Transformer winding state detection method based on temperature characteristics

Abstract

The invention provides a transformer winding state detection method based on temperature characteristics. The method comprises the following steps: (1) collecting operation data; (2) preprocessing the operation data; (3) taking the normalized transformer load current, the normalized active power, the normalized top oil temperature, the normalized ambient temperature and the normalized oil flow speed as input data, taking the normalized winding hot-spot temperature as output data, and constructing a data set; (4) establishing a VSN-TKAN-GRN deep learning network parallel computing model of the transformer winding hot-spot temperature considering the multi-head attention mechanism; (5) predicting the hot-spot temperature of the transformer winding based on a VSN-TKAN-GRN deep learning network parallel computing model; and (6) calculating the average relative error percentage of the transformer winding hot-spot temperature prediction result and the actual measurement result, and judging the winding state according to the change of the average relative error percentage. According to the method, the state of the transformer winding in operation can be efficiently and accurately judged, so that effective transformer operation and maintenance measures can be taken in time, and large faults are avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer operation monitoring, and in particular to a transformer winding state detection method based on temperature characteristics. Background Art

[0002] Power transformers are essential and critical components of power grids, playing a crucial role in power transmission and voltage conversion. In the context of developing new power systems, the integration of high proportions of renewable energy increases the randomness and uncertainty of power system operation, subjecting power equipment to more extreme and volatile operating conditions, significantly increasing the difficulty of equipment operation and maintenance. Power transformers are complex structures, subject to the combined influence of external operating environments and internal electromagnetic and thermal factors, making faults such as overheating and turn-to-turn short circuits unavoidable. If early faults in transformer windings are not detected promptly, they can develop into catastrophic failures.

[0003] Hotspot temperature is a crucial parameter reflecting the operating status of a winding. During transformer operation, the hotspot temperature should not exceed a certain value determined by the insulation grade. Otherwise, the insulation life of the transformer will be shortened, and in severe cases, the transformer may be permanently damaged. Furthermore, interturn faults are a common type of thermal fault in windings, primarily caused by aging and deterioration of the interturn insulation. During transformer operation, overload, poor winding heat dissipation, and other factors can cause localized overheating of the winding. When the hotspot temperature exceeds the specified value, the performance of the interturn insulation material degrades. If the transformer is overloaded for an extended period of time or if poor winding heat dissipation is not promptly addressed, the interturn insulation will further degrade, resulting in a decrease in interturn insulation resistance, an increase in interturn leakage current and faulty coil current, and ultimately, an interturn fault. Timely and effective monitoring of transformer winding temperature characteristics, especially hotspot temperature, is crucial for safe and reliable operation. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a transformer winding status detection method based on temperature characteristics. The method can accurately determine the transformer winding status by real-time monitoring of the transformer's load current, active power, top oil temperature, ambient temperature, and oil flow rate.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A method for detecting a transformer winding state based on temperature characteristics, the method comprising: Step 1: Collect transformer data in operation. The sampling frequency is , the length is ; Step 2: Preprocess the collected transformer data; Step 3: Use the preprocessed transformer data as input data and the normalized winding hotspot temperature data as output data to construct a data set; Step 4. Establish a VSN-TKAN-GRN deep learning network parallel computing model for transformer winding hot spot temperature considering the multi-head attention mechanism, and use the Adam optimizer to optimize the weight parameters until the optimal parameter combination of the network is found; Step 5: Select the transformer load current, active power, top oil temperature, ambient temperature, and oil flow rate within a period of time after the input data as the input of the VSN-TKAN-GRN deep learning network parallel computing model to predict the transformer winding hotspot temperature; Step 6. Calculate the average relative error percentage between the transformer winding hot spot temperature prediction results and the measured results The winding status is judged according to the change in the average relative error percentage. If the average relative error percentage is less than 3%, it is judged that the transformer winding is operating normally; if the average relative error percentage is less than or greater than 3%, it is judged that the transformer winding has changed and there may be a turn-to-turn short circuit fault.

