnosis method and system

Abstract

The application is suitable for the field of power system automation technology, and relates to a kind of intelligent distribution network fault early warning and remote diagnosis method and system, comprising: collecting SCADA real-time data, PMU synchronous phasor data, meteorological environment data and equipment state data;Screening key features and generating model input feature set;The feature set is input into the edge end lightweight early warning model for real-time reasoning, generates an early warning signal, and uploads the early warning signal and the corresponding feature set to the cloud;The cloud identifies the fault type and the confidence through the fault type classification submodel, constructs the distribution network topology, and outputs the fault positioning coordinates;Generate a diagnosis report containing fault cause, influence range and operation and maintenance suggestion, and push to the operation and maintenance terminal;Incremental sample feature knowledge is integrated into the original model using knowledge distillation technology, and incremental training and parameter updating are carried out.The application solves the problems of low fault processing efficiency and fault positioning accuracy, slow response and high cost of traditional distribution network.

Description

Technical Field

[0001] This invention belongs to the field of power system automation technology, and in particular relates to a method and system for intelligent distribution network fault early warning and remote diagnosis. Background Technology

[0002] As the core carrier of the coordinated interaction of "source, grid, load, and storage" in the new power system, the smart distribution network is a key link connecting centralized power generation, distributed new energy sources (such as photovoltaics and energy storage), and diversified user-side loads (data centers, charging piles, industrial loads, etc.). Its operational stability directly determines the power supply reliability, energy utilization efficiency, and user electricity experience of the power system. Fault early warning and remote diagnosis technologies are the core supporting means to ensure the safe operation of the smart distribution network, undertaking the important functions of early risk identification, rapid fault location, and assisted efficient handling. With the changes in the structure and operating characteristics of the distribution network, the data generated by the distribution network exhibits characteristics of massive volume, real-time nature, and multi-source heterogeneity—the daily data volume reaches the TB level, the sampling frequency needs to meet the kHz level to capture instantaneous changes in electrical quantities, and the data types cover electrical quantity data such as voltage and current, environmental data such as temperature and humidity, and status data such as equipment insulation resistance and temperature rise. This places higher demands on the real-time performance, accuracy, and adaptability of fault early warning and diagnosis technologies.

[0003] Traditional methods have the following drawbacks:

[0004] The reliance on single-type monitoring data (such as only collecting electrical quantity data like voltage and current) without integrating related data such as meteorological environment and equipment status leads to significant bias in fault feature extraction. For example, short-circuit faults caused by lightning strikes during thunderstorms are easily misjudged as line overloads because data such as the number of lightning strikes and humidity are not included; faults caused by transformer insulation aging cannot be detected in advance due to the lack of equipment status data such as insulation resistance and temperature rise, ultimately resulting in a high rate of false alarms and missed alarms.

[0005] The existing technologies suffer from poor timeliness and a lack of predictive capabilities. Most adopt a "post-fault diagnosis" model—requiring overcurrent protection devices, circuit breakers, and other equipment to trigger their actions before inferring the fault type and location based on fault data, thus lacking an early warning mechanism.

[0006] The problem of low fault location accuracy. Traditional fault location methods mainly use impedance method and traveling wave method. The former is affected by the many branches of the distribution network and the asymmetrical load, and the location error often exceeds 500 meters; the latter is affected by line attenuation and reflected wave interference, and has poor applicability in low-voltage distribution areas (10kV and below).

[0007] Existing machine learning diagnostic models mostly adopt a "static training-fixed deployment" model. When a new photovoltaic inverter is added to the distribution network, the line topology is adjusted (such as adding a new transformer branch), or a new type of fault occurs, the model's diagnostic accuracy decreases because the training data is not updated in a timely manner. Moreover, model updates require manual re-labeling of massive amounts of samples and full training, resulting in long maintenance cycles and high costs for operation and maintenance manpower and computing power, making it impossible to adapt to dynamic changes in the distribution network.

[0008] The problem of poor remote interactivity. Traditional diagnostic systems can only output basic information such as "fault location + preliminary type" and cannot generate targeted operation and maintenance suggestions; moreover, they lack a real-time interaction mechanism with on-site operation and maintenance personnel - operation and maintenance personnel cannot provide feedback on the progress of fault handling and the actual situation on site, resulting in a break in the closed loop of early warning-diagnosis-handling-feedback.

