Digital twin GIS partial discharge pattern recognition method

Abstract

The application relates to the field of electric power industry and discloses a digital twin GIS partial discharge mode recognition method, which comprises the following steps: S1: a high-fidelity digital twin simulation model is established according to the typical structure, operation parameters and potential abnormal working conditions of GIS equipment; S2: in the simulation model, the physical structure that may cause partial discharge is accurately modeled, and a simulation data set containing common and rare partial discharge fault scenes is simulated and generated; S3: time and space features are extracted, and a fault diagnosis model based on time and space data collaborative learning is constructed; S4: under laboratory conditions, the time and space features are refined in combination with laboratory GIS equipment and the digital twin model, and the data set is enriched by continuously collecting data, so that the model learns more fault types in the mode of incremental learning; and S5: the verified effective model is integrated into an actual engineering scene monitoring system, and intelligent recognition is carried out based on the continuous fusion of simulation and actual data.

Description

Technical Field

[0001] This invention relates to the field of the power industry, and specifically to a method for partial discharge pattern recognition in digital twin GIS. Background Technology

[0002] With the increasing demands for the safe and reliable operation of power systems, partial discharge (PD) monitoring and diagnostic technology for high-voltage gas-insulated metal-enclosed switchgear (GIS) has become a crucial link in ensuring the stable operation of the power grid. Partial discharge, as an early fault phenomenon, often indicates potential degradation of the equipment's insulation system. Failure to detect and address it in time can lead to catastrophic equipment damage and large-scale power outages. Traditional partial discharge monitoring methods, such as those relying on single sensor data, suffer from limitations in diagnostic accuracy and robustness under complex electromagnetic environments and equipment operating conditions. They are prone to false alarms and missed alarms, and are ill-suited to effectively addressing various subtle or rare fault modes.

[0003] In recent years, digital twin technology has demonstrated enormous potential in equipment condition monitoring and predictive maintenance due to its high-fidelity modeling capabilities. By constructing digital twin simulation models of GIS equipment, the physical behavior of the equipment under various ideal and non-ideal operating conditions can be simulated, generating massive amounts of diverse and precisely labeled simulation data, thereby effectively compensating for the limitations of actual fault data collection. However, how to fully mine the rich information in these multimodal simulation data and actual measurement data and transform it into accurate diagnostic capabilities remains a challenge. In particular, partial discharge signals often contain both spatial distribution and temporal evolution characteristics; simply processing spatial and temporal information separately makes it difficult to fully capture the essence of the fault.

[0004] Therefore, this study focuses on how to integrate multimodal data and combine digital twin simulation technology with advanced spatiotemporal data collaborative learning methods to improve the accuracy and reliability of partial discharge pattern recognition in GIS equipment. We aim to develop a novel method that first establishes a high-precision GIS digital twin simulation model, generating a simulation dataset containing common and rare fault scenarios through detailed physical modeling and operational condition simulation. Then, by comprehensively utilizing these simulation data and actual experimental data, we extract the joint spatial and temporal features of the partial discharge signal. Finally, by training a deep learning model capable of effectively learning these spatiotemporal collaborative features and performing refined hyperparameter optimization, we achieve intelligent, accurate, and efficient recognition of partial discharge patterns in GIS equipment. Ultimately, this method can be applied to practical engineering scenarios, providing strong protection for the operational safety of power equipment. Summary of the Invention

[0005] The purpose of this invention is to provide a digital twin GIS partial discharge pattern recognition method, which combines simulation data with experimental data and uses an algorithm model for spatiotemporal multimodal data fusion to form a method applicable to the field of GIS partial discharge pattern recognition.