[0006] The above Step 1 specifically includes: The voltage signal, current signal, top oil temperature, ambient temperature, oil flow rate and winding hot spot temperature of the transformer in operation are collected. The voltage signal and current signal are obtained through the voltage transformer and current transformer respectively. The sampling frequency is , the length is The top oil temperature and ambient temperature are obtained by temperature sensors; the oil flow velocity and winding hot spot temperature are obtained by optical fiber oil flow sensor and optical fiber temperature sensor respectively.

[0007] The above Step 2 specifically includes: The collected transformer voltage signal, current signal, top oil temperature, ambient temperature, oil flow velocity and winding hot spot temperature data are preprocessed, including the calculation of the effective value of the transformer voltage signal, current signal, and active power, as well as the calculation of the moving average and moving standard deviation of the top oil temperature, ambient temperature, oil flow velocity and winding hot spot temperature. The transformer's load current, active power, top oil temperature after removing abnormal values, ambient temperature, oil flow velocity and winding hot spot temperature are normalized.

[0008] The above Step 3 specifically includes: The preprocessed transformer load current, active power, top oil temperature, ambient temperature and oil flow velocity are used as input data, and the normalized winding hot spot temperature data are used as output data to construct the data set.

[0009] The VSN-TKAN-GRN network, based on the latest deep learning research, is designed to address the complex problem of transformer winding hotspot temperature prediction. This network architecture combines the advanced features of Variable Selection Networks, Temporal Kolmogorov-Arnold Networks, and Gated Residual Networks to improve the model's prediction accuracy and efficiency. The VSN selects the most useful features for the prediction task; the TKAN processes time series data and captures temporal dependencies; and the GRN introduces nonlinear transformations and gating mechanisms to enhance the model's expressive power. The establishment of the VSN-TKAN-GRN network is now explained. The network consists of three main modules and other optimization modules: (1) Variable Selection Network: Generate variable selection weights through the gated residual network and the Softmax layer to dynamically select the most relevant variables in the input data.

[0010] (2) Temporal Kolmogorov-Arnold Network: As the core module of the network, TKAN captures long-range dependencies in time series data by introducing temporal dependencies and memory management mechanisms. The TKAN layer manages information flow through recursive mechanisms and gating units.

[0011] (3) Gated Residual Network: Used to control information flow, handle nonlinear interactions and long-range dependencies. GRN enhances the flexibility and expressiveness of the model through gated linear units and residual connections.

[0012] (4) Parallel computing model: The VSN-TKAN-GRN network adopts a parallel computing architecture and captures long-range dependencies between different time steps through a multi-head attention mechanism. The multi-head attention mechanism allows the model to focus on information from multiple time steps simultaneously, thereby improving prediction accuracy.

[0013] (5) Optimizer: Use the Adam optimizer to optimize the weight parameters. The Adam optimizer combines the advantages of the momentum method and the adaptive learning rate, and can efficiently find the optimal parameter combination of the network.

[0014] (6) Training process: The network optimizes the model parameters by minimizing the prediction error until the optimal parameter combination is found. During the training process, early stopping and learning rate decay strategies are used to prevent overfitting and improve the generalization ability of the model.

[0015] The deep learning network model in the above Step 4 consists of an input layer, a VSN module, a TKAN module, a GRN module, a self-attention layer, a fully connected layer, and an output layer. The VSN module is used to perform nonlinear processing on the input data and combine variables according to weights. The TKAN module is used to capture the long-term dependencies of the data processed by the VSN module, and at the same time uses the output of the encoder as the input of the decoder to generate a predicted time series based on the GRN module. The self-attention layer uses a multi-head attention mechanism to capture the long-term relationship between different time steps, and is linked to the fully connected layer to transform the output of the self-attention layer into the final output result.