[0009] Therefore, it is imperative to address the problems of low efficiency in fault handling and fault location accuracy, slow response, and high cost in traditional power distribution networks. Summary of the Invention

[0010] To address the shortcomings of existing technologies, the purpose of this invention is to provide a smart distribution network fault early warning and remote diagnosis method to solve the problems of low efficiency and fault location accuracy, slow response and high cost in traditional distribution network fault handling; in addition, this invention also provides a smart distribution network fault early warning and remote diagnosis system.

[0011] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0012] This invention provides a method for fault early warning and remote diagnosis of smart distribution networks, comprising the following steps:

[0013] S10. Deploy edge acquisition terminals in the distribution network area to collect SCADA real-time data, PMU synchronous phasor data, meteorological environmental data and equipment status data;

[0014] S20. The data collected in S10 is cleaned and standardized, and key features are selected and a model input feature set is generated by combining wavelet transform and Fourier transform with an attention mechanism.

[0015] S30. Input the feature set into the edge lightweight early warning model for real-time inference, calculate the probability of fault occurrence, and when the probability of fault occurrence is greater than or equal to the preset early warning threshold, generate an early warning signal and upload the early warning signal and the corresponding feature set to the cloud.

[0016] S40. After receiving the uploaded data, the cloud inputs it into the cloud-based accurate diagnosis model. The fault type classification sub-model identifies the fault type and confidence level. At the same time, the distribution network topology map is constructed, and the topological relationship and feature set are input into the fault location sub-model based on graph attention network, and the fault location coordinates are output.

[0017] S50, combining fault type, location coordinates and historical operation and maintenance data, generates a diagnostic report including fault cause, impact range and operation and maintenance suggestions, and pushes it to the operation and maintenance terminal;

[0018] S60. When a new type of fault or topology adjustment occurs in the distribution network, the new data is labeled as incremental samples and stored in the sample pool. The feature knowledge of the incremental samples is integrated into the original model using knowledge distillation technology for incremental training and parameter updates.

[0019] Furthermore, in S10, the data acquisition terminal includes a DTU power distribution terminal and a smart meter; the SCADA real-time data includes three-phase voltage, three-phase current, active power and reactive power; the PMU synchronization phasor data includes phase angle and frequency; the meteorological environment data includes temperature, humidity and number of lightning strikes; and the equipment status data includes insulation resistance value and temperature rise data.

[0020] Furthermore, the specific steps of S20 are as follows:

[0021] S201. Use the 3σ criterion to remove outliers from voltage and current time series data, and use linear interpolation to fill in missing data.

[0022] S202. Z-score standardization is applied to data with different dimensions.

[0023] S203. The current signal is decomposed by wavelet transform to extract the time-domain features at the moment of fault change. The voltage signal is converted to the frequency domain by Fourier transform to extract the harmonic components. The contribution weight of each feature to the fault is calculated by combining the attention mechanism, the top 20 key features are selected, and the model input feature set is generated.

[0024] Furthermore, in step S30, an improved random forest algorithm is used to train the feature set:

[0025] The optimal feature subset is selected using the Gini coefficient, reducing the feature dimensionality by 30% and lowering the computational overhead at the edges. The formula is as follows:

[0026]

[0027] in, The number of categories includes faulty and normal. Features Corresponding category Given the sample proportion, let the model output failure probability be... ,when When this occurs, an early warning signal is triggered at the edge and pushed to the cloud.

[0028] Furthermore, in step S40, the GBDT gradient boosting tree algorithm is used to construct a fault type classifier using the feature set F pushed from the edge and historical fault samples as training data. The classifier outputs the fault type. and confidence level Conf, when At that time, it was determined to be a valid diagnostic result.

[0029] Furthermore, the model loss function uses logarithmic loss:

[0030]

[0031] in, For the sample size, For the sample The true labels are: 1 = faulty, 0 = normal. Predict samples for the model This represents the probability of failure.

