Converter transformer turn-to-turn short circuit fault identification method based on deep learning

Abstract

The invention discloses a converter transformer turn-to-turn short circuit fault identification method based on deep learning. The method comprises the following steps: 1) synchronously acquiring historical voltage and current data of a network side and a valve side of a converter transformer in different working states; 2) preprocessing the historical voltage and current data, and constructing a data set by taking the preprocessed historical voltage and current data as input and the working state as output; 3) constructing a deep learning network model, and training the deep learning network model by using the data set to obtain a trained deep learning network model; 4) collecting real-time voltage and current data of the network side and the valve side of the converter transformer, and preprocessing the real-time voltage and current data to obtain preprocessed real-time voltage and current data; and 5) inputting the preprocessed real-time voltage and current data into the trained deep learning network model, and carrying out real-time fault identification on the converter transformer. According to the designed four-layer convolutional neural network structure, accurate identification of the network side / valve side slight short circuit is realized.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, specifically to a method for identifying inter-turn short-circuit faults in converter transformers based on deep learning. Background Technology

[0002] Converter transformers are core equipment in high-voltage direct current (HVDC) transmission systems, undertaking crucial functions such as AC-DC power conversion, voltage matching, and isolation. Their operating status directly determines the stability and power supply reliability of the HVDC system. Compared to conventional AC transformers, converter transformers operate under more complex conditions: they are subjected to special loads such as DC bias, harmonic impacts (3rd / 5th / 7th harmonics generated by commutation), and non-sinusoidal voltage on the valve-side windings. This significantly increases the risk of inter-turn short-circuit faults, accounting for 65%-75% of all converter transformer faults, making it one of the main hidden dangers causing system outages.

[0003] Existing converter transformer inter-turn short circuit diagnosis technology has significant limitations:

[0004] Traditional methods have poor adaptability: conventional oil chromatography analysis and DC resistance testing require shutdown and disassembly, which not only leads to the shutdown of the HVDC system (a single shutdown can result in losses of over ten million yuan), but also cannot identify minor inter-turn short circuits (1%-5% of the turns are short-circuited); the Sweep Frequency Response Analysis (SFRA) method is affected by the complex winding structure of the converter transformer (multiple windings on the grid side / valve side, and delta connection on the extended side), and the fault location accuracy is less than 70%.

[0005] The circuit model method has low versatility: converter transformers need to consider special parameters such as DC bias and harmonic losses, making it extremely difficult to establish an accurate equivalent model. Traditional identification methods based on circuit models are limited by equipment models and wiring methods, and cannot be adapted to converter transformers of different voltage levels (±500kV / ±800kV).

[0006] Minor fault identification is difficult: The voltage waveform of the valve side winding of the converter transformer is not sinusoidal. After the addition of harmonic components, the changes in port electrical quantities (voltage distortion rate <0.8%, current fluctuation <3%) caused by minor inter-turn short circuits are easily masked by harmonic noise. The measurement error of existing current transformers (0.2-0.5 class) further exacerbates the difficulty of feature extraction, resulting in a fault misdiagnosis rate of over 40%.

[0007] The protection system has a slow response: the differential protection of the converter transformer is affected by the unbalanced current caused by DC bias, and a high action threshold needs to be set. It cannot respond to minor inter-turn short circuits (short circuit current increase <10%), making it difficult to achieve fault early warning.

[0008] Therefore, there is an urgent need for a fault identification technology that can adapt to the complex operating conditions of converter transformers, requires no shutdown, and can accurately identify minor inter-turn short circuits, so as to ensure the safe and stable operation of HVDC systems. Summary of the Invention

[0009] The purpose of this invention is to provide a deep learning-based method for identifying inter-turn short-circuit faults in converter transformers, comprising the following steps:

[0010] 1) Synchronously acquire historical voltage and current data of the converter transformer grid side and valve side under different operating conditions.

[0011] 2) Preprocess the historical voltage and current data, and use the preprocessed historical voltage and current data as input and the working status as output to construct a dataset.