[0016] The preprocessing in Step 2 above specifically includes: Step 2.1. Calculate the effective value of the transformer voltage signal, current signal, and active power respectively. The corresponding calculation formula is: ; ; ; Where: and Represent the effective values of voltage signal and current signal respectively; Indicates power factor; Indicates the phase angle of voltage and current signals; Step 2.2. Calculate the moving average and moving standard deviation of the top oil temperature, ambient temperature, oil flow velocity, and winding hotspot temperature, respectively. Set thresholds to identify abnormal values in the top oil temperature, ambient temperature, and oil flow velocity. The calculation formulas for the moving average and moving standard deviation of the top oil temperature are:

[0017]

[0018] ; Where, Indicates time Moving average of Indicates time The moving standard deviation of It's time The data value of Indicates the number of data points; Represents the transformer top oil temperature data; the moving average and moving standard deviation of the ambient temperature, oil flow velocity and winding hot spot temperature are calculated according to the formula of the top oil temperature; Represents the normalized top oil temperature value, indicating the difference between the current top oil temperature value and the moving average; the formula is normalized by subtracting the moving average and dividing by the moving standard deviation , so that data at different time points can be compared on the same scale; by calculating Values can identify abnormal data points that deviate significantly from the normal range; Step 2.3. Normalize the transformer's load current, active power, top oil temperature after removing abnormal values, ambient temperature, oil flow velocity, and winding hot spot temperature. The normalized calculation formula for the transformer's load current is: ; Where: is the normalized effective value of the transformer load current; and They represent the minimum and maximum effective values of the transformer load current respectively; the active power, top oil temperature after removing abnormal values, ambient temperature, oil flow velocity and normalized winding hot spot temperature are calculated according to the load current formula.

[0019] The specific steps for using the Adam optimizer to optimize weight parameters in Step 4 above include: Step 4.1, Initialize parameters; Before starting training, the following parameters need to be initialized: learning rate (usually set to 0.001); decay rate of the first-order moment estimate (usually set to 0.9); decay rate of the second-order moment estimate (usually set to 0.999); a small constant to prevent division by zero errors (Usually set to 10 −8 ); Initialize the first-order moment estimate ; Initialize the second-order moment estimate ; Initialize time step ; Step 4.2. Calculate the gradient of the loss function with respect to the parameters. At each time step , calculate the loss function About parameters Gradient: ; Step 4.3, update the first-order moment estimate; Use the current gradient , update the first-order moment estimate : ; Step 4.4, update the second-order moment estimate; Use the current gradient , update the second-order moment estimate : ; Step 4.5, correct the first-order moment estimate and the second-order moment estimate; because and It is zero at initialization, so there will be a deviation in the first few time steps. To correct this deviation, the following formula is used: Bias-corrected first moment: ; Where, is the bias-corrected first-order moment estimate; Bias-corrected second moment: ; Where, is the bias-corrected second-order moment estimate; Step 4.6, Update parameters: Update parameters based on the revised estimate; Use the corrected first moment and second-order moment Update parameters : ; Where, is the time step Parameter value when is the time step Parameter value when ; Step 4.7: After completing the parameter update, Adam enters the next time step and repeats the above steps. Each iteration calculates new gradients, updates the first- and second-order moment estimates, performs bias correction, and updates the parameters based on the new estimates. As training progresses, the parameter updates gradually approach the minimum value of the loss function. Step 4.8: After the optimization is completed, the network model with the optimal weight parameters is obtained.

[0020] The calculation formula for the average relative error percentage in Step 6 above is: ; Where, Indicates the predicted value of the hot spot temperature of the transformer winding; Indicates the measured value of the hot spot temperature of the transformer winding; The number of data points representing the hot spot temperature of the winding.