[0032] Furthermore, construct the distribution network topology. ,in A set of nodes, including towers and transformers. Given a set of lines, including line impedance and length, a graph attention network algorithm is used to integrate topological relationships with the feature set. The fusion process, through node feature matching, path probability calculation, and maximum probability node location, outputs the coordinates of the fault point and the path probability. The calculation is as follows:

[0033]

[0034] in, For nodes The set of neighboring nodes, ) is a node and The edge weights between lines are determined by both line impedance and characteristic similarity.

[0035] Furthermore, in S60, only 10%-20% of the model parameters are updated to avoid full retraining. The training time is controlled within 1 hour. The model's early warning accuracy, false alarm rate, and false negative rate are calculated daily. When any indicator deviates from the target value for 3 consecutive days, incremental updates are automatically triggered. Manually triggered updates are also supported.

[0036] Secondly, the present invention also provides a smart distribution network fault early warning and remote diagnosis system, comprising:

[0037] The perception layer is equipped with several sensors to collect real-time SCADA data, PMU synchronization phasor data, meteorological environmental data, and equipment status data of the power distribution network.

[0038] The transmission layer, including edge nodes and communication modules, is used to aggregate and transmit data to the platform layer. The communication modules include a power line carrier PLC module, a 5G communication module, and an optical fiber module.

[0039] The platform layer includes edge computing terminals and cloud data centers. The edge computing terminals deploy an improved random forest model for real-time early warning, and the cloud data centers include storage servers and computing servers that run a CNN-LSTM hybrid model and a GAT fault location model to provide computing power support.

[0040] The edge computing terminal establishes a collaborative channel for data uploading and model distribution with the cloud data center.

[0041] Furthermore, it also includes an encryption chip and a diagnostic report generation unit. The encryption chip is used to encrypt the data storage and processing, and the diagnostic report generation unit is used to automatically generate a diagnostic report containing the cause of the fault, the scope of impact, and maintenance suggestions by combining the fault type, location results, and historical operation and maintenance data, so as to assist in operation and maintenance decision-making.

[0042] Compared with existing technologies, the intelligent distribution network fault early warning and remote diagnosis method and system provided by this invention have at least the following advantages:

[0043] Traditional power distribution networks suffer from low fault handling efficiency, low fault location accuracy, slow response, and high costs. This invention addresses these issues by integrating SCADA, PMU, meteorological, and equipment status data through a three-dimensional feature fusion system encompassing temporal, spatial, and environmental aspects. By combining wavelet transform, Fourier transform, and attention mechanisms, the accuracy of key feature extraction is significantly improved, and the false alarm and false negative rates for fault warnings are drastically reduced. This effectively avoids misdiagnosing overload as short circuits and missing equipment aging faults. Furthermore, the improved random forest model at the edge of this invention controls the warning response time to within 100ms, enabling real-time monitoring at the distribution area level. The cloud-based CNN-LSTM hybrid model, combined with GAT topology mapping, can predict complex faults in advance, breaking away from the traditional post-fault diagnosis model and effectively shortening the warning lag time, thus securing a "golden window" for fault handling. The fault location error is reduced, and the coverage of fault type diagnosis is improved, accurately identifying new types of faults such as harmonic exceedances caused by distributed photovoltaic grid connection and voltage fluctuations caused by energy storage charging and discharging switching. This invention, based on an incremental learning mechanism using knowledge distillation, shortens model update and maintenance cycles. The closed-loop interactive process of "early warning-diagnosis-handling-feedback," combined with automatically generated operation and maintenance suggestions, effectively reduces the mean time to repair (MTTR) of the distribution network and lowers the power supply reliability index (SAIDI). Simultaneously, it is compatible with existing SCADA and electricity consumption information collection systems, requiring no large-scale infrastructure modifications and reducing deployment costs. This invention employs the SM4 national cryptographic algorithm for encryption throughout the entire chain, ensuring high security for data transmission and storage, and the model architecture supports flexible expansion. In summary, this invention, through systematic technological innovation, achieves "precision, real-time, intelligent, and closed-loop" early warning and diagnosis of distribution network faults, providing core technical support for the safe and stable operation of smart distribution networks, and possesses significant engineering application value and economic benefits. Attached Figure Description

[0044] To more clearly illustrate the solutions of the present invention, a brief introduction will be given to the drawings used in the description of the embodiments below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0045] Figure 1 A flowchart illustrating a method for early warning and remote diagnosis of faults in a smart distribution network, as provided in an embodiment of the present invention;