[0006] The objective of this invention, a digital twin GIS partial discharge pattern recognition method, is achieved through the following technical solution:

[0007] A method for partial discharge pattern recognition in digital twin GIS, the specific implementation steps of which are as follows:

[0008] S1: Establish a high-fidelity digital twin simulation model based on the typical structure, operating parameters and potential abnormal working conditions of GIS equipment;

[0009] S2: In the simulation model, accurate modeling is performed on the physical structures that may cause partial discharge, and simulation datasets containing common and rare partial discharge fault scenarios are generated. This part regarding physical structures that may cause partial discharge refers to the accurate modeling of electric field distortion points generated at metal connections of equipment, metal-gas-solid insulation interfaces, and internal defects in equipment (scratches, sharp points, metal particles, internal bubbles, cracks, etc.) that may lead to partial discharge. The simulation dataset for common partial discharge fault scenarios includes, but is not limited to, points of concentrated field strength inside GIS equipment caused by structure or defects. The simulation dataset for rare partial discharge fault scenarios includes, but is not limited to, rare gas discharge phenomena induced by factors such as gas insulation quality (purity and gas pressure) and abnormal operating environment (drastic temperature changes, moisture intrusion, overload).

[0010] S3: Extract spatiotemporal features and construct a fault diagnosis model based on spatiotemporal data collaborative learning;

[0011] S4: In laboratory conditions, combined with laboratory GIS equipment and digital twin simulation models, the spatiotemporal characteristics are refined, and the dataset is enriched by continuously collecting data, so that the model can learn more fault types through incremental learning.

[0012] S5: Integrate proven and effective models into the monitoring system for actual engineering scenarios, and perform intelligent identification, early warning and optimization based on the continuous fusion of simulation and actual data.

[0013] Furthermore, step S1 includes the following sub-steps:

[0014] S11: Collect typical structural information of GIS equipment, including but not limited to the equipment's external dimensions, internal structural layout, insulation medium type and parameters, geometric shape of conductive components, and installation method;

[0015] S12: Obtain key operating indicators of the GIS equipment, including but not limited to operating voltage, current, temperature, humidity, pressure, and frequency;

[0016] S13: Analyze potential abnormal operating conditions that may occur in GIS equipment, and identify physical factors that may induce partial discharge, including insulation defects, metallic foreign objects, surface roughness, contaminant accumulation, overvoltage or undervoltage conditions;

[0017] S14: Based on the collected structural information, operating parameters, and analyzed abnormal operating conditions, a high-fidelity three-dimensional geometric model and material property model of the GIS are constructed in COMSOL simulation software, and multi-physics coupling simulation is performed to accurately simulate the physical behavior of GIS equipment under different operating conditions.

[0018] Furthermore, step S2 includes the following sub-steps:

[0019] S21: Based on the established simulation model, detailed physical principle modeling is performed on the common fault physical mechanisms that cause partial discharge in GIS equipment to obtain common fault models. The common fault models include, but are not limited to, electric field distribution, insulation dielectric breakdown characteristics, and charge accumulation and release processes.

[0020] S22: For rare fault conditions that may cause partial discharge in GIS equipment, we conduct an in-depth analysis of their occurrence conditions and development process, and set corresponding boundary conditions and initial conditions for these rare fault conditions in the established simulation model to simulate their potential occurrence process.

[0021] S23: Combining the common fault modeling in S21 and the rare fault condition simulation in S22, several simulation calculations are initiated. Each simulation is for a specific fault mode, condition or combination thereof, and outputs multimodal simulation data including partial discharge phenomena.

[0022] S24: Collect, organize, and label the obtained multimodal simulation data to form a structured simulation dataset. The simulation dataset should contain multimodal simulation data under different common and rare partial discharge fault scenarios. The multimodal simulation data includes surface discharge, suspension discharge, air gap discharge, and tip discharge. Each data record is labeled with the corresponding fault mode, occurrence location, and operating parameters.

[0023] Furthermore, step S3 includes the following sub-steps:

[0024] S31: Preprocess the acquired raw dataset: Apply wavelet denoising technology to the raw data, using the db4 wavelet as the wavelet basis function to decompose the data into 5 levels; after thresholding the high-frequency coefficients after decomposition, reconstruct the denoised dataset through inverse wavelet transform to remove random noise, sensor noise, background signal interference, etc.; apply linear normalization to the denoised data to map it to the numerical range of [0,1].

