A power distribution network state perception and fault diagnosis method based on embedded edge computing

Abstract

The application provides a power distribution network state sensing and fault diagnosis method based on embedded edge computing, which comprises deploying an embedded edge computing node on a power distribution network terminal device, collecting multi-element operation data of the power distribution network in real time by the edge computing node through high-precision sensors, and performing local preprocessing on the collected multi-element operation data; deploying a lightweight neural network model on the edge computing node, and realizing real-time sensing of the operation state of the power distribution network and AI / ML fault diagnosis through real-time inference; and interacting the edge computing node with a central station through an edge-cloud collaborative mechanism, wherein the edge-cloud collaborative mechanism comprises a data selective uploading strategy, a federated learning type model optimization, and a central station deep diagnosis supplement, so as to optimize the diagnosis result and the model parameter. The application can realize high-precision real-time sensing of the operation state of the power distribution network and rapid and accurate diagnosis of faults.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, specifically to a distribution network status perception and fault diagnosis method based on embedded edge computing. By introducing embedded edge computing technology, the method enables real-time perception of the distribution network's operating status and rapid and accurate fault diagnosis, thereby improving the intelligence level and power supply reliability of the distribution network. Background Technology

[0002] As the end-user segment of the power system, the distribution network's operational status perception and fault diagnosis are crucial for ensuring power supply reliability and rapid recovery. With the integration of distributed energy resources and a high proportion of renewable energy into the grid, the distribution network structure is becoming increasingly complex, and the amount of operational data is exploding. Traditional diagnostic methods relying on central station cloud processing face numerous challenges. 1. High data transmission latency: When massive amounts of sensor data are uploaded to the cloud for processing, the latency can reach hundreds of milliseconds when the network is congested, which cannot meet the requirements for millisecond-level fault response.

[0003] 2. Bandwidth and privacy issues: Uploading full data consumes a lot of bandwidth and poses a risk of data leakage.

[0004] 3. Centralized computing resources: The cloud is overloaded and cannot cope with sudden peak failures.

[0005] While existing edge computing technologies have been applied in the Industrial Internet of Things (IIoT), there is still a lack of systematic local computing frameworks, lightweight AI model deployment, and edge-cloud collaboration mechanisms for state perception and fault diagnosis in power distribution networks. In recent years, the development of lightweight machine learning models (such as MobileNet and TinyML) and embedded AI chips has made real-time inference possible at the edge. Summary of the Invention

[0006] This invention provides a distribution network status perception and fault diagnosis method based on embedded edge computing. Through localized data acquisition, real-time preprocessing, lightweight artificial intelligence / machine learning (AI / ML) model inference, and intelligent collaboration mechanism with the central station, it achieves high-precision real-time perception of the distribution network operation status and rapid and accurate fault diagnosis.

[0007] The present invention achieves the above objectives through the following technical solutions: A method for power distribution network status perception and fault diagnosis based on embedded edge computing, applied in power distribution network automation devices, includes: Embedded edge computing nodes are deployed on the distribution network terminal equipment. The edge computing nodes collect multi-dimensional operation data of the distribution network in real time through high-precision sensors and perform local preprocessing on the collected multi-dimensional operation data. Deploy lightweight neural network models on edge computing nodes to achieve real-time perception of the power distribution network's operating status and AI / ML fault diagnosis through real-time inference; The edge computing nodes interact with the central station through the edge-cloud collaboration mechanism. The edge-cloud collaboration mechanism includes a selective data upload strategy, federated learning model optimization, and in-depth diagnostic supplementation by the central station to optimize diagnostic results and model parameters. Among them, the edge computing nodes only upload complex / uncertain fault samples and their feature vectors, new unknown fault events, and periodically summarized statistical features. The central station performs global model retraining or incremental learning and distributes the updated model parameters to the edge computing nodes.

[0008] According to the present invention, a distribution network status perception and fault diagnosis method based on embedded edge computing is provided. An embedded edge computing node is integrated in the distribution network terminal equipment. The edge computing node includes a processor, a memory, a sensor interface and a communication module. The processor is selected as a domestic dedicated edge AI acceleration chip that supports floating-point operations and matrix acceleration or a high-performance processor compatible with ARM Cortex-M / A / R series and RISC-V architecture. A variety of heterogeneous sensors, including at least high-precision voltage transformers, current transformers, miniature phasor measurement units, temperature sensors, vibration sensors, and partial discharge sensors, are connected through a high-precision sensor interface to collect multi-dimensional operating data of the line in real time. Configure sensor sampling parameters. For critical fault transient signals, set the sampling rate to no less than 50kHz to capture high-frequency transient characteristics. Real-time synchronous acquisition and caching of sensor data is achieved on edge computing nodes. The timing consistency of multi-source data is ensured through hardware timers or software triggering mechanisms, and the acquired raw data is temporarily stored in local memory.

[0009] According to the present invention, a method for power distribution network status perception and fault diagnosis based on embedded edge computing is provided, which performs local preprocessing on the collected multi-dimensional operational data at the edge computing node, including the following steps: A bandpass filter is used to filter the collected multivariate operating data to remove irrelevant frequency bands other than the 50Hz power frequency and retain the effective frequency bands containing key fault characteristics. Denoising of the filtered data can be achieved by applying wavelet threshold denoising or adaptive Kalman filtering algorithms. The denoised data is synchronized and aligned, and the consistency of multi-source data in time is ensured by timestamp matching or data interpolation methods. The data after synchronization and alignment is normalized to map it to a uniform range and remove outliers.

[0010] The present invention provides a method for power distribution network status perception and fault diagnosis based on embedded edge computing, which deploys a lightweight neural network model on edge computing nodes to achieve real-time monitoring and evaluation of the normal operation of the power distribution network, including the following steps: Deploy a lightweight load forecasting model: Based on historical data and real-time current / power sequences, use a lightweight LSTM or Transformer-mini model optimized with 8-bit / 4-bit integer quantization, structured pruning, and knowledge distillation techniques to perform short-term load forecasting. Deploy a voltage quality assessment module: calculate key voltage quality indicators in real time, including voltage deviation, three-phase unbalance, harmonic distortion rate, and flicker, determine whether each voltage quality indicator exceeds the limit by preset thresholds, and issue an early warning signal when the limit is exceeded. Deployment of equipment health assessment model: Integrating multi-dimensional signals and using unsupervised anomaly detection algorithm to assess the health status (SOH) of key equipment.

[0011] According to the present invention, a method for power distribution network status perception and fault diagnosis based on embedded edge computing is provided, which deploys a lightweight neural network model on the edge computing node to realize AI / ML fault diagnosis of the normal operation of the power distribution network, including the following steps: A lightweight CNN-LSTM hybrid fault diagnosis model is deployed. The fault diagnosis model includes depthwise separable convolutional layers and lightweight LSTM layers. The depthwise separable convolutional layers are used to extract local time-frequency features of the preprocessed signal, and the lightweight LSTM layers are used to capture the temporal dependencies of transient processes. A fast criterion system based on mutation start-up and zero-order direction criterion rules is deployed on the edge computing nodes. The fault diagnosis model is only activated for further diagnosis when a fault feature that meets the start-up conditions is detected. During the fault diagnosis process, the fault diagnosis model outputs a confidence score for each diagnosis result; when the confidence score is lower than a preset threshold, the sample is marked as an uncertain sample and relevant feature information is recorded. A traveling wave-based double-ended or single-ended ranging algorithm is used, combined with a fault diagnosis model to correct the ranging results.

