A method for assessing the health status of power transformers

Abstract

This invention discloses a method for assessing the health status of power transformers, comprising: first, acquiring the operating status data of the power transformer; second, establishing a digital twin model of the power transformer, including a geometric model, a physical model, a behavioral model, and a rule model; third, calculating whether there is a deviation between the operating status data of the physical power transformer and the operating status data of the virtual power transformer in the digital twin model; and fourth, simulating the health degradation path of the power transformer under different operating environments and loads in the digital twin model to generate health degradation data of the power transformer, and assessing the health status of the power transformer based on the health degradation data. This invention applies digital twin technology to the health status assessment of power transformers, and through virtual-real interaction and data fusion, it can provide rich operating data for the physical power transformer, reflecting its operating status more comprehensively.

Description

Technical Field

[0001] This invention relates to a method for assessing the health status of power transformers, which is a method for assessing the health status of power transformers based on a spatiotemporal attention mechanism network under digital twins. Background Technology

[0002] Power transformers are one of the core pieces of equipment in a power system. Their health is crucial for the safe and stable operation of the power system; any fault directly impacts its normal operation and safety. Therefore, assessing the health status of power transformers based on their operational information allows for timely detection and repair of faults, ensuring their safe and stable operation and minimizing economic losses caused by transformer failures.

[0003] Currently, the health status assessment methods for power transformers mainly rely on health indices and scoring tables. These methods are based on historical data of the power transformers and typically only use currently sampled data, facing the problem of data gaps. This results in low efficiency and accuracy in power transformer health status assessment. Digital twin technology can effectively solve these problems by creating a virtual power transformer identical to the physical one in a virtual environment, simulating the operation of the physical power transformer. Through virtual-real interaction and data fusion, rich operational data can be provided to the physical power transformer, more comprehensively reflecting its operating status. Therefore, applying digital twin technology to the health status assessment of power transformers is an important research direction. Summary of the Invention

[0004] The purpose of this invention is to propose a method for assessing the health status of power transformers, which is a power transformer health status assessment method based on spatiotemporal attention mechanism networks, in order to overcome the above-mentioned shortcomings and improve the accuracy of power transformer health status assessment.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A method for assessing the health status of a power transformer, wherein the method includes:

[0007] Step 1: Obtain the operating status data of the power transformer, including:

[0008] The percentage data of five dissolved gases (H2, CH4, C2H6, C2H4, and C2H2) in the insulating oil, along with voltage, current, temperature, and oil level operating status data, will be stored in a database.

[0009] Step 2: Establish a digital twin model of the power transformer, including: geometric model, physical model, behavioral model, and rule model; among which:

[0010] Geometric model: Based on the shape, size and positional relationship of each component and shell of the physical power transformer, the winding method of the primary and secondary windings, the assembly and insertion relationship of the leads and iron core, and the bushing assembly structure parameters, a 3D model of the power transformer with three dimensions of x-axis, y-axis and z-axis is formed.

[0011] Physical model: Based on the geometric model, it incorporates strain analysis test data of transformer coil windings and insulation under the action of impulse voltage and electric field force, as well as performance test data of transformer tank and various components under external force, such as compressive strength and damage resistance.

[0012] Behavioral model: Based on the mapping relationship between the operating status data and health status of the power transformer entity, simulated operating status data of its changes over operating time is generated;

[0013] Rule-based model: Construct a power transformer state assessment model based on a spatiotemporal attention mechanism network, including an encoder, decoder, and classifier;

[0014] Step 3: Calculate whether there is a discrepancy between the operating status data of the actual power transformer and the operating status data of the virtual power transformer in the digital twin model. If there is a discrepancy, adjust the parameters of the digital twin model according to the actual operating status data. Repeat this step until the digital twin model of the power transformer matches the actual power transformer.

[0015] Step 4: In the digital twin model, simulate the health degradation path of power transformers under different operating environments and loads, generate health degradation data of power transformers, and evaluate the health status of power transformers based on the health degradation data.

[0016] A further aspect of the solution is that the power transformer condition assessment model is constructed by training a neural network with known samples as input.

[0017] The solution further includes: the network structure of the encoder is a gated recurrent unit network based on spatial attention mechanism, and the structure of the decoder is a gated recurrent unit network based on temporal attention mechanism.

[0018] The beneficial effects of this invention are:

[0019] 1. Applying digital twin technology to the health status assessment of power transformers: By simulating the real-time operation of power transformers through digital twin technology, and through virtual-real interaction, data fusion and other means, rich operating data can be provided for physical power transformers, which can more comprehensively reflect the operating status of power transformers.