[0012] 3) Construct a deep learning network model and train it using the dataset to obtain a trained deep learning network model.

[0013] 4) Collect real-time voltage and current data from the grid side and valve side of the converter transformer, and preprocess the real-time voltage and current data to obtain preprocessed real-time voltage and current data.

[0014] 5) Input the preprocessed real-time voltage and current data into the trained deep learning network model to perform real-time fault identification of the converter transformer.

[0015] Furthermore, the working status includes normal working status and fault status.

[0016] Furthermore, the fault states include inter-turn short-circuit faults with different numbers of short-circuit turns.

[0017] Furthermore, the historical voltage and current data are acquired through the TA / TV sensor of the HVDC system.

[0018] Furthermore, the preprocessing includes harmonic filtering, normalization, and window truncation.

[0019] Furthermore, the harmonic filtering method includes wavelet transform.

[0020] Furthermore, the normalization process is as follows:

[0021] (1)

[0022] In the formula, This represents the data after normalization. This represents the original data. , These represent the maximum and minimum values ​​of the original data, respectively.

[0023] Furthermore, the deep learning network model includes an input layer, a convolutional pooling layer, an attention fully connected layer, and a dual output layer.

[0024] The input layer is used to input voltage and current data.

[0025] The convolutional pooling layer includes two convolutional layers and two pooling layers.

[0026] The first convolutional layer is used to extract local features from the voltage and current data to obtain a local feature map.

[0027] The first pooling layer is used to compress the local feature map to obtain the key feature map.

[0028] The second convolutional layer is used to extract features from the key feature map, resulting in a secondary extracted feature map.

[0029] The second pooling layer is used to compress the secondary extracted feature map to obtain a secondary compressed feature map.

[0030] The attention-based fully connected layer comprises three fully connected layers, which are used to map the secondary compressed feature map to a low-dimensional space, learn the correlation between features, and identify the fault state of the converter transformer.

[0031] The dual output layer is used to output the fault status of the converter transformer.

[0032] Furthermore, the deep learning network model uses the Leaky ReLU activation function, as shown below:

[0033] (2)

[0034] In the formula, x is the input to the activation function. The leakage coefficient is in the negative range. This is the output of the activation function.

[0035] Furthermore, the deep learning network model uses the Adam optimization algorithm.

[0036] The technical effect of this invention is beyond doubt. This invention integrates the electrical characteristics (time domain + frequency domain) of the grid-side / valve-side dual windings with the converter operating parameters (firing angle, DC bias current), and uses wavelet transform harmonic filtering + harmonic suppression normalization processing to extract samples with two converter cycles as a sliding window. This solves the problem of difficult fault feature extraction under non-sinusoidal signals and is suitable for converter transformers of different voltage levels.

[0037] This invention designs a 4-layer Convolutional Neural Network (CNN) structure consisting of "large kernel denoising (5×5) + small kernel feature extraction (3×3) + attention mechanism". It selects the Leaky ReLU activation function to strengthen the weight of valve-side fault features, thereby achieving accurate identification of minor short circuits on the network side / valve side without the need to build a complex equivalent model.

[0038] This invention designs a multi-level anti-interference scheme for DC bias, firing angle fluctuation, and transformer error. Through sample expansion and feature correction, it improves the identification stability under complex operating conditions. Based on the existing monitoring equipment data of the HVDC system, it realizes the integrated identification of fault type, short circuit location and severity, and fills the gap of differential protection response lag.

[0039] The beneficial effects of this invention include:

[0040] 1. High identification accuracy: The accuracy rate for identifying minor inter-turn short circuits (2%-10% of turns) in the grid-side / valve-side windings of ±500kV / ±800kV converter transformers reaches 92%-96%, and it can distinguish the short circuit location and the number of short-circuited turns. The test set accuracy is 94.5%.