[0025] S32: For the preprocessed data, its inherent spatial structure-related features are considered. These features include the sensor array distribution pattern of partial discharge signals at different locations of the equipment and the influence of the equipment's geometry on the signal propagation path. A convolutional neural network-based algorithm is used as the spatial feature extractor, specifically:

[0026] When extracting spatial features, convolutional neural network models use convolutional kernels to extract data features by sliding across feature vectors, as shown in the following formula:

[0027] ;

[0028] ;

[0029] In the formula, This represents the output after convolution. I Represents the input matrix, K Indicates size is m × n The convolution kernel;

[0030] S33: For the time series characteristics of the preprocessed data, the time series characteristics include the variation trend of the partial discharge signal over time, the time interval of the pulse sequence, and the rising and falling edges of the signal; an algorithm based on an LSTM network is used as the time feature extractor.

[0031] S34: The spatial features extracted by CNN and the time series features extracted by LSTM network are fused to form a spatiotemporal co-representation. This representation is then concatenated with the original preprocessed data and used as the input to the spatiotemporal co-learning model. In the feature fusion process, a combination of feature weighted fusion and multimodal attention mechanism is used. A fully connected layer is used to set the weights of spatiotemporal information, so that the final model can learn the most critical parts of the spatiotemporal features.

[0032] S35: Divide the processed input dataset into three parts: 70% for training, 20% for validation, and 10% for testing. Use a systematic hyperparameter optimization method to tune hyperparameters of the model, such as learning rate and batch size, until the training results of the model can meet the expected accuracy and robustness requirements.

[0033] Furthermore, in step S31, the linear normalization method is applied, and its calculation method is shown in the following formula:

[0034]

[0035] in, and These represent the values ​​of each variable in the data before and after normalization, respectively. For and These represent the minimum and maximum values ​​of each variable in the data, respectively.

[0036] Furthermore, the CNN model includes an input layer, convolutional blocks, flattening layers, and fully connected layers; each convolutional block consists of a convolutional layer, an activation function, and a pooling layer; the first convolutional block contains 32 filters in its convolutional layer, with a filter size of 3×3. The local receptive field extracts features through weight sharing and feature mapping. The activation function uses the ReLU function to introduce non-linearity. The pooling layer uses max pooling to reduce dimensionality, and the pooling window size is set to 2×2; the second convolutional block changes the number of filters to 64, and the third convolutional block has 128.

[0037] Furthermore, the LSTM network consists of two stacked LSTM layers. The first LSTM layer consists of 128 neuron units, with return_sequences set to True, the activation function for internal gate control set to Sigmoid, and the activation functions for state and output set to tanh. The second LSTM layer changes the number of neuron units to 64 based on the first layer.

[0038] Furthermore, step S4 includes the following sub-steps:

[0039] S41: Based on the weight dependence of the spatiotemporal collaborative model on different partial discharge pattern recognition tasks, dynamically adjust the digital twin simulation model to enhance the targeting of feature acquisition; for the constructed spatiotemporal collaborative learning model, perform feature weight analysis and probe configuration optimization, and quantify the weight of spatial features and temporal features in different types of partial discharge pattern recognition tasks.

[0040] S42: Based on the weighted quantitative analysis of spatiotemporal features, for patterns highly dependent on spatial features, the digital twin simulation model will be used to specifically refine the modeling of these key areas and precisely deploy virtual probes to collect denser, higher-resolution spatial distribution data. For patterns highly dependent on temporal features, the range and sampling frequency of the virtual probes will be adjusted in the simulation model to ensure accurate capture of these fine temporal characteristics. If a high sensitivity to a specific frequency component is identified, the sampling frequency of the simulation model will be set to at least twice the Nyquist frequency of the target frequency component, and the sampling window will be optimized based on the instantaneous rate of change of the signal. For key areas in real GIS equipment where effective data is difficult to collect directly due to complex structure, inconvenient installation, or insufficient accuracy of existing sensors, high-quality supplementary data will be generated using the digital twin simulation model. This supplementary data will be used to specifically simulate the discharge phenomena that may occur in these areas based on the analysis results of the previous steps, and generate corresponding spatiotemporal feature data to compensate for the deficiencies in actual data collection.