[0012] The distribution network state perception and fault diagnosis method based on embedded edge computing provided by the present invention further performs a comprehensive judgment and output step, including: Edge computing nodes integrate local AI diagnostic results, rule-based judgment results, and feedback information from the central station. The integration strategy employs weighted voting or Bayesian integration algorithms, assigning weights based on the reliability and importance of each information source to generate the final status report and fault alarm. The output includes the distribution network operating status level, divided into four levels: normal, attention, abnormal, and severe; fault type, occurrence time, confidence level, scope of impact, and recommended isolation scheme; and also provides the health score of key equipment and specific maintenance recommendations. The final generated status report and fault alarm are quickly reported to the main station system via standard communication protocols; Based on the final diagnostic results, the edge computing node automatically triggers the corresponding local automated actions.

[0013] According to the present invention, a method for power distribution network state perception and fault diagnosis based on embedded edge computing is provided, the construction of a lightweight load prediction model includes: Long Short-Term Memory Networks or lightweight Transformer-mini models are selected as the base models; For the LSTM model, its gating mechanism is improved by reducing the number of neurons in the hidden layer and using a projection layer to map the high-dimensional hidden state to the low-dimensional output space. For the Transformer-mini model, a sparse self-attention mechanism is used instead of a fully connected self-attention mechanism, and linear positional encoding or learnable positional embedding is used to limit the length of the input sequence to adapt to the short-time prediction window. The model's weights, activations, and gradients are quantized using 8-bit / 4-bit integer quantization techniques. Quantization-aware training is introduced to simulate quantization error during training, and gradients are backpropagated through a pass-through estimator. For critical layers, a mixed precision strategy is adopted, that is, some high-precision calculations are retained to ensure numerical stability, while low-bit integer operations are used for other layers. Implement importance-based structured pruning, pruning at the neuron or convolutional kernel level; use L1 norm or L2 norm to evaluate the importance of each channel or neuron, and remove redundant structures whose contribution to the prediction results is lower than a preset threshold. Historical load data, real-time current sequences, active / reactive power sequences, and time characteristics are fused using multi-modal methods and used as model inputs. An external feature embedding layer is introduced to map meteorological data or electricity price signals into a vector space, and then weighted and fused into time series features through an attention mechanism.

[0014] According to the present invention, a distribution network state perception and fault diagnosis method based on embedded edge computing includes a voltage quality assessment module comprising the following units: Real-time performance indicator calculation unit: This unit contains multiple parallel calculation submodules, each responsible for the real-time calculation of different voltage quality indicators. Voltage deviation calculation submodule: Real-time acquisition of instantaneous three-phase voltage values, calculation of the percentage deviation from the rated voltage using the formula Δ. U =( U real U nom ) / U nom ×100%; Three-phase unbalance calculation submodule: Based on the symmetrical component method, it calculates the ratio of negative sequence component to positive sequence component in real time and generates three-phase unbalance index. The harmonic distortion rate calculation submodule uses Fast Fourier Transform or Wavelet Transform algorithms to perform frequency domain analysis on the real-time voltage signal, extract the 2nd to 50th harmonic components, and calculate the total harmonic distortion rate. ; Voltage flicker calculation submodule: Based on IEC 61000-4-15 standard, it calculates the short-time flicker value Pst and the long-time flicker value Plt through square detection and filtering. Limit Exceeding Criteria and Threshold Management Unit: Built-in threshold database set by national voltage quality standards and enterprise operation and maintenance specifications; A dynamic threshold mechanism is adopted to automatically adjust the sensitivity of the criterion according to different runtime stages; The system compares the various indicators calculated in real time with the preset thresholds and outputs a Boolean flag indicating that the limit has been exceeded. Warning signal generation and classification unit: Based on the over-limit flag and the degree to which the indicator exceeds the limit, different levels of early warning signals are generated; A time-sliding window smoothing algorithm is used to eliminate false alarms caused by instantaneous interference, ensuring that alarms are triggered only when limits are continuously exceeded.

[0015] According to the present invention, a power distribution network state perception and fault diagnosis method based on embedded edge computing is provided. The construction of a lightweight CNN-LSTM hybrid fault diagnosis model includes the following steps: By replacing standard convolutional layers with depthwise separable convolutional layers, the convolutions of spatial and channel dimensions are decoupled, significantly reducing the computational cost when extracting local features in the temporal and frequency domains. We designed a parallel structure of multi-scale convolutional kernels and used 3×3, 5×5 and 7×7 convolutional kernels to capture fault features under different receptive fields, focusing on extracting fault traveling wave head steepness, high-frequency energy distribution and transient spectrum features. A batch normalization layer and the GELU activation function are introduced after the convolutional layer; Flatten the high-dimensional feature map output by the CNN module into a sequence vector and input it into a lightweight LSTM layer; An improved LSTM gating structure is used, employing a projection layer to compress the hidden state dimension from high-dimensional to low-dimensional. Introducing residual connections across LSTM layers alleviates the gradient vanishing problem in deep networks; The model weights and activation values ​​are symmetrically quantized using 8-bit / 4-bit quantization, and quantization-aware training QAT is used to simulate quantization noise during the training phase. The importance of convolutional kernel channels in CNN modules is scored based on the L1 norm, and redundant channels with low contribution to fault classification are removed. A channel attention module, SE-Block, is inserted between CNN and LSTM to automatically weight the importance of features in different frequency domains. The output layer uses a Softmax classifier to output the confidence probability distribution for each fault type.

[0016] The present invention provides a method for power distribution network state perception and fault diagnosis based on embedded edge computing. The specific implementation of a fast criterion system based on abrupt initiation and zero-sequence direction criterion includes: Configure a high-speed sampling interface to acquire in real time the instantaneous values ​​of three-phase current, three-phase voltage, zero-sequence current, and zero-sequence voltage collected by the edge computing node; The real-time data is subjected to sliding window mean filtering to calculate the amplitude and phase at the current moment and align them with the corresponding values ​​at the previous moment or the previous period. A phase-separate discrimination strategy is adopted to calculate the rate of change of amplitude of current and voltage in each phase in real time. d i / d t , d u / d t (or the absolute value of the difference between sampled values); Set a fixed start threshold and a floating start threshold. When the current change or voltage change of any phase is continuous... N When a sampling point exceeds a preset threshold value, a high-level signal is output to trigger the sudden change. An interlocking logic is introduced. When normal load fluctuation characteristics are detected, the automatic interlocking start signal is determined by harmonic characteristics or duration. For low-current grounding or arc suppression coil grounding systems, zero-sequence current and zero-sequence voltage signals are acquired in real time. Calculate the direction of zero-sequence power or the phase difference angle between zero-sequence current and zero-sequence voltage. ; Setting direction criteria: When the zero-sequence power flows from the line to the bus, it is determined to be a forward fault; otherwise, it is determined to be a reverse fault or a bus fault. The direction enable signal is output only when a positive fault is detected; The output of the fast criterion system serves as the enable signal for the fault diagnosis model, directly controlling the opening and closing of the model's power domain or clock domain. The threshold value for triggering sudden changes is dynamically adjusted based on the current operating load rate and background noise level of the distribution network.