[0041] 2. Strong adaptability to operating conditions: Under conditions of DC bias (0-5% of rated current), firing angle fluctuation (15°-30°), and transformer error (0.2-0.5 level), the identification accuracy remains above 92%, adapting to the actual operating scenarios of HVDC systems;

[0042] 3. Excellent economic performance: Based on existing current transformer / voltage transformer (TA / TV) data, there is no need to shut down for disassembly or install sensors, avoiding system downtime losses of over ten million yuan and reducing operation and maintenance costs by more than 70%.

[0043] 4. Good real-time performance: The total time for data preprocessing + network inference is ≤45ms, which meets the millisecond-level response requirements of HVDC system and can provide early warning of minor faults to prevent the fault from escalating. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of fault diagnosis based on deep learning.

[0045] Figure 2 A schematic diagram of the data collection and population process. Detailed Implementation

[0046] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0047] Example 1:

[0048] See Figures 1 to 2 A deep learning-based method for identifying inter-turn short-circuit faults in converter transformers includes the following steps:

[0049] 1) Synchronously acquire historical voltage and current data of the converter transformer grid side and valve side under different operating conditions.

[0050] 2) Preprocess the historical voltage and current data, and use the preprocessed historical voltage and current data as input and the working status as output to construct a dataset.

[0051] 3) Construct a deep learning network model and train it using the dataset to obtain a trained deep learning network model.

[0052] 4) Collect real-time voltage and current data from the grid side and valve side of the converter transformer, and preprocess the real-time voltage and current data to obtain preprocessed real-time voltage and current data.

[0053] 5) Input the preprocessed real-time voltage and current data into the trained deep learning network model to perform real-time fault identification of the converter transformer.

[0054] Example 2:

[0055] A method for identifying inter-turn short-circuit faults in converter transformers based on deep learning is described in Embodiment 1. Further, the working state includes normal working state and fault state.

[0056] Example 3:

[0057] A deep learning-based method for identifying inter-turn short-circuit faults in converter transformers is provided. The main technical contents are described in any one of Embodiments 1 and 2. Furthermore, the fault states include inter-turn short-circuit faults with different numbers of short-circuit turns.

[0058] Example 4:

[0059] A deep learning-based method for identifying inter-turn short-circuit faults in converter transformers is described in any one of Embodiments 1 to 3. Furthermore, the historical voltage and current data are acquired through the TA / TV sensor of the HVDC system.

[0060] Example 5:

[0061] A method for identifying inter-turn short-circuit faults in converter transformers based on deep learning is described in any one of Examples 1 to 4. Further, the preprocessing includes harmonic filtering, normalization processing, and window truncation.

[0062] Example 6:

[0063] A method for identifying inter-turn short-circuit faults in converter transformers based on deep learning is described in any one of Examples 1 to 5. Furthermore, the harmonic filtering method includes wavelet transform.

[0064] Example 7:

[0065] A deep learning-based method for identifying inter-turn short-circuit faults in converter transformers, the main technical contents of which are described in any one of Embodiments 1 to 6, further wherein the normalization process is as follows:

[0066] (1)

[0067] In the formula, This represents the data after normalization. This represents the original data. , These represent the maximum and minimum values ​​of the original data, respectively.

[0068] Example 8:

[0069] A method for identifying inter-turn short-circuit faults in converter transformers based on deep learning is described in any one of Examples 1 to 7. Further, the deep learning network model includes an input layer, a convolutional pooling layer, an attention fully connected layer, and a dual output layer.

[0070] The input layer is used to input voltage and current data.

[0071] The convolutional pooling layer includes two convolutional layers and two pooling layers.

[0072] The first convolutional layer is used to extract local features from the voltage and current data to obtain a local feature map.

[0073] The first pooling layer is used to compress the local feature map to obtain the key feature map.

[0074] The second convolutional layer is used to extract features from the key feature map, resulting in a secondary extracted feature map.

[0075] The second pooling layer is used to compress the secondary extracted feature map to obtain a secondary compressed feature map.