[0041] S43: Model verification and iteration are conducted on a laboratory testing platform, and model performance is continuously improved through incremental data learning. The constructed laboratory testing platform will simultaneously run a digital twin simulation model and actual GIS equipment. By accurately simulating various typical and rare partial discharge conditions, real sensor data deployed on the laboratory GIS equipment will be collected, and these real data will be compared and analyzed with the simulation data generated by the digital twin simulation model. The focus will be on verifying the recognition accuracy, positioning accuracy, and fault mode classification capability of the spatiotemporal collaborative model constructed in the previous step under different conditions. By comparing the model's prediction results for real and simulated data, the model's performance indicators will be quantified, including accuracy, precision, recall, and F1 score.

[0042] S44: Incremental Data Learning and Fault Type Expansion: During laboratory testing, real sensor data will be continuously and over a long period of time, and this newly acquired data will be periodically and continuously input into the spatiotemporal collaborative model to achieve incremental learning. The incremental learning strategy will ensure that the model can continuously learn and adapt, and gradually master new fault types and operating scenarios. After each incremental learning, the model's performance will be re-evaluated on the test set, and its performance improvement will be compared with that of the initial model.

[0043] S45: Sensor Optimization Based on Digital Twin Assistance: Iteratively optimize the sensor distribution strategy in actual GIS equipment using the assistance of a digital twin simulation model; evaluate the impact of different sensor layout schemes on the recognition performance of the spatiotemporal collaborative model through simulation experiments, and identify deficiencies in the number, location, type, and sampling parameters of sensors; apply the optimization suggestions based on the simulation results to the sensor layout of the laboratory test platform.

[0044] Furthermore, step S5 includes the following sub-steps:

[0045] S51: Deployment and integration of the model in actual engineering monitoring systems; The spatiotemporal collaborative model, validated and optimized in a laboratory environment to meet accuracy and robustness requirements, will be integrated into a GIS equipment status monitoring system in actual engineering scenarios; This GIS equipment status monitoring system will have the functions of real-time acquisition of actual engineering data, intelligent identification by calling the model, and generation of fault early warning information;

[0046] S52: Intelligent Identification, Early Warning, and Fault Location: The integrated model will perform real-time and intelligent identification of GIS equipment operation data collected in actual engineering scenarios; when an abnormal signal that may indicate partial discharge is detected, the model will immediately trigger an early warning mechanism and provide high-precision fault location information based on the identified fault mode, location, and confidence level; maintenance personnel will receive a detailed warning report, including the fault type, possible root cause, risk level, and recommended handling measures;

[0047] S53: Simulation-assisted operation and maintenance decision-making and continuous model optimization; operation and maintenance personnel will proactively arrange equipment maintenance and troubleshooting based on the fault warning and location results of the digital twin simulation model, thereby minimizing the probability of further faults and reducing power outage losses; during the operation of actual engineering scenarios, real fault signals and normal operation data will be continuously collected and accumulated, and these actually collected data will be fed back into the digital twin simulation model to update and optimize the parameters and physical model of the simulation model; the actually collected fault signals will be added as new samples to the training dataset of the spatiotemporal collaborative model, and through periodic retraining or online learning, the model's identification and diagnostic capabilities will be continuously improved and optimized, ultimately forming a mature and dynamically evolving GIS partial discharge fault diagnosis and operation and maintenance method.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] This invention combines digital twin simulation technology with advanced spatiotemporal data collaborative learning methods to improve the accuracy and reliability of partial discharge pattern recognition in GIS equipment. Our aim is to develop a novel method that first establishes a high-precision GIS digital twin simulation model, generating a simulation dataset containing common and rare fault scenarios through detailed physical modeling and operational condition simulation. Then, by comprehensively utilizing these simulation data and actual experimental data, joint spatial and temporal features of the partial discharge signal are extracted. Finally, a deep learning model capable of effectively learning these spatiotemporal collaborative features is trained and its hyperparameters are finely optimized to achieve intelligent, accurate, and efficient recognition of partial discharge patterns in GIS equipment. Ultimately, this method can be applied to practical engineering scenarios, providing strong protection for the safe operation of power equipment. Attached Figure Description

[0050] The accompanying drawings of this invention are described below.