[0020] 2. This invention proposes a spatiotemporal attention mechanism network to distinguish the relative importance of different sensors in the health status of power transformers and to consider different degrees of temporal dependencies in historical data, so as to improve the accuracy of power transformer health status assessment.

[0021] 3. The method proposed in this invention can be applied to the health status assessment of power transformers in various scenarios, and can quickly and accurately assess the status of power transformers, thus having high application value.

[0022] The present invention will now be described in detail with reference to the embodiments. Detailed Implementation

[0023] A method for assessing the health status of a power transformer, wherein the method includes:

[0024] Step 1: Obtain the operating status data of the power transformer, including:

[0025] The percentage data of five dissolved gases (H2, CH4, C2H6, C2H4, and C2H2) in the insulating oil, along with voltage, current, temperature, and oil level operating status data, will be stored in a database.

[0026] Step 2: Establish a digital twin model of the power transformer, including: geometric model, physical model, behavioral model, and rule model; among which:

[0027] Geometric model: Based on the shape, size and positional relationship of each component and shell of the physical power transformer, the winding method of the primary and secondary windings, the assembly and insertion relationship of the leads and iron core, and the bushing assembly structure parameters, a 3D model of the power transformer with three dimensions of x-axis, y-axis and z-axis is formed.

[0028] Physical model: Based on the geometric model, it incorporates strain analysis test data of transformer coil windings and insulation under the action of impulse voltage and electric field force, as well as performance tests such as compressive strength and damage resistance of transformer tank and various components under the action of external force;

[0029] Behavioral model: Based on the mapping relationship between the operating status data and health status of the power transformer entity, simulated operating status data of its changes over operating time is generated;

[0030] Rule-based model: Construct a power transformer state assessment model based on a spatiotemporal attention mechanism network, including an encoder, decoder, and classifier;

[0031] Step 3: Calculate whether there is a discrepancy between the operating status data of the actual power transformer and the operating status data of the virtual power transformer in the digital twin model. If there is a discrepancy, adjust the parameters of the digital twin model according to the actual operating status data. Repeat this step until the digital twin model of the power transformer matches the actual power transformer.

[0032] Step 4: In the digital twin model, simulate the health degradation path of power transformers under different operating environments and loads, generate health degradation data of power transformers, and evaluate the health status of power transformers based on the health degradation data.

[0033] Wherein: the power transformer condition assessment model is constructed by training a neural network with known samples as input; the encoder network structure is a gated recurrent unit network based on spatial attention mechanism, and the decoder structure is a gated recurrent unit network based on temporal attention mechanism.

[0034] The following is a detailed description of the method:

[0035] Step 1: Use sensors to collect signals from the physical power transformer to obtain operating status data of the power transformer in the physical system.

[0036] Step 1.1: Collect operational status data of the physical power transformer. This mainly includes using gas chromatography to analyze the percentage of five dissolved gases (H2, CH4, C2H6, C2H4, and C2H2) in the insulating oil (dissolved gas analysis in oil refers to analyzing the gases dissolved in the insulating oil of power equipment, and diagnosing abnormal phenomena of the equipment based on the composition, content, and changes of the gases. For example, when abnormal phenomena such as local overheating or partial discharge occur in the transformer, gases such as H2, C2H6, CH4, C2H4, and C2H2 are generated, most of which are dissolved in the insulating oil. Therefore, the gases are extracted from the insulating oil, and the composition of the gases is analyzed using gas chromatography to obtain the percentage of H2, C2H6, CH4, C2H4, and C2H2 in all dissolved gases). Also, collect operational status data such as voltage, current, temperature, and oil level, and store the collected data in a database.

[0037] Step 2: Based on the working principle of a physical power transformer, establish a digital twin model of the power transformer, mainly including a geometric model, a physical model, a behavioral model, and a rule model, for the simulation, optimization, evaluation, and real-time monitoring of the working process of the physical power transformer.

[0038] Step 2.1: For the geometric model mentioned in Step 2, based on the key characteristic parameters of the physical power transformer, including the shape, size and positional relationship of each component and shell of the power transformer, the winding method of the primary and secondary windings, the assembly and insertion relationship of the leads and core, the bushing assembly structure, etc., a virtual 3D model of the power transformer is established using modeling software from three dimensions: x-axis, y-axis and z-axis.

[0039] Step 2.2: For the physical model described in Step 2, obtain the structural parameters of the power transformer entity. This mainly includes strain analysis and testing of the transformer coil winding and insulation under the action of impulse voltage and electric field force, as well as performance tests such as pressure resistance and damage resistance of the transformer tank and its components under the action of external force. Then, collect the structural parameter data of the power transformer entity and add it to the geometric model.