[0017] Therefore, compared with existing technologies, the power distribution network status perception and fault diagnosis method based on embedded edge computing proposed in this invention can achieve efficient task scheduling, real-time load balancing, and energy efficiency optimization in power distribution network automation devices. Through the technical solution of this invention, power distribution network automation devices can achieve the following beneficial effects: 1. Traditional cloud-based diagnostic methods require uploading massive amounts of sensor data to the cloud for processing, resulting in data transmission delays of up to hundreds of milliseconds during network congestion, which cannot meet the millisecond-level fault response requirements of distribution networks. This invention, however, introduces embedded edge computing capabilities into distribution network terminal equipment, enabling local data preprocessing and lightweight AI / ML model inference. By offloading critical computing tasks to the edge, it significantly reduces data transmission latency, allowing fault diagnosis and status awareness to be completed within milliseconds, greatly improving the system's real-time performance.

[0018] 2. Uploading all data to the cloud not only consumes a large amount of bandwidth resources but also increases network transmission costs. This invention, through local preprocessing and feature extraction at edge nodes, uploads only key feature vectors or complex fault samples to the central station, significantly reducing the amount of data uploaded, effectively saving bandwidth resources, and optimizing the utilization efficiency of network resources.

[0019] 3. In cloud-based processing, sensitive data needs to be uploaded to remote servers, posing risks of data leakage and privacy violations. This invention utilizes edge computing to complete data preprocessing and preliminary analysis locally, uploading only non-sensitive or anonymized feature data, effectively reducing the risk of data leakage and enhancing user privacy protection. Simultaneously, the optimized federated learning model in the edge-cloud collaboration mechanism further ensures data security during transmission and storage.

[0020] 4. This invention leverages the local computing capabilities of edge nodes to offload some computational tasks from the central station to the edge, significantly reducing the computational burden on the central station. This distributed computing architecture not only enhances the overall processing power of the system but also strengthens its scalability, enabling the system to easily cope with the challenges of future expansion and increased complexity of the power distribution network.

[0021] 5. The lightweight AI / ML model used in this invention has undergone deep optimization, enabling high-precision state perception and fault diagnosis at the edge. By integrating multiple sensor data and intelligent algorithms, the system can accurately identify various fault types, including single-phase grounding, phase-to-phase short circuits, and open-circuit faults, and achieve rapid location and isolation. Simultaneously, the edge-cloud collaboration mechanism, through in-depth diagnostic supplementation from the central station and continuous model optimization, further improves the accuracy of fault diagnosis and the system's self-healing capability.

[0022] 6. This invention fully utilizes the hardware resources of existing power distribution network terminal equipment, and achieves low-cost, high-efficiency intelligent transformation by introducing embedded edge computing modules and lightweight AI models. Compared to the traditional cloud-based diagnostic model, which requires large-scale upgrades to network infrastructure and central station computing resources, this invention significantly reduces deployment costs and improves economic efficiency. Simultaneously, the system's high real-time performance and accuracy reduce power outage time and economic losses, further enhancing social benefits.

[0023] 7. This invention is particularly suitable for power distribution network scenarios with limited hardware resources, complex communication conditions, and stringent real-time requirements. Through lightweight model design and edge-cloud collaboration mechanism, the system can operate stably under low power consumption and low bandwidth conditions and effectively cope with various complex faults and interference situations.

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0025] Figure 1 This is a flowchart of an embodiment of a power distribution network status perception and fault diagnosis method based on embedded edge computing according to the present invention.

[0026] Figure 2 This is an overall flowchart of state perception and fault diagnosis in an embodiment of the distribution network state perception and fault diagnosis method based on embedded edge computing of the present invention.

[0027] Figure 3 This is a flowchart illustrating the inference process of deploying a lightweight neural network model on an edge computing node in an embodiment of a power distribution network state perception and fault diagnosis method based on embedded edge computing according to the present invention.

[0028] Figure 4 This is a schematic diagram of the structure of a lightweight CNN-LSTM hybrid fault diagnosis model in an embodiment of a power distribution network state perception and fault diagnosis method based on embedded edge computing according to the present invention.

[0029] Figure 5 This is a schematic diagram of the system framework of the edge-cloud collaboration mechanism in an embodiment of a power distribution network state perception and fault diagnosis method based on embedded edge computing of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0031] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0032] See Figures 1 to 5 This embodiment provides a method for power distribution network status perception and fault diagnosis based on embedded edge computing, including the following steps: Embedded edge computing nodes are deployed on the distribution network terminal equipment. The edge computing nodes collect multi-dimensional operation data of the distribution network in real time through high-precision sensors and perform local preprocessing on the collected multi-dimensional operation data. Deploy lightweight neural network models on edge computing nodes to achieve real-time perception of the power distribution network's operating status and AI / ML fault diagnosis through real-time inference; The edge computing nodes interact with the central station through the edge-cloud collaboration mechanism. The edge-cloud collaboration mechanism includes a selective data upload strategy, federated learning model optimization, and in-depth diagnostic supplementation by the central station to optimize diagnostic results and model parameters. Among them, the edge computing nodes only upload complex / uncertain fault samples and their feature vectors, new unknown fault events, and periodically summarized statistical features. The central station performs global model retraining or incremental learning and distributes the updated model parameters to the edge computing nodes.

[0033] In this embodiment, a comprehensive judgment and output step is also performed, including: Edge computing nodes integrate local AI diagnostic results, rule-based criterion results, and feedback information from the central station. The integration strategy employs weighted voting or Bayesian fusion algorithms, assigning weights based on the reliability and importance of each information source to generate the final status report and fault alarm. The output includes the distribution network operation status level, divided into four levels: normal, attention, abnormal, and severe; fault type, occurrence time, confidence level, scope of impact, and recommended isolation scheme; it also provides health scores of key equipment and specific maintenance recommendations, providing comprehensive decision support for operation and maintenance personnel. The final generated status reports and fault alarms are quickly reported to the main station system via standard communication protocols such as IEC 61850 GOOSE, MMS, or DL / T 860 to ensure the real-time and accuracy of the information. Triggering local automated actions: Based on the final diagnostic results, the edge computing node automatically triggers corresponding local automated actions, including but not limited to sending trip signals to isolate the fault area, starting fault waveform storage to record fault details, and lighting up indicator lights to alert on-site personnel. Achieving closed-loop control: Through the above steps, this method realizes a complete closed-loop control process from "perception-diagnosis-decision-execution", ensuring the timeliness, effectiveness and intelligence level of power distribution network fault handling.