[0046] Figure 2 This is a deployment diagram of an application scenario for an intelligent power distribution network fault early warning and remote diagnosis system provided in an embodiment of the present invention. Detailed Implementation

[0047] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0048] 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 invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0049] This invention provides a method for fault early warning and remote diagnosis of intelligent distribution networks, focusing on the sub-field of operation monitoring and fault management of intelligent distribution networks. It can be further subdivided into intelligent diagnostic technology that integrates machine learning and power systems. Its application scenarios cover low-voltage distribution networks of 10kV and below (including urban distribution networks and rural distributed distribution networks) and complex distribution networks with distributed photovoltaic / energy storage access. It can serve the development of fault management systems for power grid operation and maintenance companies and power automation equipment manufacturers, while being compatible with existing distribution network SCADA, electricity consumption information collection systems, and other infrastructure, possessing cross-scenario adaptability. The method for fault early warning and remote diagnosis of intelligent distribution networks includes the following steps:

[0050] S10. Deploy edge acquisition terminals in the distribution network area to collect SCADA real-time data, PMU synchronization phasor data, meteorological environmental data, and equipment status data; S20. Clean and standardize the data collected in S10, and use wavelet transform and Fourier transform, combined with an attention mechanism, to select key features and generate a model input feature set; S30. Input the feature set into the lightweight early warning model at the edge for real-time inference, calculate the probability of fault occurrence, and generate an early warning signal when the probability of fault occurrence is greater than or equal to the preset early warning threshold, and upload the early warning signal and the corresponding feature set to the cloud; S40. The cloud receives the uploaded data. Then, the data is input into a cloud-based precision diagnostic model. The fault type classification sub-model identifies the fault type and confidence level. Simultaneously, a distribution network topology map is constructed, and the topological relationships and feature sets are input into a fault location sub-model based on graph attention networks, outputting fault location coordinates. S50: Combining the fault type, location coordinates, and historical operation and maintenance data, a diagnostic report containing the fault cause, impact range, and operation and maintenance suggestions is generated and pushed to the operation and maintenance terminal. S60: When a new type of fault or topology adjustment occurs in the distribution network, the new data is labeled as incremental samples and stored in the sample pool. Knowledge distillation technology is used to integrate the feature knowledge of the incremental samples into the original model for incremental training and parameter updates.

[0051] This invention solves the problems of low efficiency and accuracy of fault location, slow response and high cost in traditional power distribution network fault handling.

[0052] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0053] This invention provides a method for fault early warning and remote diagnosis of intelligent distribution networks, focusing on the sub-field of operation monitoring and fault management of intelligent distribution networks. It can be further subdivided into intelligent diagnostic technology that integrates machine learning and power systems. Its application scenarios cover low-voltage distribution networks of 10kV and below (including urban distribution networks and rural distributed distribution networks) and complex distribution networks with distributed photovoltaic / energy storage access. It can serve the fault management system development of power grid operation and maintenance companies and power automation equipment manufacturers, while being compatible with existing distribution network SCADA, electricity consumption information collection systems, and other infrastructure, possessing cross-scenario adaptability. Figure 1 As shown, in this embodiment, the smart distribution network fault early warning and remote diagnosis method includes the following steps:

[0054] S10. Deploy edge acquisition terminals in the distribution network area to collect SCADA real-time data, PMU synchronous phasor data, meteorological environmental data and equipment status data.

[0055] Specifically, in this embodiment, the data acquisition terminal includes a DTU distribution terminal and a smart meter. The SCADA real-time data includes a sampling frequency of 1kHz, three-phase voltage, three-phase current, active power, and reactive power. The PMU synchronous phasor data includes a sampling frequency of 50Hz, phase angle, and frequency. Meteorological sensors are deployed around the transformer area to collect temperature (accuracy ±0.5℃), humidity (accuracy ±3%RH), and lightning strike count (resolution 1 strike / minute). Status monitoring sensors are installed on key equipment (such as transformers and circuit breakers) to collect insulation resistance (range 0-1000MΩ) and temperature rise (range -40℃-125℃) data.