[0051] Figure 1 This is a flowchart of the method according to an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of a GIS simulation model according to an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of the spatial feature extraction process according to an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of the time feature extraction process according to an embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram of L1 and L2 regularization in an embodiment of the present invention. Detailed Implementation

[0056] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0057] Example:

[0058] The flowchart of the solution is as follows Figure 1 As shown.

[0059] Taking a typical SF6 gas-insulated metal-enclosed switchgear (GIS) as an example, its three-dimensional structural data was collected. Specifically, the external dimensions of the GIS were first obtained; then, the internal structure was analyzed to obtain the geometry and material properties of the post insulators and wall bushings, as well as the geometry and surface roughness of the high-voltage conductors; simultaneously, parameters such as the density and humidity of the SF6 gas inside the equipment, and the installation method of each component were recorded. Key operating parameters of the GIS under normal operating conditions were monitored and recorded. These parameters include, but are not limited to: phase voltage, phase current, grounding current, internal temperature, SF6 gas pressure and humidity, and grid frequency. Using COMSOL simulation software, based on the collected structural information, operating parameters, and analyzed abnormal operating conditions, a high-fidelity three-dimensional geometric model of the GIS was constructed. Accurate material properties were assigned to different components in the model, and multiphysics coupling simulation was performed on this basis. A schematic diagram of the GIS simulation model is shown below. Figure 2 As shown.

[0060] For common partial discharge physical mechanisms, detailed physical principle modeling is performed in the simulation model. The model accurately simulates micro-cracks on the insulator surface, adjusting their size and shape, and combining this with electric field analysis to calculate the induced electric field enhancement effect and predict whether the local electric field intensity will reach the breakdown threshold of the insulating medium. Sharp edges on the conductor surface are accurately simulated, analyzing the resulting electric field concentration and calculating potential corona discharge. For rare fault conditions that may occur, their occurrence conditions are analyzed in depth. Appropriate boundary and initial conditions are set for these rare conditions in the simulation model. For example, if a rare surface contaminant causes a localized decrease in insulation performance, a contaminant layer with low dielectric strength or high conductivity can be simulated in the simulation model, with corresponding surface charge density and surface conductivity set as partial boundary conditions. Then, several simulation calculations are systematically initiated. Each simulation targets one or a set of specific fault modes, conditions, or combinations thereof. For each simulation, multimodal simulation data containing partial discharge phenomena is output. The output data of each simulation are collected, organized and labeled to form a structured simulation dataset.

[0061] Preprocessing was performed on the generated simulation data and existing GIS partial discharge experimental data. First, the raw data was denoised and filtered. Then, the data was normalized to allow the model to efficiently extract features from data of different magnitudes. A spatial feature extractor was used to extract spatial structure-related features, such as… Figure 3 As shown, for example, the distribution of partial discharge signals at different locations of the equipment, the influence of the equipment's geometry on signal propagation, etc., are extracted using a time-series feature extractor to extract relevant features, such as... Figure 4As shown, the spatiotemporal features include, for example, the trend of partial discharge signal changes over time, the time interval of the pulse sequence, the rising and falling edges of the signal, and the periodic changes of the signal. The extracted spatiotemporal features are fused, and the fused spatiotemporal collaborative features, along with the preprocessed data, are used as input to the spatiotemporal data collaborative learning model. The model output is set as specific modes of different types of partial discharge, including classification labels such as "surface discharge," "suspended discharge," "air gap discharge," and "tip discharge." Furthermore, based on the dimensionality and complexity of the data, the hierarchical structure and number of layers of the spatiotemporal data collaborative learning model are initially designed.