[0034] Embedded edge computing nodes are integrated into distribution network terminal equipment (such as feeder terminal unit, distribution terminal unit, or station terminal unit). The edge computing node includes a processor, memory, sensor interface, and communication module. The processor is selected from domestically produced dedicated edge AI acceleration chips that support floating-point operations and matrix acceleration or high-performance processors compatible with ARM Cortex-M / A / R series and RISC-V architecture. The high-precision sensor interface connects to various heterogeneous sensors, including but not limited to high-precision voltage transformers, current transformers, miniature phasor measurement units, temperature sensors, vibration sensors, and partial discharge sensors, to collect multi-dimensional operating data such as three-phase voltage, current, phasor, frequency, harmonics, temperature, vibration, and partial discharge of the line in real time. Configure sensor sampling parameters to ensure that the sampling frequency of normal operating data is not less than 10kHz, and for critical fault transient signals, set the sampling rate to not less than 50kHz to capture high-frequency transient characteristics; Real-time synchronous acquisition and caching of sensor data is achieved on edge computing nodes. The timing consistency of multi-source data is ensured through hardware timers or software triggering mechanisms, and the acquired raw data is temporarily stored in local memory for subsequent preprocessing and analysis.

[0035] Local preprocessing of the collected multi-dimensional operational data at the edge computing node includes the following steps: A bandpass filter is used to filter the collected multivariate operating data, removing irrelevant frequency bands other than the 50Hz power frequency and retaining the effective frequency bands containing key fault characteristics. Applying wavelet threshold denoising or adaptive Kalman filtering algorithms to denoise the filtered data effectively suppresses white noise and impulse interference, and improves the signal-to-noise ratio of the data. The denoised data is synchronized and aligned, and the consistency of multi-source data in time is ensured by timestamp matching or data interpolation methods. The data after synchronization and alignment are normalized to map the data to a uniform range and outliers are removed to improve the quality of subsequent model input data. As can be seen, the above preprocessing process is all completed locally on the edge computing node, avoiding the bandwidth pressure caused by uploading the original massive data to the central station, while providing a clean and high signal-to-noise ratio data foundation for subsequent feature extraction and model inference.

[0036] Deploying lightweight neural network models on edge computing nodes to achieve real-time monitoring and evaluation of the normal operation of the distribution network includes the following steps: Deploy a lightweight load prediction model: Based on historical data and real-time current / power sequences, use a lightweight LSTM or Transformer-mini model optimized with 8-bit / 4-bit integer quantization, structured pruning and knowledge distillation techniques to perform short-term load prediction, such as 5 to 30 minutes. Deploy a voltage quality assessment module: calculate key voltage quality indicators such as voltage deviation, three-phase unbalance, harmonic distortion rate, and flicker in real time, determine whether each voltage quality indicator exceeds the limit by preset thresholds, and issue an early warning signal when the limit is exceeded; this module is integrated into the edge computing node to ensure real-time performance and low latency. Deployment of equipment health assessment model: Integrating multi-dimensional signals such as temperature, vibration, and partial discharge, and using unsupervised anomaly detection algorithms such as isolated forest or support vector machine, to assess the health status (SOH) of key equipment such as switchgear, transformers, and line joints; The model is also optimized through quantization, pruning, and knowledge distillation, and the number of parameters and inference latency meet the requirements of low-power embedded platforms, enabling predictive maintenance; All models run in parallel or serial on edge computing nodes. Through efficient resource scheduling and management strategies, the smooth operation of each model on the low-power embedded platform is ensured, while meeting the needs of real-time monitoring and evaluation of the normal operation of the power distribution network.

[0037] Deploying lightweight neural network models on edge computing nodes to achieve AI / ML fault diagnosis for routine operation of distribution networks includes the following steps: A lightweight CNN-LSTM hybrid fault diagnosis model is deployed. This model comprises depthwise separable convolutional layers and lightweight LSTM layers. The depthwise separable convolutional layers extract local time-frequency features of the preprocessed signal, including but not limited to the steepness of the fault traveling wave head and energy distribution. The lightweight LSTM layers capture the temporal dependencies of transient processes. The fault diagnosis model is optimized with parameters controlled within the range of 500k to 2M to ensure efficient operation on edge computing nodes. The model can identify and classify various common and complex faults, including but not limited to single-phase grounding faults, phase-to-phase short circuits (two-phase and three-phase), open-circuit faults, overloads, ferroresonance, and high-resistance grounding. A fast criterion system based on mutation start-up and zero-order direction criterion rules is deployed on the edge computing node. The fault diagnosis model is only awakened for further diagnosis when a fault feature that meets the start-up condition is detected, thereby reducing average power consumption and improving diagnostic efficiency. During the fault diagnosis process, the fault diagnosis model outputs a confidence score for each diagnosis result; when the confidence score is lower than a preset threshold (e.g., 0.85), the sample is marked as an uncertain sample and relevant feature information is recorded to provide data support for subsequent edge-cloud collaborative optimization. A traveling wave-based dual-end or single-end ranging algorithm is adopted, and the ranging results are corrected by a fault diagnosis model to strictly control the fault location error within 50m; the entire positioning process is completed independently on the edge computing node to ensure rapid response; Through the above optimized design, the entire fault diagnosis process is completed independently on the edge side, with a typical fault diagnosis delay of less than 10ms, which fully meets the millisecond-level protection and control requirements of the distribution network.

[0038] In this embodiment, the implementation steps of the edge-cloud collaboration mechanism are as follows: According to preset upload rules, edge computing nodes only upload three types of data to the central station, including: complex or uncertain fault samples and their feature vectors, which are fault events with model confidence below a preset threshold; novel unknown fault events, i.e. fault types that appear for the first time or have not been identified by the model; and periodically summarized statistical features, such as equipment load rate and voltage quality indicators, rather than raw waveform data. This strategy significantly reduces the amount of data uploaded. The central station aggregates feature and gradient information uploaded from multiple edge computing nodes, and uses this information for global model retraining or incremental learning to optimize overall model performance. Subsequently, the central station distributes the updated lightweight model parameters to each edge computing node in a differential form to achieve continuous model evolution. During this process, there is no need to upload original sensitive data, effectively protecting user privacy and security. For complex fault events marked by edge computing nodes, the central station uses full historical data and high-precision large models for secondary diagnosis to improve diagnostic accuracy. The central station feeds back the final diagnosis to the corresponding edge computing node to correct local decisions and improve fault handling efficiency. During edge-cloud collaboration, lightweight communication protocols such as MQTT or CoAP are used, supporting multiple high-speed transmission methods such as 5G, fiber optics, and Ethernet to ensure low-latency and reliable data transmission between edge computing nodes and the central station, meeting the real-time requirements of the power distribution network.