[0021] The present invention provides a method for detecting the state of a transformer winding based on temperature characteristics. A VSN-TKAN-GRN deep learning network calculation model is established based on the transformer's load current, active power, top oil temperature, ambient temperature, oil flow velocity, and winding hotspot temperature. The transformer winding state is judged according to the change in the average relative error percentage between the predicted and measured hotspot temperatures of the transformer winding. This judgment method is efficient, accurate, and easy to implement, making it easy for operators to promptly detect abnormal transformer winding conditions such as turn-to-turn short circuits.

[0022] The benefit of this invention lies in the proposed transformer winding state identification method, which utilizes a parallel VSN-TKAN-GRN deep learning network. This network accurately captures the key features inherent in the transformer's multi-source time series—load current, active power, top oil temperature, ambient temperature, and oil flow velocity—through a multi-head self-attention mechanism. This enhances the ability to capture long-term dependencies in the time series and improves prediction accuracy. Furthermore, the constructed parallel deep learning network boasts a flexible architecture, strong parallel computing and generalization capabilities, and robust resistance to overfitting, enabling computational analysis of transformer winding hotspot temperatures and accurate identification of changes in transformer winding states.

[0023] Since the transformer mentioned in the present invention adopts the above-mentioned technical solution, it can accurately monitor the transformer winding status through the transformer's load current, active power, top oil temperature, ambient temperature, and oil flow rate, and then take effective transformer operation and maintenance measures, greatly reducing the transformer's failure damage rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The present invention will be further described below with reference to the accompanying drawings and examples: Figure 1 Schematic diagram of the process of the present invention; Figure 2 This is the parallel VSN-TKAN-GRN deep learning network structure diagram of the present invention. DETAILED DESCRIPTION

[0025] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0026] Example 1: A method and system for detecting transformer winding status based on temperature characteristics, comprising the following steps: (1) Collect the voltage signal, current signal, top oil temperature, ambient temperature, oil flow velocity and winding hot spot temperature of the transformer in operation. The voltage signal and current signal are obtained through the voltage transformer and current transformer respectively. The sampling frequency is , the length is The top oil temperature and ambient temperature are acquired through temperature sensors; the oil flow velocity and the winding hot spot temperature are acquired through optical fiber oil flow sensors and optical fiber temperature sensors respectively; (2) Preprocess the collected transformer voltage signal, current signal, top oil temperature, ambient temperature, oil flow velocity and winding hot spot temperature. The specific steps are as follows: 2.1 Calculate the effective value of the transformer voltage signal, current signal, and active power respectively. The corresponding calculation formula is:

[0027]

[0028]

[0029] Where: and Represent the effective values of voltage signal and current signal respectively; Indicates power factor; Indicates the phase angle of voltage and current signals.

[0030] 2.2 Calculate the moving average and moving standard deviation of the top oil temperature, ambient temperature, oil flow velocity, and winding hot spot temperature respectively, and set thresholds to identify abnormal values in the top oil temperature, ambient temperature, and oil flow velocity. Taking the top oil temperature as an example, the calculation formulas for the moving average and moving standard deviation are:

[0031]

[0032]

[0033] Where, Indicates time Moving average of Indicates time The moving standard deviation of It's time The data value of Indicates the number of data points; Indicates the transformer top oil temperature data, Represents the normalized top oil temperature value, indicating the difference between the current top oil temperature value and the moving average; the formula is normalized by subtracting the moving average and dividing by the moving standard deviation , so that data at different time points can be compared on the same scale; by calculating values, can identify abnormal data points that deviate significantly from the normal range.

[0034] 2.3 Normalize the transformer's load current, active power, top oil temperature after removing abnormal values, ambient temperature, oil flow rate, and winding hot spot temperature. Take the transformer's load current as an example to illustrate. The normalized calculation formula for the load current is:

[0035] Where: is the normalized effective value of the transformer load current; and Respectively represent the minimum and maximum effective values of the transformer load current; (3) Using the normalized transformer load current, active power, top oil temperature, ambient temperature, and oil flow velocity as input data and the normalized winding hotspot temperature as output data, a data set was constructed; (4) A VSN-TKAN-GRN deep learning network parallel computing model for transformer winding hot spot temperature considering the multi-head attention mechanism is established, and the weight parameters are optimized using the Adam optimizer until the optimal parameter combination of the network is found. The deep learning network model mainly consists of an input layer, a VSN module, a TKAN module, a GRN module, a self-attention layer, a fully connected layer, and an output layer. Among them, the VSN module is used to perform nonlinear processing on the input data and combine variables according to the weights; the TKAN module is used to capture the long-term dependencies of the data processed by the VSN module, and at the same time, the output of the encoder is used as the input of the decoder to generate a predicted time series based on the GRN module; the self-attention layer uses a multi-head attention mechanism to capture the long-term relationship between different time steps, and is linked to the fully connected layer to transform the output of the self-attention layer into the final output result.

[0036] The Adam optimizer optimizes the parameters of the deep learning network calculation model. The specific calculation steps are as follows: 4.1、Initialization parameters; Before starting training, the following parameters need to be initialized: learning rate (usually set to 0.001); decay rate of the first-order moment estimate (usually set to 0.9); decay rate of the second-order moment estimate (usually set to 0.999); a small constant to prevent division by zero errors (Usually set to 10 −8 ); Initialize the first-order moment estimate ; Initialize the second-order moment estimate ; Initialize time step ; 4.2. Calculate the gradient of the loss function with respect to the parameters; At each time step , calculate the loss function About parameters Gradient: ; 4.3. Update the first-order moment estimate; Use the current gradient , update the first-order moment estimate : ; 4.4, Update the second-order moment estimate; Use the current gradient , update the second-order moment estimate : ; 4.5. Correct the first-order moment estimate and the second-order moment estimate; because and It is zero at initialization, so there will be a deviation in the first few time steps. To correct this deviation, the following formula is used: Bias-corrected first moment: ; Where, is the bias-corrected first-order moment estimate; Bias-corrected second moment: ; Where, is the bias-corrected second-order moment estimate; 4.6. Update parameters: Update parameters based on the revised estimate; Use the corrected first moment and second-order moment Update parameters : ; Where, is the time step Parameter value when is the time step Parameter value when ; 4.7. After completing the parameter update, Adam enters the next time step and repeats the above steps. Each iteration calculates new gradients, updates the first-order and second-order moment estimates, performs bias correction, and updates the parameters based on the new estimates. As training progresses, the parameter updates gradually approach the minimum value of the loss function. 4.8. After the optimization is completed, the network model with the optimal weight parameters is obtained.

[0037] The calculation formula for the average relative error percentage in Step 6 above is: ; Where, Indicates the predicted value of the hot spot temperature of the transformer winding; Indicates the measured value of the hot spot temperature of the transformer winding; The number of data points representing the hot spot temperature of the winding.

[0038] (5) The transformer load current, active power, top oil temperature, ambient temperature, and oil flow velocity within a period of time after the input data are selected as the input of the VSN-TKAN-GRN deep learning network parallel computing model to predict the hot spot temperature of the transformer winding; (6) Calculate the average relative error percentage between the transformer winding hot spot temperature prediction results and the measured results The winding status is determined based on the change in the average relative error percentage. If the average relative error percentage is less than 3%, the transformer winding is considered to be operating normally. If the average relative error percentage is less than or greater than 3%, the transformer winding is considered to have changed, and there may be a turn-to-turn short circuit fault. The calculation formula for the average relative error percentage is:

[0039] Where, Indicates the predicted value of the hot spot temperature of the transformer winding; Indicates the measured value of the hot spot temperature of the transformer winding; The number of data points representing the hot spot temperature of the winding.

[0040] In other words, this technical solution establishes a VSN-TKAN-GRN deep learning network calculation model based on the transformer's load current, active power, top oil temperature, ambient temperature, oil flow velocity and winding hotspot temperature, and judges the transformer winding status according to the change in the average relative error percentage between the transformer winding hotspot temperature prediction results and the measured results. This judgment method is efficient, accurate, and easy to implement, which facilitates operators to promptly detect whether there are abnormal transformer winding conditions such as inter-turn short circuits.