[0062] The preprocessed and merged dataset is divided into training, validation, and test sets according to a preset ratio. A suitable ratio is 70% for training, 15% for validation, and 15% for testing. The training set is used for learning model parameters, the validation set for hyperparameter tuning and model performance monitoring, and the test set for final model evaluation. On the training set, the Adam optimizer is selected, employing a gradient descent-type optimization algorithm. For classification tasks, the cross-entropy loss function is used, and training data is input into the model in batches with a preset batch size. The model iteratively updates the weight parameters of each layer to minimize the error between the model's predicted values ​​and the true labels, enabling the model to learn patterns in the data. During model training, performance metrics on the validation set are used to dynamically monitor model performance. If the model's performance on the validation set no longer improves, the training process is adjusted or a regularization step is added according to a preset strategy, such as... Figure 5 As shown, to avoid overfitting and thus improve the model's generalization ability, a systematic hyperparameter optimization method is employed to tune the key hyperparameters of the model. The optimization goal is to achieve the highest accuracy in partial discharge pattern recognition on the validation set, or to meet a preset accuracy requirement.

[0063] The optimized model is deployed to a specially built laboratory testing platform, and monitoring data representing various typical and special partial discharge scenarios generated by the platform itself are input. Under controlled laboratory conditions, the model undergoes rigorous performance evaluation. The model's effectiveness and robustness are verified by comparing its identification results with actual introduced faults. The validated model is then integrated into a monitoring system in a real-world engineering scenario. Specifically, this can be achieved by deploying the model in a monitoring unit next to the GIS equipment or on a cloud server. This system connects to various sensors on the GIS equipment to collect real-time operational data. The deployed model intelligently identifies and continuously monitors the collected real-time operational data. In this way, early warning of partial discharge in GIS equipment is achieved. Once an abnormal signal is detected, an alarm is issued promptly, and fault location is attempted, providing decision support for maintenance personnel. Based on the test results and operational feedback in real-world engineering scenarios, the model is further optimized. By retraining the model, its identification and diagnostic capabilities are continuously improved and optimized. This continuous learning and optimization process ultimately leads to a more mature and reliable GIS partial discharge fault diagnosis method to cope with increasingly complex and changing real-world operating environments.

[0064] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for partial discharge pattern recognition of digital twin GIS, characterized in that, The specific implementation steps are as follows: S1: According to the typical structure, operating parameters and potential abnormal conditions of GIS equipment, a high-fidelity digital twin simulation model is established; the step S1 includes the following sub-steps: S11: Collect the typical structure information of GIS equipment, including the external dimensions, internal structure layout, insulation medium type and parameters, geometric shape of conductive components and installation method; S12: Obtain the key operating indicators of GIS equipment, including operating voltage, current, temperature, humidity, pressure and frequency; S13: Analyze the potential abnormal conditions that may occur in GIS equipment, identify the physical factors that may induce partial discharge, including insulation defects, metal foreign matter, surface roughness, pollutant accumulation, overvoltage or under-voltage state; S14: Based on the collected structure information, operating parameters and analyzed abnormal conditions, a high-fidelity three-dimensional geometric model and material attribute model of GIS is constructed in COMSOL simulation software, and multi-physical field coupling simulation is performed to accurately simulate the physical behavior of GIS equipment under different conditions; S2: In the digital twin simulation model, the physical structure that may cause partial discharge is accurately modeled, and a simulation data set containing common and rare partial discharge fault scenarios is simulated; the step S2 includes the following sub-steps: S21: Based on the established simulation model, detailed physical principle modeling is performed for common fault physical mechanisms that cause partial discharge in GIS equipment to obtain common fault modeling, including electric field distribution, insulation medium breakdown characteristics, charge accumulation and release process; S22: For rare fault conditions that may cause partial discharge in GIS equipment, the generation conditions and development process are analyzed in depth, and appropriate boundary conditions and initial conditions are set for these rare fault conditions in the established simulation model to simulate their potential occurrence process; S23: Combined with the common fault modeling in S21 and the rare fault condition simulation in S22, several simulation calculations are started, each simulation is for a specific fault mode, condition or combination, and multi-modal simulation data containing partial discharge phenomena are output; S24: Collect, organize and label the obtained multi-modal simulation data of each time to form a structured simulation data set, which should contain multi-modal simulation data under different common and rare partial discharge fault scenarios; The multi-modal simulation data includes surface discharge, floating discharge, air gap discharge and sharp tip discharge, and each data record is labeled with the corresponding fault mode, occurrence position and condition parameter; S3: Extract the spatio-temporal features and build a fault diagnosis model based on spatio-temporal data collaborative learning; S4: Under laboratory conditions, combine laboratory GIS equipment and digital twin simulation model to refine spatio-temporal features, and enrich the data set by continuously collecting data, so that the model learns more fault types through incremental learning; S5: Integrate the verified effective model into the actual engineering scene monitoring system, and based on the continuous fusion of simulation and actual data, perform intelligent identification, early warning and optimization.