[0040] Step 2.3: For the behavioral model described in Step 2, the electromagnetic induction principle and the coil turns ratio are used to realize the voltage step-up and step-down functions to simulate the operation of the power transformer. Based on the mapping relationship between the operating status data and health status of the power transformer entity, simulated operating status data of its changes over time is generated as a supplement to the data of the physical transformer. This mainly includes the percentage of the five dissolved gases H2, CH4, C2H6, C2H4 and C2H2 in the dissolved gases of the insulating oil under the digital twin module, as well as operating status data such as voltage, current, temperature and oil level.

[0041] Step 2.4: Calculate whether there is a deviation between the operating status data of the physical power transformer and the operating status data of the virtual power transformer in the digital twin module. If there is a deviation, adjust the parameters of the digital twin model according to the actual operating status data. Repeat this step until the digital twin model of the power transformer matches the physical power transformer.

[0042] Step 2.5: In the digital twin module, simulate the health degradation path of power transformers under different operating environments and different loads to generate health degradation data of power transformers;

[0043] Step 3: Preprocess the data and construct input samples. Create a health status assessment model based on a spatiotemporal attention mechanism in the health status assessment module to assess the status of power transformers and verify the performance of the model.

[0044] Step 3.1: Merge the data collected in Step 1.1 and Step 2.5, which can be represented as D = {(x1,y1),(x2,y2),...,(x...} i ,y i ),...,(x n ,y n )},in, Let be a vector composed of sensor data from the i-th sample. For x i The j-th sensor data, y i Let R represent the transformer health status of the i-th sample, n represent the number of data samples collected, and m represent the number of sensors, where 1 ≤ i ≤ n and 1 ≤ j ≤ m.

[0045] Step 3.2: Preprocess the data. Due to the dynamic dependencies in the data, a sliding window is used to process the data and construct the model input samples. Assuming the length of the sliding window is P, the processed data can be denoted as... Among them, X i =[x i ,x i+1 ,...,x i+p-1 ,...,x i+P-1 [x] represents the i-th input sample. i+p-1 Y represents the p-th data point in the i-th input sample, where 1 ≤ p ≤ P; i The transformer health status for the i-th input sample;

[0046] Step 3.3: For the rule model described in Step 2, construct a power transformer state assessment model based on a spatiotemporal attention mechanism network, mainly including an encoder, a decoder, and a classifier. The encoder consists of gated recurrent units based on a spatial attention mechanism, the decoder consists of gated recurrent units based on a temporal attention mechanism, and the classifier consists of fully connected layers. The calculation process of input samples in the power transformer state assessment model based on the spatiotemporal attention mechanism network is as described in Steps 3.4-3.12.

[0047] Step 3.4: The encoder's network structure is a gated recurrent unit network based on a spatial attention mechanism, which is used to calculate the relative importance of different sensors. The step size is the length P of the sliding window. The network parameters are initialized, and the i-th input sample X is... i =[x i ,x i+1 ,...,x i+p-1 ,...,x i+P-1 The input sample is fed into the encoder, and the calculation process of the input sample in the encoder is as described in steps 3.5-3.7;

[0048] Step 3.5: When time step p = 1, initialize the hidden state vector of the gated loop unit as follows: Will The query is used as the i-th input sample in the spatial attention mechanism; when p>1, the hidden state vector of the gated recurrent unit in the previous time step is used. In the spatial attention mechanism, the query is the i-th input sample, and the p-th data set in the i-th input sample is used as the key and value. The similarity between the query and the key is calculated as follows:

[0049]

[0050]

[0051] Where V1, W1, and b1 are learnable parameters in the spatial attention mechanism, tanh is the hyperbolic tangent function, and exp is an exponential function with the natural constant e as its base. The weight vectors of different sensors for the p-th group of data in the i-th input sample can be obtained after normalization.

[0052] Step 3.6, With x i+p-1 Performing a dot product yields a weighted vector. Will The input is fed into the gated recurrent unit, and the hidden state vector of the gated recurrent unit at the p-th time step can be obtained through calculation. p = p + 1. When 1 ≤ p ≤ p, proceed to step 3.5; otherwise, proceed to step 3.7.

[0053] Step 3.7: Repeat steps 3.5 and 3.6 to obtain the set of hidden state vectors obtained by the encoder for the i-th input sample.