[0076] The attention-based fully connected layer comprises three fully connected layers, which are used to map the secondary compressed feature map to a low-dimensional space, learn the correlation between features, and identify the fault state of the converter transformer.

[0077] The dual output layer is used to output the fault status of the converter transformer.

[0078] Example 9:

[0079] A deep learning-based method for identifying inter-turn short-circuit faults in converter transformers is described in any one of Examples 1 to 8. Further, the deep learning network model uses the Leaky ReLU activation function, as shown below:

[0080] (2)

[0081] In the formula, x is the input to the activation function (the weighted sum of neurons + bias). The leakage coefficient in the negative range is usually a small constant (such as 0.01 or 0.001) to control the degree of signal attenuation in the negative input range. This is the output of the activation function.

[0082] LeakyReLU is an improved version of the ReLU activation function that avoids the problem of neuron "death" by introducing a non-zero slope in the negative interval.

[0083] Example 10:

[0084] A method for identifying inter-turn short-circuit faults in converter transformers based on deep learning is described in any one of Examples 1 to 9. Furthermore, the deep learning network model uses the Adam optimization algorithm.

[0085] The Adam optimization algorithm initializes the model parameters and first- and second-order moment estimates, iteratively calculates the current gradient and updates the moment estimates, and after bias correction, updates the parameters by combining the global learning rate and adaptive adjustment term until the termination condition is met.

[0086] Example 11:

[0087] See Figures 1 to 2 A deep learning-based method for identifying inter-turn short-circuit faults in converter transformers, the main technical contents of which include:

[0088] 1. Converter transformer-specific multi-dimensional dataset construction technology: Integrating the electrical characteristics (time domain + frequency domain) of the grid-side / valve-side dual windings with converter operating parameters (firing angle, DC bias current), and through wavelet transform harmonic filtering + harmonic suppression normalization processing, samples are extracted with two converter cycles as a sliding window to solve the problem of difficult fault feature extraction under non-sinusoidal signals, and adapt to converter transformers of different voltage levels;

[0089] 2. Harmonic Anti-interference Convolutional Neural Network Architecture: A 4-layer CNN structure of "large kernel denoising (5×5) + small kernel feature extraction (3×3) + attention mechanism" is designed. The Leaky ReLU activation function is selected to strengthen the weight of valve-side fault features, so as to achieve accurate identification of slight short circuits on the network side / valve side without the need to build a complex equivalent model.

[0090] 3. Robustness optimization mechanism for converter operation: A multi-level anti-interference scheme is designed to address DC bias, firing angle fluctuation, and transformer error. Through sample expansion and feature correction, the identification stability under complex operating conditions is improved.

[0091] 4. Non-intrusive uninterrupted power outage identification solution: Based on the existing monitoring equipment data of the HVDC system, it realizes integrated identification of fault type, short circuit location and severity, making up for the gap of differential protection response lag.

[0092] Example 12:

[0093] See Figures 1 to 2 A deep learning-based method for identifying inter-turn short-circuit faults in converter transformers, the main technical contents of which include:

[0094] 1. Equipment parameters: ±500kV converter transformer (grid-side voltage 220kV, valve-side voltage 50kV, rated capacity 315MVA, grid-side turns 1200, valve-side turns 300), TA / TV accuracy class 0.2;

[0095] 2. Data Acquisition: The existing TA / TV of the HVDC system is used to acquire grid-side and valve-side voltage and current data, and the trigger angle α=20° and DC bias current 3% of rated current are recorded simultaneously; the sampling frequency is 10kHz, each sample contains 400 data points (2 commutation cycles), covering normal operation and valve-side winding 2%-8% turn short circuit conditions;

[0096] 3. Dataset Construction: Wavelet transform was applied to the collected data to filter out the 5th harmonic, and harmonic suppression normalization was used to eliminate the order-of-magnitude difference. The method is shown in formula (1):

[0097] (1)

[0098] Construct a dataset (300 512×512 grayscale images for each working condition), and divide it into training, validation, and test sets in a 6:2:2 ratio;

[0099] 4. Network Training: A 4-layer CNN was built with convolutional kernels configured as 16 / 32 / 64 / 128 (5×5 for the first 2 layers and 3×3 for the last 2 layers). The Leaky ReLU activation function and Adam optimization algorithm were used (initial learning rate 0.001, batch size 64). After 45 iterations, the training set accuracy reached 96%.