[0039] In this embodiment, the construction of the lightweight load prediction model includes: Long Short-Term Memory (LSTM) network or a lightweight Transformer-mini model was selected as the base model. For the LSTM model, its gating mechanism is improved, the number of neurons in the hidden layer is reduced, and a projection layer is used to map the high-dimensional hidden state to the low-dimensional output space. For the Transformer-mini model, a sparse self-attention mechanism is used instead of a fully connected self-attention mechanism to reduce computational complexity, and linear positional encoding or learnable positional embedding is used to limit the length of the input sequence to adapt to the short-term prediction window. The model's weights, activations, and gradients are quantized using 8-bit / 4-bit integer quantization techniques. Quantization-aware training is introduced to simulate quantization error during training, and gradients are backpropagated through a straight-through estimator (STE). For critical layers (such as input and embedding layers), a mixed precision strategy is adopted, that is, some high-precision calculations are retained to ensure numerical stability, while low-bit integer operations are used for the remaining layers. Implement importance-based structured pruning, pruning at the neuron or convolutional kernel level; The importance of each channel or neuron is evaluated using the L1 norm or L2 norm, and redundant structures whose contribution to the prediction results is lower than a preset threshold are eliminated. After pruning, the interlayer connection matrix is ​​readjusted to ensure the regularity of the model structure and avoid the hardware acceleration difficulties caused by unstructured sparsity. Historical load data, real-time current sequences, active / reactive power sequences, and time characteristics (such as time, day of the week, and holiday markers) are fused using a multimodal approach and used as model input. An external feature embedding layer is introduced to map meteorological data (such as temperature and humidity) or electricity price signals into a vector space. The data is then weighted and fused into time series features through an attention mechanism to improve the model's generalization ability under the fluctuation of distributed energy output. Through the above-mentioned architectural improvements and optimization techniques, the number of parameters in the final lightweight load prediction model is strictly controlled between 500k and 2M, ensuring that the single inference latency on the embedded edge computing node is less than 5ms, while maintaining the prediction accuracy to meet the requirements of power distribution network scheduling.

[0040] In this embodiment, the voltage quality assessment module is constructed by including the following units: Real-time performance indicator calculation unit: This unit contains multiple parallel calculation submodules, each responsible for the real-time calculation of different voltage quality indicators. Voltage deviation calculation submodule: Real-time acquisition of instantaneous three-phase voltage values, calculation of the percentage deviation from the rated voltage using the formula Δ. U =( U real U nom ) / U nom ×100%; The three-phase unbalance calculation submodule: Based on the symmetrical component method, it calculates the ratio of the negative sequence component (or zero sequence component) to the positive sequence component in real time and generates a three-phase unbalance index. Harmonic Distortion Rate (THD) Calculation Submodule: Utilizes Fast Fourier Transform (FFT) or wavelet transform algorithms to perform frequency domain analysis on the real-time voltage signal, extracting the 2nd to 50th harmonic components, and calculating the total harmonic distortion rate. ; Voltage flicker calculation submodule: Based on IEC 61000-4-15 standard, it calculates the short-time flicker value Pst and the long-time flicker value Plt through square detection and filtering. Limit Exceeding Criteria and Threshold Management Unit: Built-in threshold database of national voltage quality standards (such as GB / T 12325) and enterprise operation and maintenance specifications; A dynamic threshold mechanism is adopted to automatically adjust the sensitivity of the criterion according to different operating periods (such as peak and valley periods); The system compares the various indicators calculated in real time with the preset thresholds and outputs a Boolean flag indicating that the limit has been exceeded. Warning signal generation and classification unit: Based on the over-limit flag and the degree of exceeding the limit, different levels of early warning signals are generated, such as general alarm, serious alarm, and emergency. A time-sliding window smoothing algorithm is used to eliminate false alarms caused by instantaneous interference, ensuring that alarms are triggered only when limits are continuously exceeded. Module collaboration and output interface: The calculation cycle of the voltage quality assessment module is synchronized with the data acquisition frequency to ensure millisecond-level response; The out-of-limit information output by the module not only triggers local audible and visual alarms, but is also pushed synchronously to the fault diagnosis model of the edge node as a feature vector to assist in comprehensive decision-making; The voltage quality statistical reports and transient waveform data are packaged and uploaded to the main station using standard communication protocols.

[0041] In this embodiment, the construction of the lightweight CNN-LSTM hybrid fault diagnosis model includes the following steps: Depthwise separable convolution is used instead of standard convolution to decouple convolution from spatial and channel dimensions, which significantly reduces computation when extracting local features in the temporal and frequency domains. We designed a parallel structure of multi-scale convolutional kernels and used 3×3, 5×5 and 7×7 convolutional kernels to capture fault features under different receptive fields, focusing on extracting fault traveling wave head steepness, high-frequency energy distribution and transient spectrum features. Batch normalization layers and GELU activation functions are introduced after convolutional layers to accelerate model convergence and improve nonlinear expressiveness. Flatten the high-dimensional feature map output by the CNN module into a sequence vector and input it into a lightweight LSTM layer; An improved LSTM gating structure is adopted, using a projection layer to compress the hidden state dimension from high dimension (e.g., 256 dimensions) to low dimension (e.g., 64 dimensions), thereby reducing the number of parameters while preserving temporal dependencies; Introducing residual connections across LSTM layers alleviates the gradient vanishing problem in deep networks; The model weights and activation values ​​are symmetrically quantized using 8-bit / 4-bit methods. Quantization-aware training is used to simulate quantization noise during the training phase to ensure inference accuracy at low bit depths. The importance of convolutional kernel channels in CNN modules is scored based on the L1 norm, and redundant channels with low contribution to fault classification are removed to make the model structure more compact. A channel attention module, SE-Block, is inserted between CNN and LSTM to automatically weight the importance of features in different frequency domains, thereby enhancing the sensitivity to weak fault features such as high-impedance grounding and ferroresonance. Through the above structural improvements and compression techniques, the total number of model parameters is strictly controlled within the range of 500k to 2M, ensuring that the single inference latency on the embedded edge computing node is less than 10ms, which meets the millisecond-level protection requirements of the power distribution network. The output layer uses a Softmax classifier to classify fault types such as single-phase grounding, phase-to-phase short circuit, open circuit, overload, ferroresonance, and high-resistance grounding, and outputs the confidence probability distribution of each category.

[0042] In this embodiment, the specific implementation of the fast criterion system based on mutation initiation and zero-order direction criteria includes: Configure a high-speed sampling interface to acquire in real time the instantaneous values ​​of three-phase current, three-phase voltage, zero-sequence current, and zero-sequence voltage collected by the edge computing node; The real-time data is subjected to sliding window (window length is 1 / 4 power frequency cycle) mean filtering to calculate the amplitude and phase at the current moment and align them with the corresponding values ​​at the previous moment or the previous cycle. A phase-separate discrimination strategy is adopted to calculate the rate of change of amplitude of current and voltage in each phase in real time. d i / d t , d u / d t (or the absolute value of the difference between sampled values); Set a fixed start threshold and a floating start threshold. When the current change or voltage change of any phase is continuous... N sampling points ( N When the value exceeds the preset threshold (2-5), a high-level "mutation start" signal is output. By introducing interlocking logic, when normal load fluctuations (such as the starting of a large motor) are detected, the automatic interlocking start signal is determined by harmonic characteristics or duration to prevent false operation. For low-current grounding or arc suppression coil grounding systems, zero-sequence current and zero-sequence voltage signals are acquired in real time; the direction of zero-sequence power or the phase difference angle between zero-sequence current and zero-sequence voltage is calculated. ; When the zero-sequence power flows from the line to the bus (or the phase difference angle) Falling within a specific range, such as 90 90 If the fault is positive, it is determined to be a forward fault; otherwise, it is determined to be a reverse fault or a bus fault. The direction enable signal is output only when a positive fault is detected; The following decision criterion is adopted: AND-OR-OR combination criterion: Mode 1 (High Sensitivity): The model is awakened when either the mutation start signal or the zero-sequence direction action signal is satisfied. Mode 2 (High Reliability): The model is only activated when both the mutation start signal and the zero-sequence direction action signal are satisfied simultaneously. The output of the rapid judgment system serves as the enable signal for the fault diagnosis model, directly controlling the opening and closing of the model's power domain or clock domain; it dynamically adjusts the threshold value for triggering sudden changes based on the current operating load rate and background noise level of the distribution network; it reduces sensitivity during off-peak periods and increases sensitivity during peak periods to ensure reliability in high-noise environments. Under normal operating conditions, the rapid judgment system remains operational, while the fault diagnosis model is in a deep dormant state. Once a wake-up signal is output, the loading and inference of the fault diagnosis model are immediately triggered via hardware interrupt or DMA request; After the fault diagnosis is completed, if the fault disappears or is reset, the system will automatically cancel the wake-up signal, and the fault diagnosis model will re-enter the sleep state after the set hold time (e.g., 500ms). As can be seen, through the above hardware-level parallel computing and logical judgment, millisecond-level fault screening is achieved, reducing the average power consumption of the fault diagnosis model by more than 70%, and ensuring low-power operation of edge nodes during fault-free periods.