[0056] Furthermore, in this embodiment, the edge acquisition terminal transmits data to the regional edge nodes via power line carrier (PLC) or 5G private network. After being aggregated by the edge nodes, the data is transmitted to the cloud data center via optical fiber. The transmission latency is controlled within 50ms, and encryption protocols (such as the SM4 national cryptographic algorithm) are used to ensure data security.

[0057] S20. The data collected in S10 is cleaned and standardized. Then, key features are selected and a model input feature set is generated by using wavelet transform and Fourier transform, combined with an attention mechanism.

[0058] Specifically, in this embodiment, the "3σ criterion" is used to remove outliers (such as sudden values ​​caused by sensor failure) from time-series data such as voltage and current. Linear interpolation is used to fill in missing data (when the missing rate is ≤5%), and a sensor failure warning is triggered when the missing rate exceeds 5%. Data of different dimensions (such as voltage kV and temperature ℃) are processed by Z-score standardization (converting the data into a distribution with a mean of 0 and a standard deviation of 1) to avoid the impact of dimension differences on model training. The current signal is decomposed by wavelet transform (using the db4 wavelet basis) to extract the time-domain features (such as peak factor and kurtosis) at the moment of fault change. The voltage signal is converted to the frequency domain by Fourier transform to extract harmonic components (such as the 3rd and 5th harmonic content). The contribution weight of each feature to the fault is calculated by combining the attention layer mechanism, and the top 20 key features are selected to generate the model input feature set.

[0059] S30. Input the feature set into the lightweight early warning model at the edge for real-time inference, calculate the probability of fault occurrence, and generate an early warning signal when the probability of fault occurrence is greater than or equal to the preset early warning threshold. Then, upload the early warning signal and the corresponding feature set to the cloud.

[0060] Specifically, in this embodiment, an improved random forest algorithm is used to train the feature set, and the optimal feature subset is selected by the Gini coefficient to reduce the feature dimension by 30% and reduce the computational overhead at the edge (the smaller the Gini coefficient, the stronger the ability of the feature to distinguish between fault and normal states).

[0061]

[0062] in, Number of categories (faulty / normal). Features Corresponding category The sample proportion. Model training and early warning triggering: Set the number of decision trees to 100 (to balance model accuracy and overfitting risk), and the training objective is to identify fault precursor features (such as voltage fluctuations exceeding ±5%, current harmonic content exceeding 5%). Model output fault probability. ,when When the warning signal is triggered at the edge, it is pushed to the cloud; the warning response time is controlled within 100ms to meet the real-time monitoring requirements of the distribution area.

[0063] S40. After receiving the uploaded data, the cloud inputs it into the cloud-based accurate diagnostic model. The fault type classification sub-model identifies the fault type and confidence level. At the same time, a distribution network topology map is constructed, and the topological relationships and feature sets are input into the fault location sub-model based on graph attention network, outputting the fault location coordinates.

[0064] Specifically, in this embodiment, the GBDT gradient boosting tree algorithm is used to push the feature set from the edge. Historical fault samples (including 10 common fault types: single-phase grounding, three-phase short circuit, line overload, transformer fault, etc.) are used as training data to construct a fault type classifier. The classifier outputs the fault type. and confidence level Conf, when When the result is deemed a valid diagnosis, the model uses logarithmic loss as its loss function to ensure classification accuracy.

[0065]

[0066] in, For the sample size, For the sample The true label (1 = fault, 0 = normal). Predict samples for the model This represents the probability of failure.

[0067] Furthermore, in this embodiment, a distribution network topology diagram is constructed. ( It is a set of nodes, including towers, transformers, etc. For a set of lines (containing attributes such as line impedance and length), the Graph Attention Network (GAT) algorithm is used to connect the topological relationships with the feature set. Fusion. The fault point coordinates are output through node feature matching, path probability calculation, and maximum probability node localization, with a localization error ≤ 50 meters. Path probability. The calculation is as follows:

[0068]

[0069] in, For nodes The set of neighboring nodes. ) is a node and The edge weights between lines are determined by both line impedance and characteristic similarity.

[0070] S50, combining fault type, location coordinates and historical operation and maintenance data, generates a diagnostic report including fault cause, impact range and operation and maintenance suggestions, and pushes it to the operation and maintenance terminal.