[0054] Step 3.8: The decoder structure is a gated recurrent unit network based on a temporal attention mechanism. The temporal attention mechanism is used to calculate the relative importance of the hidden state vectors at different time steps. Assuming the decoder's step size is Q, the hidden state vector obtained by passing the i-th input sample through the encoder... The input is fed into the decoder, and the calculation process is as described in steps 3.9-3.10;

[0055] Step 3.9: When time step q = 1, initialize the hidden state vector of the gated loop unit as follows: Will The query is used as the hidden state vector of the i-th input sample in the time attention mechanism; when q>1, the hidden state vector of the gated recurrent unit of the previous time step is used. In the time attention mechanism, the query is the hidden state vector of the i-th input sample, and the hidden state vector of the i-th input sample is used as the key and values. The similarity between the query and the key is calculated as follows:

[0056]

[0057]

[0058] Where V2, W2, and b2 are learnable parameters in the time attention mechanism, tanh is the hyperbolic tangent function, and exp is an exponential function with the natural constant e as its base. The weights of all hidden state vectors at time step q in the i-th input sample can be obtained after normalization.

[0059] Step 3.10, and Performing a dot product yields a weighted vector. Will The input is fed into the gated recurrent unit, and the hidden state vector of the gated recurrent unit at the q-th time step can be obtained through calculation. q = q + 1. When 1 ≤ q ≤ Q, proceed to step 3.9; otherwise, proceed to step 3.11.

[0060] Step 3.11: The classifier is a fully connected classification layer; initialize the network parameters. Use the hidden state vector obtained in Step 3.10. The input is fed into a fully connected classification layer, and after passing through a softmax layer, the output result of the classifier for the i-th input sample can be obtained. The calculation method is as follows:

[0061]

[0062]

[0063] Here, W3 and b3 are the learnable parameters in the fully connected layer. Given the probability of the power transformer state of the i-th input sample, select the category with the highest probability as the classification result of the power transformer;

[0064] Step 3.12: Select cross-entropy as the loss function for training the model, and use gradient descent to find the optimal network parameters. The calculation method for the cross-entropy loss function is as follows:

[0065]

[0066] Where C represents the number of healthy states of the power transformer. To determine whether the power transformer of the i-th input sample belongs to the j-th category, if so, then... otherwise, Let be the probability that the power transformer of the i-th input sample belongs to the j-th category. Calculate the loss value for all samples and update the parameters in the network using gradient descent.

[0067] Compared with the prior art, the above embodiments have the following advantages:

[0068] 1. Applying digital twin technology to the health status assessment of power transformers: By simulating the real-time operation of power transformers through digital twin technology, and through virtual-real interaction, data fusion and other means, rich operating data can be provided for physical power transformers, which can more comprehensively reflect the operating status of power transformers.

[0069] 2. A spatiotemporal attention mechanism network is proposed to distinguish the relative importance of different sensors in the health status of power transformers and to consider the different degrees of temporal dependence in historical data, so as to improve the accuracy of power transformer health status assessment.

[0070] 3. The proposed method is applicable to the health status assessment of power transformers in various scenarios, and can quickly and accurately assess the status of power transformers, thus having high application value.

Claims

1. A method for assessing the health status of a power transformer, characterized in that, The method includes: Step 1: Obtain the operating status data of the power transformer, including: The percentage data of five dissolved gases (H2, CH4, C2H6, C2H4, and C2H2) in the insulating oil, along with voltage, current, temperature, and oil level operating status data, will be stored in a database. Step 2: Establish a digital twin model of the power transformer, including: geometric model, physical model, behavioral model, and rule model; among which: Geometric model: Based on the shape, size and positional relationship of each component and shell of the physical power transformer, the winding method of the primary and secondary windings, the assembly and insertion relationship of the leads and iron core, and the bushing assembly structure parameters, a 3D model of the power transformer with three dimensions of x-axis, y-axis and z-axis is formed. Physical model: Based on the geometric model, it incorporates strain analysis test data of transformer coil windings and insulation under the action of impulse voltage and electric field force, as well as performance test data of transformer tank and various components under external force, such as compressive strength and damage resistance. Behavioral model: Based on the mapping relationship between the operating status data and health status of the power transformer entity, simulated operating status data of its changes over operating time is generated; Rule-based model: Construct a power transformer state assessment model based on a spatiotemporal attention mechanism network, including an encoder, decoder, and classifier; Step 3: Calculate whether there is a discrepancy between the operating status data of the actual power transformer and the operating status data of the virtual power transformer in the digital twin model. If there is a discrepancy, adjust the parameters of the digital twin model according to the actual operating status data. Repeat this step until the digital twin model of the power transformer matches the actual power transformer. Step 4: In the digital twin model, simulate the health degradation path of power transformers under different operating environments and loads, generate health degradation data of power transformers, and evaluate the health status of power transformers based on the health degradation data.

2. The method for assessing the health status of power transformers according to claim 1, characterized in that, The power transformer condition assessment model is constructed by training a neural network with known samples as input.

3. The method for assessing the health status of power transformers according to claim 1, characterized in that, The encoder's network structure is a gated recurrent unit network based on spatial attention mechanism, and the decoder's structure is a gated recurrent unit network based on temporal attention mechanism.

Images