[0043] In practical applications, the hardware foundation of this invention is based on embedded edge computing nodes, and its design is optimized for the typical hardware constraints (such as low power consumption and limited memory) of power distribution network terminals (e.g., FTU, DTU, TTU). Regarding processor selection, high-performance, cost-effective domestic chips are prioritized, such as the Fuxi series edge AI processors or Cambricon's Siyuan NPU acceleration chips. These chips have a main frequency of 1.5GHz to 2.5GHz, support floating-point operations and matrix acceleration, and have power consumption controlled below 5W, making them suitable for long-term stable operation in harsh outdoor environments. Simultaneously, they are compatible with international ARM Cortex-A72 / A53 series processors and equipped with at least 512MB of RAM and 4GB of eMMC storage to support model loading and data caching.

[0044] The sensor equipment configuration includes: high-precision voltage and current transformers with a measurement accuracy of 0.2%, supporting three-phase imbalance detection; a miniature phasor measurement unit with a sampling frequency of no less than 50kHz, providing real-time data such as phasor, frequency, and power factor; and auxiliary sensors such as NTC temperature sensors, MEMS vibration sensors, and UHF partial discharge sensors for multi-dimensional condition monitoring. All sensors are connected to the edge processor via analog-to-digital converters, and the sampling period can be dynamically adjusted, such as 1 second / sample in normal mode and 10 ms / sample in fault mode.

[0045] The AI ​​model deployment framework uses TensorFlow Lite Micro or Edge Impulse embedded versions, supporting edge-side model optimization and inference acceleration. The communication module integrates LoRa, NB-IoT, or 5G modules for edge-cloud data interaction, ensuring reliable transmission in weak signal environments. The power system employs a dual-backup design (main power supply + lithium battery backup) to ensure continuous fault diagnosis. The overall hardware size is controlled within 200mm × 150mm × 100mm, with an IP67 protection rating, adaptable to extreme temperature ranges from -40°C to +70°C.

[0046] The experimental process in this embodiment is divided into five stages, employing a combination of simulation and field testing to ensure comprehensive verification of the method. Key parameters and performance indicators are recorded at each stage to facilitate subsequent optimization.

[0047] 1. Data Acquisition and Simulation Environment Setup Multi-scenario distribution network models were constructed using professional power simulation software such as PSCAD / EMTDC or MATLAB / Simulink, including typical 10kV / 35kV urban distribution networks, rural ring networks, and microgrids with distributed photovoltaic / wind power. Model parameters were based on real power grid data, simulating load fluctuations, intermittent power output from renewable energy sources, and noise interference. Fault simulations covered various types, including single-phase grounding (high resistance / low resistance), phase-to-phase short circuits (two-phase / three-phase), open circuits, overloads, and resonance, generating at least 10,000 sets of multi-source datasets (voltage, current, temperature, vibration, etc.). Data acquisition was achieved through virtual sensors at a sampling rate of 50kHz, covering the period from 5 seconds before to 10 seconds after a fault. Simultaneously, real noise sources (such as electromagnetic interference and harmonic pollution) were introduced to enhance the realism of the datasets.

[0048] 2. Data Preprocessing and Feature Extraction In the edge node simulation environment, the acquired data undergoes local preprocessing. First, a digital bandpass filter (Butterworth filter, cutoff frequency 40Hz–2kHz) is applied to remove power frequency interference and high-frequency noise. Second, a db4 wavelet transform is used for multi-level threshold denoising, with adaptive threshold adjustment based on noise variance estimation. Then, data synchronization (based on GPS timestamps or NTP protocol) and normalization (Min-Max or Z-score) are performed. During the feature extraction stage, time-domain features (such as RMS value, peak value, waveform factor, and kurtosis) and frequency-domain features (such as FFT harmonic content, energy spectral density, and wavelet coefficients) are calculated, with the total feature dimension controlled within 50 dimensions. The preprocessing delay does not exceed 2ms to ensure real-time performance of subsequent inference.

[0049] 3. AI Model Training and Optimization Cloud training utilizes high-performance servers (GPUs such as NVIDIA A100), and the dataset is divided into training (70%), validation (15%), and test (15%) sets. The core model is a CNN-LSTM hybrid architecture: the CNN part uses 3 convolutional layers (3×3 kernels, 32 / 64 / 128 channels) to extract local features; the LSTM part has 2 layers (128 hidden units) to handle temporal dependencies; the output layer is a Softmax classifier that supports multi-label output. The loss function combines cross-entropy and focus loss to address imbalanced samples; the optimizer AdamW has an initial learning rate of 0.001 and uses cosine annealing scheduling. After 200 training iterations, the accuracy reached over 98%. Subsequent model optimizations included 8-bit quantization to reduce parameter precision, structured pruning to remove 30% redundant weights, and knowledge distillation to transfer knowledge from the large model (ResNet50-LSTM). The final model size was compressed to less than 1MB, and the inference speed was improved by 5 times.

[0050] 4. Algorithm Deployment and Testing The optimized model is deployed to edge nodes via OTA (Over-The-Air) and combined with a local rule engine (such as a fault criterion library based on Drools) to achieve hybrid decision-making. Testing includes unit testing (for individual modules such as preprocessing and inference) and integration testing (the entire process). Under high load scenarios, concurrent failures are simulated to evaluate system stability; noise perturbations are introduced to test robustness. In edge-cloud collaborative testing, edge nodes report feature summaries hourly, and the central station sends model updates every 24 hours. Performance monitoring tools such as TensorFlow Profiler are used to quantify latency and power consumption.

[0051] 5. Results Analysis and Iterative Optimization Collect experimental logs and analyze key metrics such as latency, accuracy, F1 score, and resource usage. If the accuracy is below 95%, backtrack to optimize model parameters or add data augmentation (rotation, flipping, noise injection). Repeat the iterative process 3-5 times until engineering standards are met.