[0071] Specifically, in this embodiment, the cause of the fault is such as "single-phase grounding due to line aging," and the affected scope involves the number of users and load capacity. The maintenance suggestion is such as "the 10kV line section needs to be replaced, and the spare part model is XX." The edge-end early warning signal is pushed to the maintenance personnel via SMS and APP. The cloud diagnostic report is synchronized to the distribution network dispatch center in real time. The maintenance personnel receive the diagnostic report through the mobile APP and can provide feedback on information such as "fault confirmation / troubleshooting" and "handling progress." The feedback data is transmitted back to the cloud in real time. The cloud system automatically records the entire process data of "early warning time - diagnostic result - handling measures - repair time" to form a fault handling file for subsequent model training.

[0072] S60. When a new type of fault or topology adjustment occurs in the distribution network, the new data is labeled as incremental samples and stored in the sample pool. Knowledge distillation technology is used to integrate the feature knowledge of the incremental samples into the original model for incremental training and parameter updates.

[0073] Specifically, in this embodiment, when a new type of fault (such as a fault not included in historical samples) or topology adjustment occurs in the distribution network, the system automatically labels the new data as "incremental samples" and stores them in the sample pool; using knowledge distillation technology, the feature knowledge of the incremental samples is integrated into the original model (the teacher model is the original model, and the student model is the updated model), updating only 10%-20% of the model parameters to avoid full retraining, and the training time is controlled within 1 hour; the model's early warning accuracy (target ≥ 95%), false alarm rate (target ≤ 5%), and missed alarm rate (target ≤ 3%) are calculated daily, and incremental updates are automatically triggered when any indicator deviates from the target value for 3 consecutive days; manual triggering of updates is also supported (such as after major topology adjustments in the distribution network).

[0074] This invention also provides a smart distribution network fault early warning and remote diagnosis system, combined with Figure 1 and Figure 2 In this embodiment, the intelligent distribution network fault early warning and remote diagnosis system includes:

[0075] The sensing layer (bottom layer) is equipped with several sensors to collect real-time SCADA data, PMU synchronization phasor data, meteorological environmental data, and equipment status data of the power distribution network.

[0076] The transport layer (middle layer) includes edge nodes and communication modules, which are used to aggregate data and transmit it to the platform layer. The communication modules include power line carrier PLC modules, 5G communication modules and fiber optic modules.

[0077] The platform layer (top layer) includes edge computing terminals and cloud data centers. The edge computing terminals deploy an improved random forest model for real-time early warning. The cloud data centers include storage servers and computing servers, running a CNN-LSTM hybrid model and a GAT fault location model to provide computing power support.

[0078] Among them, a collaborative channel for data uploading and model distribution is established between the edge computing terminal and the cloud data center.

[0079] Specifically, in this embodiment, the sensing layer includes a SCADA sensor, a PMU device, a meteorological sensor, and an equipment status sensor. The SCADA sensor is used to collect three-phase voltage, current, and active / reactive power at a sampling frequency of 1kHz to capture real-time changes in electrical quantities. The PMU device is used to collect synchronization phasors (phase angle and frequency) at a sampling frequency of 50Hz to ensure data time synchronization. The meteorological sensor is used to collect temperature (accuracy ±0.5℃), humidity (accuracy ±3%RH), wind speed, and lightning strike count to provide data on environmental influencing factors. The equipment status sensor is used to collect the temperature and insulation resistance of key equipment such as transformers and circuit breakers to detect early signs of equipment aging and failure.

[0080] Specifically, in this embodiment, in the transmission layer, the PLC module transmits data through power line carrier technology, adapting to existing power distribution network line resources and reducing deployment costs; the 5G communication module relies on the low latency characteristics of the 5G private network, with a transmission latency of ≤50ms, meeting the real-time early warning data transmission requirements; the edge node (including the data aggregation unit) aggregates multi-source transmission data, completes data format conversion and preliminary cleaning, and reduces invalid data transmission; the fiber optic transmission module undertakes high-speed data transmission between the edge node and the cloud data center, ensuring stable uploading of massive amounts of data.