2. The digital twin GIS partial discharge pattern recognition method of claim 1, wherein: The step S3 comprises the following sub-steps: S31: preprocessing the obtained original data set: applying wavelet denoising technology to the original data, decomposing the data to 5 layers; after threshold processing of the decomposed high-frequency coefficients, the denoised data set is reconstructed by inverse wavelet transform; applying linear normalization method to the denoised data, mapping it to the numerical interval [0, 1]; S32: considering the internal spatial structure related characteristics of the preprocessed data; using the algorithm based on the CNN model as a spatial feature extractor; S33: for the time series characteristics of the preprocessed data, the time series characteristics include the trend of the partial discharge signal over time, the time interval of the pulse sequence, the rising edge and the falling edge of the signal; using the algorithm based on the LSTM network as a time feature extractor; S34: the spatial features extracted by the CNN model are fused with the time series features extracted by the LSTM network to form a spatio-temporal collaborative feature representation, and the original preprocessed data is dimensionally spliced as the input of the spatio-temporal collaborative learning model; in the process of feature fusion, the form of feature weighted fusion combined with multi-modal attention mechanism fusion is adopted, and a fully connected layer is used to set the weight of spatio-temporal information, so that the final model can learn the most critical part of the spatio-temporal features; S35: dividing the obtained input data set after processing; using a systematic hyperparameter optimization method to optimize the hyperparameters of the model until the training results of the model can meet the expected accuracy and robustness requirements.

3. The digital twin GIS partial discharge pattern recognition method of claim 2, wherein: In step S31, the linear normalization method is used, and its calculation method is as follows: in, and These represent the values ​​of each variable in the data before and after normalization, respectively. For and These represent the minimum and maximum values ​​of each variable in the data, respectively.

4. The digital twin GIS partial discharge pattern recognition method of claim 2, wherein: The CNN model comprises an input layer, a convolution block, a flattening layer and a fully connected layer; each convolution block is composed of a convolution layer, an activation function and a pooling layer; the convolution layer of the first convolution block comprises 32 filters, and the filter size is set to 3x3; the local receptive field therein extracts features through weight sharing and feature mapping, the activation function adopts ReLU function to introduce nonlinearity, and the pooling layer adopts maximum pooling to reduce dimension, and the pooling window size is set to 2x2; the number of filters of the second convolution block is changed to 64 based on the first convolution block, and the number of filters of the third convolution block is changed to 128.

5. The digital twin GIS partial discharge pattern recognition method of claim 2, wherein: The LSTM network is composed of two stacked LSTM layers, the first LSTM layer is composed of 128 neuron units, return_sequences is set to True, the activation function of the inner gate is set to Sigmoid, and the activation functions of the state and output are set to tanh; the number of neuron units of the second LSTM layer is changed to 64 based on the first layer.

Images