[0052] As shown in Table 1, the following table expands upon the experimental results, covering more test conditions and comparative analyses. The data are based on 1000 repeated trials with a 95% confidence interval.

[0053] Table 1: Comparative experimental data with traditional cloud methods

[0054] Experimental results show that the method provided in this embodiment performs excellently under various conditions, especially in high-noise and complex fault scenarios, outperforming traditional cloud methods. In terms of resource consumption, the average CPU utilization of edge nodes is <30%, memory usage is <200MB, and power consumption is <3W.

[0055] Ten edge-enhanced FTUs were deployed in a 10kV ring main unit in a provincial capital city, covering a 5km line. During testing, a single-phase ground fault (injected resistance 100Ω) was simulated. The edge nodes extracted features and ran a CNN-LSTM model within 7ms, diagnosing it as a single-phase A-phase ground fault with a confidence level of 0.96. An isolation relay was automatically triggered, and only a fault summary was reported to the central station. After confirmation by the central station, optimized parameters were issued to improve the model's sensitivity to high-resistance ground faults. After three months of actual operation, the false alarm rate was <1%, the power restoration time was reduced from 30 minutes using traditional methods to 5 minutes, and economic losses were reduced by approximately 40%.

[0056] In a chemical industrial area with strong electromagnetic interference (SNR approximately 8dB), a DTU was deployed in a 35kV substation, integrating vibration and partial discharge sensors. Multiple simulations of phase-to-phase short circuits and overload faults were performed. Edge nodes used wavelet denoising to filter interference, achieving a diagnostic accuracy of 96.5% and a location error of <45m. In edge-cloud collaboration, uncertain samples were uploaded from the edge, and the central station updated the model using federated learning; after iteration, the accuracy increased to 98%. During the testing period, three false fault trips were avoided, improving production continuity.

[0057] In remote rural areas, TTU nodes are deployed along 10kV overhead lines, integrating distributed photovoltaic (PV) grid connections. Simulating voltage fluctuations caused by PV inverter failures, edge nodes sense load changes in real time, predict overload risks, and diagnose latency <10ms. Under the edge-cloud mechanism, the central station analyzes network-wide data and distributes PV output prediction sub-models to further optimize local decision-making. After implementation, the renewable energy absorption rate increased by 15%, and fault response time was reduced by 70%.

[0058] In summary, this invention is applicable to various power distribution network scenarios, such as urban high-voltage distribution networks, rural low-voltage networks, industrial park microgrids, and new energy cluster access points. It can be extended to areas such as electric vehicle charging station monitoring, demand-side response management, and smart home energy optimization. With the rise of 6G communication and quantum computing, advanced encryption (ensuring data privacy) and multimodal fusion (video + sensors) can be further integrated.

[0059] This invention, by introducing embedded edge computing technology, achieves efficient, reliable, and intelligent distribution network status perception and fault diagnosis, significantly improving the system's real-time performance, accuracy, security, and economy, and providing strong support for the intelligent transformation and sustainable development of the power system.

[0060] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0061] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for power distribution network status perception and fault diagnosis based on embedded edge computing, characterized in that, Includes the following steps: Embedded edge computing nodes are deployed on the distribution network terminal equipment. The edge computing nodes collect multi-dimensional operation data of the distribution network in real time through high-precision sensors and perform local preprocessing on the collected multi-dimensional operation data. Deploy lightweight neural network models on edge computing nodes to achieve real-time perception of the power distribution network's operating status and AI / ML fault diagnosis through real-time inference; The edge computing nodes interact with the central station through the edge-cloud collaboration mechanism. The edge-cloud collaboration mechanism includes a selective data upload strategy, federated learning model optimization, and in-depth diagnostic supplementation by the central station to optimize diagnostic results and model parameters. Among them, the edge computing nodes only upload complex / uncertain fault samples and their feature vectors, new unknown fault events, and periodically summarized statistical features. The central station performs global model retraining or incremental learning and distributes the updated model parameters to the edge computing nodes.

2. The method according to claim 1, characterized in that: Embedded edge computing nodes are integrated into the distribution network terminal equipment. The edge computing node includes a processor, memory, sensor interface and communication module. The processor is selected as a domestic dedicated edge AI acceleration chip that supports floating-point operation and matrix acceleration or a high-performance processor compatible with ARM Cortex-M / A / R series and RISC-V architecture. A variety of heterogeneous sensors, including at least high-precision voltage transformers, current transformers, miniature phasor measurement units, temperature sensors, vibration sensors, and partial discharge sensors, are connected through a high-precision sensor interface to collect multi-dimensional operating data of the line in real time. Configure sensor sampling parameters. For critical fault transient signals, set the sampling rate to no less than 50kHz to capture high-frequency transient characteristics. Real-time synchronous acquisition and caching of sensor data is achieved on edge computing nodes. The timing consistency of multi-source data is ensured through hardware timers or software triggering mechanisms, and the acquired raw data is temporarily stored in local memory.

3. The method according to claim 1, characterized in that, The collected multi-dimensional operational data is preprocessed locally at the edge computing node, including the following steps: A bandpass filter is used to filter the collected multivariate operating data to remove irrelevant frequency bands other than the 50Hz power frequency and retain the effective frequency bands containing key fault characteristics. Denoising of the filtered data can be achieved by applying wavelet threshold denoising or adaptive Kalman filtering algorithms. The denoised data is synchronized and aligned, and the consistency of multi-source data in time is ensured by timestamp matching or data interpolation methods. The data after synchronization and alignment is normalized to map it to a uniform range and remove outliers.

4. The method according to claim 1, characterized in that, Deploying lightweight neural network models on edge computing nodes to achieve real-time monitoring and evaluation of the normal operation of the distribution network includes the following steps: Deploy a lightweight load forecasting model: Based on historical data and real-time current / power sequences, use a lightweight LSTM or Transformer-mini model optimized with 8-bit / 4-bit integer quantization, structured pruning, and knowledge distillation techniques to perform short-term load forecasting. Deploy a voltage quality assessment module: calculate key voltage quality indicators in real time, including voltage deviation, three-phase unbalance, harmonic distortion rate, and flicker, determine whether each voltage quality indicator exceeds the limit by preset thresholds, and issue an early warning signal when the limit is exceeded. Deployment of equipment health assessment model: Integrating multi-dimensional signals and using unsupervised anomaly detection algorithm to assess the health status (SOH) of key equipment.

5. The method according to claim 1, characterized in that, Deploying lightweight neural network models on edge computing nodes to achieve AI / ML fault diagnosis for routine operation of distribution networks includes the following steps: A lightweight CNN-LSTM hybrid fault diagnosis model is deployed. The fault diagnosis model includes depthwise separable convolutional layers and lightweight LSTM layers. The depthwise separable convolutional layers are used to extract local time-frequency features of the preprocessed signal, and the lightweight LSTM layers are used to capture the temporal dependencies of transient processes. A fast criterion system based on mutation start-up and zero-order direction criterion rules is deployed on the edge computing nodes. The fault diagnosis model is only activated for further diagnosis when a fault feature that meets the start-up conditions is detected. During the fault diagnosis process, the fault diagnosis model outputs a confidence score for each diagnosis result; when the confidence score is lower than a preset threshold, the sample is marked as an uncertain sample and relevant feature information is recorded. A traveling wave-based double-ended or single-ended ranging algorithm is used, combined with a fault diagnosis model to correct the ranging results.