[0081] Specifically, in this embodiment, in the platform layer, the edge computing terminal (including the lightweight model running unit) deploys an improved random forest model to achieve real-time early warning of ≤100ms and reduce reliance on cloud computing power; the encryption chip (SM4) is integrated into various hardware components of the platform layer to encrypt data storage and computation processes, ensuring the security of core data; the diagnostic report generation unit automatically generates a diagnostic report containing the cause of the fault, the scope of impact, and operational suggestions by combining the fault type, location results, and historical operation and maintenance data to assist in operation and maintenance decisions.

[0082] Compared with existing technologies, the intelligent distribution network fault early warning and remote diagnosis method and system described in the above embodiments suffer from problems such as low fault handling efficiency and fault location accuracy, slow response, and high cost in traditional distribution networks. This invention integrates SCADA, PMU, meteorological, and equipment status data through a three-dimensional feature fusion system of "time-space-environment," combining wavelet transform, Fourier transform, and attention mechanisms. This significantly improves the accuracy of key feature extraction and greatly reduces the false alarm and missed alarm rates for fault early warning, effectively avoiding problems such as misjudging overload as short circuit and missing equipment aging faults. The improved random forest model at the edge of this invention controls the early warning response time to within 100ms, achieving real-time monitoring at the distribution area level. The cloud-based CNN-LSTM hybrid model combined with GAT topology mapping can predict complex faults in advance, breaking the traditional post-fault diagnosis mode and effectively shortening the early warning lag time, thus securing a "golden window" for fault handling. The fault location error is reduced, and the fault type diagnosis coverage is improved, accurately identifying new faults such as harmonic exceedances caused by distributed photovoltaic grid connection and voltage fluctuations caused by energy storage charging and discharging switching. This invention, based on an incremental learning mechanism using knowledge distillation, shortens model update and maintenance cycles. The closed-loop interactive process of "early warning-diagnosis-handling-feedback," combined with automatically generated operation and maintenance suggestions, effectively reduces the mean time to repair (MTTR) of the distribution network and lowers the power supply reliability index (SAIDI). Simultaneously, it is compatible with existing SCADA and electricity consumption information collection systems, requiring no large-scale infrastructure modifications and reducing deployment costs. This invention employs the SM4 national cryptographic algorithm for encryption throughout the entire chain, ensuring high security for data transmission and storage, and the model architecture supports flexible expansion. In summary, this invention, through systematic technological innovation, achieves "precision, real-time, intelligent, and closed-loop" early warning and diagnosis of distribution network faults, providing core technical support for the safe and stable operation of smart distribution networks, and possesses significant engineering application value and economic benefits.

[0083] Obviously, the embodiments described above are merely preferred embodiments of the present invention, and not all embodiments. The accompanying drawings illustrate preferred embodiments of the present invention, but do not limit the scope of the patent. The present invention can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this invention.

Claims

1. A method for early warning and remote diagnosis of faults in a smart distribution network, characterized in that, Includes the following steps: S10. Deploy edge acquisition terminals in the distribution network area to collect SCADA real-time data, PMU synchronous phasor data, meteorological environmental data and equipment status data; S20. The data collected in S10 is cleaned and standardized, and key features are selected and a model input feature set is generated by combining wavelet transform and Fourier transform with an attention mechanism. S30. Input the feature set into the edge lightweight early warning model for real-time inference, calculate the probability of fault occurrence, and when the probability of fault occurrence is greater than or equal to the preset early warning threshold, generate an early warning signal and upload the early warning signal and the corresponding feature set to the cloud. S40. After receiving the uploaded data, the cloud inputs it into the cloud-based accurate diagnosis model. The fault type classification sub-model identifies the fault type and confidence level. At the same time, the distribution network topology map is constructed, and the topological relationship and feature set are input into the fault location sub-model based on graph attention network, and the fault location coordinates are output. S50, combining fault type, location coordinates and historical operation and maintenance data, generates a diagnostic report including fault cause, impact range and operation and maintenance suggestions, and pushes it to the operation and maintenance terminal; S60. When a new type of fault or topology adjustment occurs in the distribution network, the new data is labeled as incremental samples and stored in the sample pool. The feature knowledge of the incremental samples is integrated into the original model using knowledge distillation technology for incremental training and parameter updates.