[0100] 5. Identification test: After preprocessing the 4% short-circuit data of the valve side, the network is input and the output "valve side winding inter-turn short circuit, short circuit turns 4%" is output within 42ms, with an accuracy of 94.8%.

Claims

1. A method for identifying inter-turn short-circuit faults in converter transformers based on deep learning, characterized in that, Includes the following steps: 1) Synchronously acquire historical voltage and current data of the converter transformer grid side and valve side under different operating conditions; 2) Preprocess the historical voltage and current data, and use the preprocessed historical voltage and current data as input and the working status as output to construct a dataset; 3) Construct a deep learning network model and train it using the dataset to obtain a trained deep learning network model; 4) Collect real-time voltage and current data from the grid side and valve side of the converter transformer, and preprocess the real-time voltage and current data to obtain preprocessed real-time voltage and current data. 5) Input the preprocessed real-time voltage and current data into the trained deep learning network model to perform real-time fault identification of the converter transformer.

2. The method for identifying inter-turn short-circuit faults in converter transformers based on deep learning according to claim 1, characterized in that, The operating status includes normal operating status and fault status.

3. The method for identifying inter-turn short-circuit faults in converter transformers based on deep learning according to claim 2, characterized in that, The fault states include inter-turn short-circuit faults with different numbers of short-circuit turns.

4. The method for identifying inter-turn short-circuit faults in converter transformers based on deep learning according to claim 1, characterized in that, The historical voltage and current data were acquired through the TA / TV sensor of the HVDC system.

5. The method for identifying inter-turn short-circuit faults in converter transformers based on deep learning according to claim 1, characterized in that, The preprocessing includes harmonic filtering, normalization, and window truncation.

6. The method for identifying inter-turn short-circuit faults in converter transformers based on deep learning according to claim 5, characterized in that, The harmonic filtering method includes wavelet transform.

7. The method for identifying inter-turn short-circuit faults in converter transformers based on deep learning according to claim 5, characterized in that, The normalization process is as follows: （1） In the formula, This represents the data after normalization. Represents the original data; , These represent the maximum and minimum values ​​of the original data, respectively.

8. The method for identifying inter-turn short-circuit faults in converter transformers based on deep learning according to claim 1, characterized in that, The deep learning network model includes an input layer, a convolutional pooling layer, an attention fully connected layer, and a dual output layer. The input layer is used to input voltage and current data; The convolutional pooling layer includes two convolutional layers and two pooling layers; The first convolutional layer is used to extract local features from the voltage and current data to obtain a local feature map; The first pooling layer is used to compress the local feature map to obtain the key feature map; The second convolutional layer is used to extract features from the key feature map, resulting in a secondary extracted feature map; The second pooling layer is used to compress the secondary extracted feature map to obtain a secondary compressed feature map; The attention-based fully connected layer includes three fully connected layers, which are used to map the secondary compressed feature map to a low-dimensional space, learn the correlation between features, and identify the fault state of the converter transformer. The dual output layer is used to output the fault status of the converter transformer.

9. The method for identifying inter-turn short-circuit faults in converter transformers based on deep learning according to claim 1, characterized in that, The deep learning network model uses the Leaky ReLU activation function, as shown below: （2） In the formula, x is the input to the activation function; The leakage coefficient is in the negative range; This is the output of the activation function.

10. The method for identifying inter-turn short-circuit faults in converter transformers based on deep learning according to claim 1, characterized in that, The deep learning network model uses the Adam optimization algorithm.

Images