6. The method according to claim 1, characterized in that, It also performs: a comprehensive judgment and output step, including: Edge computing nodes integrate local AI diagnostic results, rule-based judgment results, and feedback information from the central station. The integration strategy employs weighted voting or Bayesian integration algorithms, assigning weights based on the reliability and importance of each information source to generate the final status report and fault alarm. The output includes the distribution network operating status level, divided into four levels: normal, attention, abnormal, and severe; fault type, occurrence time, confidence level, scope of impact, and recommended isolation scheme; and also provides the health score of key equipment and specific maintenance recommendations. The final generated status report and fault alarm are quickly reported to the main station system via standard communication protocols; Based on the final diagnostic results, the edge computing node automatically triggers the corresponding local automated actions.

7. The method according to claim 4, characterized in that, The construction of a lightweight load forecasting model includes: Long Short-Term Memory Networks or lightweight Transformer-mini models are selected as the base models; For the LSTM model, its gating mechanism is improved by reducing the number of neurons in the hidden layer and using a projection layer to map the high-dimensional hidden state to the low-dimensional output space. For the Transformer-mini model, a sparse self-attention mechanism is used instead of a fully connected self-attention mechanism, and linear positional encoding or learnable positional embedding is used to limit the length of the input sequence to adapt to the short-time prediction window. The model's weights, activations, and gradients are quantized using 8-bit / 4-bit integer quantization techniques. Quantization-aware training is introduced to simulate quantization error during training, and gradients are backpropagated through a pass-through estimator. For critical layers, a mixed precision strategy is adopted, that is, some high-precision calculations are retained to ensure numerical stability, while low-bit integer operations are used for other layers. Implement importance-based structured pruning, pruning at the neuron or convolutional kernel level; use L1 norm or L2 norm to evaluate the importance of each channel or neuron, and remove redundant structures whose contribution to the prediction results is lower than a preset threshold. Historical load data, real-time current sequences, active / reactive power sequences, and time characteristics are fused using multi-modal methods and used as model inputs. An external feature embedding layer is introduced to map meteorological data or electricity price signals into a vector space, and then weighted and fused into time series features through an attention mechanism.

8. The method according to claim 4, characterized in that, The voltage quality assessment module specifically includes the following units: Real-time performance indicator calculation unit: This unit contains multiple parallel calculation submodules, each responsible for the real-time calculation of different voltage quality indicators. Voltage deviation calculation submodule: Real-time acquisition of instantaneous three-phase voltage values, calculation of the percentage deviation from the rated voltage using the formula Δ. U =( U real U nom ) / U nom ×100%; Three-phase unbalance calculation submodule: Based on the symmetrical component method, it calculates the ratio of negative sequence component to positive sequence component in real time and generates three-phase unbalance index. The harmonic distortion rate calculation submodule uses Fast Fourier Transform or Wavelet Transform algorithms to perform frequency domain analysis on the real-time voltage signal, extract the 2nd to 50th harmonic components, and calculate the total harmonic distortion rate. ; Voltage flicker calculation submodule: Based on IEC 61000-4-15 standard, it calculates the short-time flicker value Pst and the long-time flicker value Plt through square detection and filtering. Limit Exceeding Criteria and Threshold Management Unit: Built-in threshold database set by national voltage quality standards and enterprise operation and maintenance specifications; A dynamic threshold mechanism is adopted to automatically adjust the sensitivity of the criterion according to different runtime stages; The system compares the various indicators calculated in real time with the preset thresholds and outputs a Boolean flag indicating that the limit has been exceeded. Warning signal generation and classification unit: Based on the over-limit flag and the degree to which the indicator exceeds the limit, different levels of early warning signals are generated; A time-sliding window smoothing algorithm is used to eliminate false alarms caused by instantaneous interference, ensuring that alarms are triggered only when limits are continuously exceeded.

9. The method according to claim 5, characterized in that, The construction of a lightweight CNN-LSTM hybrid fault diagnosis model includes the following steps: By replacing standard convolutional layers with depthwise separable convolutional layers, the convolutions of spatial and channel dimensions are decoupled, significantly reducing the computational cost when extracting local features in the temporal and frequency domains. We designed a parallel structure of multi-scale convolutional kernels and used 3×3, 5×5 and 7×7 convolutional kernels to capture fault features under different receptive fields, focusing on extracting fault traveling wave head steepness, high-frequency energy distribution and transient spectrum features. A batch normalization layer and the GELU activation function are introduced after the convolutional layer; Flatten the high-dimensional feature map output by the CNN module into a sequence vector and input it into a lightweight LSTM layer; An improved LSTM gating structure is used, employing a projection layer to compress the hidden state dimension from high-dimensional to low-dimensional. Introducing residual connections across LSTM layers alleviates the gradient vanishing problem in deep networks; The model weights and activation values ​​are symmetrically quantized using 8-bit / 4-bit quantization, and quantization-aware training QAT is used to simulate quantization noise during the training phase. The importance of convolutional kernel channels in CNN modules is scored based on the L1 norm, and redundant channels with low contribution to fault classification are removed. A channel attention module, SE-Block, is inserted between CNN and LSTM to automatically weight the importance of features in different frequency domains. The output layer uses a Softmax classifier to output the confidence probability distribution for each fault type.

10. The method according to claim 5, characterized in that, The specific implementation of the fast criterion system based on mutation initiation and zero-order direction criteria includes: Configure a high-speed sampling interface to acquire in real time the instantaneous values ​​of three-phase current, three-phase voltage, zero-sequence current, and zero-sequence voltage collected by the edge computing node; The real-time data is subjected to sliding window mean filtering to calculate the amplitude and phase at the current moment and align them with the corresponding values ​​at the previous moment or the previous period. A phase-separate discrimination strategy is adopted to calculate the rate of change of amplitude of current and voltage in each phase in real time. d i / d t , d u / d t (or the absolute value of the difference between sampled values); Set a fixed start threshold and a floating start threshold. When the current change or voltage change of any phase is continuous... N When a sampling point exceeds a preset threshold value, a high-level signal is output to trigger the sudden change. An interlocking logic is introduced. When normal load fluctuation characteristics are detected, the automatic interlocking start signal is determined by harmonic characteristics or duration. For low-current grounding or arc suppression coil grounding systems, zero-sequence current and zero-sequence voltage signals are acquired in real time. Calculate the direction of zero-sequence power or the phase difference angle between zero-sequence current and zero-sequence voltage. ; Setting direction criteria: When the zero-sequence power flows from the line to the bus, it is determined to be a forward fault; otherwise, it is determined to be a reverse fault or a bus fault. The direction enable signal is output only when a positive fault is detected; The output of the fast criterion system serves as the enable signal for the fault diagnosis model, directly controlling the opening and closing of the model's power domain or clock domain. The threshold value for triggering sudden changes is dynamically adjusted based on the current operating load rate and background noise level of the distribution network.

Images