2. The method for intelligent distribution network fault early warning and remote diagnosis according to claim 1, characterized in that, In S10, the data acquisition terminal includes a DTU power distribution terminal and a smart meter. The SCADA real-time data includes three-phase voltage, three-phase current, active power and reactive power. The PMU synchronization phasor data includes phase angle and frequency. The meteorological environment data includes temperature, humidity and number of lightning strikes. The equipment status data includes insulation resistance value and temperature rise data.

3. The method for intelligent distribution network fault early warning and remote diagnosis according to claim 1, characterized in that, The specific steps of S20 are as follows: S201. Use the 3σ criterion to remove outliers from voltage and current time series data, and use linear interpolation to fill in missing data. S202. Z-score standardization is applied to data with different dimensions. S203. The current signal is decomposed by wavelet transform to extract the time-domain features at the moment of fault change. The voltage signal is converted to the frequency domain by Fourier transform to extract the harmonic components. The contribution weight of each feature to the fault is calculated by combining the attention mechanism, the top 20 key features are selected, and the model input feature set is generated.

4. The method for intelligent distribution network fault early warning and remote diagnosis according to claim 1, characterized in that, In step S30, an improved random forest algorithm is used to train the feature set: The optimal feature subset is selected using the Gini coefficient, reducing the feature dimensionality by 30% and lowering the computational overhead at the edges. The formula is as follows: in, The number of categories includes faulty and normal. Features Corresponding category Given the sample proportion, let the model output failure probability be... ,when When this occurs, an early warning signal is triggered at the edge and pushed to the cloud.

5. The method for intelligent distribution network fault early warning and remote diagnosis according to claim 1, characterized in that, In step S40, the GBDT gradient boosting tree algorithm is used to construct a fault type classifier using the feature set F pushed from the edge and historical fault samples as training data. The classifier outputs the fault type. and confidence level Conf, when When the result is obtained, it is considered a valid diagnosis.

6. The method for intelligent distribution network fault early warning and remote diagnosis according to claim 5, characterized in that, The model loss function uses logarithmic loss: in, For the sample size, For the sample The true labels are: 1 = faulty, 0 = normal. Predict samples for the model This represents the probability of failure.

7. The method for intelligent distribution network fault early warning and remote diagnosis according to claim 5, characterized in that, Constructing the power distribution network topology ,in A set of nodes, including towers and transformers. Given a set of lines, including line impedance and length, a graph attention network algorithm is used to integrate topological relationships with the feature set. The fusion process, through node feature matching, path probability calculation, and maximum probability node location, outputs the coordinates of the fault point and the path probability. The calculation is as follows: in, For nodes The set of neighboring nodes, ) is a node and The edge weights between lines are determined by both line impedance and characteristic similarity.

8. The method for intelligent distribution network fault early warning and remote diagnosis according to claim 1, characterized in that, In the S60, only 10%-20% of the model parameters are updated to avoid full retraining. The training time is controlled within 1 hour. The model's early warning accuracy, false alarm rate, and false negative rate are calculated daily. When any indicator deviates from the target value for 3 consecutive days, incremental updates are automatically triggered. Manually triggered updates are also supported.

9. A system employing the method as described in any one of claims 1 to 8, characterized in that, include: The perception layer is equipped with several sensors to collect real-time SCADA data, PMU synchronization phasor data, meteorological environmental data, and equipment status data of the power distribution network. The transmission layer, including edge nodes and communication modules, is used to aggregate and transmit data to the platform layer. The communication modules include a power line carrier PLC module, a 5G communication module, and an optical fiber module. The platform layer includes edge computing terminals and cloud data centers. The edge computing terminals deploy an improved random forest model for real-time early warning, and the cloud data centers include storage servers and computing servers that run a CNN-LSTM hybrid model and a GAT fault location model to provide computing power support. The edge computing terminal establishes a collaborative channel for data uploading and model distribution with the cloud data center.

10. The system according to claim 9, characterized in that, It also includes an encryption chip and a diagnostic report generation unit. The encryption chip is used to encrypt the data storage and operation process, and the diagnostic report generation unit is used to automatically generate a diagnostic report containing the cause of the fault, the scope of impact, and operation and maintenance suggestions by combining the fault type, the location result, and historical operation and maintenance data to assist operation and maintenance decision-making.

Images