[0041] Example 2: like Figure 1 As shown in , a 110kV transformer is used as the object for condition monitoring. The winding status of the transformer is determined according to the following steps: (1) Collect the voltage signal, current signal, top oil temperature, ambient temperature, oil flow rate and winding hot spot temperature of the transformer in operation. The voltage signal and current signal are obtained through the voltage transformer and current transformer respectively, and the sampling frequency is , the length is The top oil temperature and ambient temperature are obtained by temperature sensors; the oil flow velocity and winding hot spot temperature are obtained by optical fiber oil flow sensor and optical fiber temperature sensor respectively; (2) Preprocess the collected transformer voltage signal, current signal, top oil temperature, ambient temperature, oil flow velocity and winding hot spot temperature. The specific steps are as follows: 2.1 Calculate the effective value of the transformer voltage signal, current signal, and active power respectively. The corresponding calculation formula is:

[0042]

[0043]

[0044] Where: and Represent the effective values of voltage signal and current signal respectively; Indicates power factor; Indicates the phase angle of voltage and current signals.

[0045] 2.2 Calculate the moving average and moving standard deviation of the top oil temperature, ambient temperature, oil flow velocity, and winding hot spot temperature respectively, and set thresholds to identify abnormal values in the top oil temperature, ambient temperature, and oil flow velocity. Taking the top oil temperature as an example, the calculation formulas for the moving average and moving standard deviation are: =

[0046] =

[0047]

[0048] Where, Indicates time Moving average of Indicates time The moving standard deviation of It's time The data value of Indicates the number of data points; Indicates the transformer top oil temperature data. Represents the normalized top oil temperature value, indicating the difference between the current top oil temperature value and the moving average; the formula is normalized by subtracting the moving average and dividing by the moving standard deviation , so that data at different time points can be compared on the same scale; by calculating values, can identify abnormal data points that deviate significantly from the normal range.

[0049] 2.3 Normalize the transformer's load current, active power, top oil temperature after removing abnormal values, ambient temperature, oil flow rate, and winding hot spot temperature. Take the transformer's load current as an example. The normalized calculation formula for the load current is:

[0050] Where: is the normalized effective value of the transformer load current; and Respectively represent the minimum and maximum effective values of the transformer load current; (3) Using the normalized transformer load current, active power, top oil temperature, ambient temperature, and oil flow velocity as input data and the normalized winding hotspot temperature as output data, a data set was constructed; (4) A VSN-TKAN-GRN deep learning network parallel computing model for transformer winding hot spot temperature considering the multi-head attention mechanism is established, and the weight parameters are optimized using the Adam optimizer until the optimal parameter combination of the network is found. The deep learning network model mainly consists of an input layer, a VSN module, a TKAN module, a GRN module, a self-attention layer, a fully connected layer, and an output layer. Among them, the VSN module is used to perform nonlinear processing on the input data and combine variables according to weights; the TKAN module is used to capture the long-term dependencies of the data processed by the VSN module, and at the same time, the output of the encoder is used as the input of the decoder to generate a predicted time series based on the GRN module; the self-attention layer uses a multi-head attention mechanism to capture the long-term relationship between different time steps, and is linked to the fully connected layer to transform the output of the self-attention layer into the final output result.

[0051] The Adam optimizer performs parameter optimization of the deep learning network calculation model. The specific calculation steps are as follows: 4.1 Initialization parameters; 4.2 Calculate the gradient of the loss function with respect to the parameters; 4.3 Updating the First-Order Moment Estimate 4.4 Updating the Second-Order Moment Estimate 4.5 Correct the first-order moment estimates and the second-order moment estimates.

[0052] 4.6 Update parameters: Update the parameters based on the revised data.

[0053] After completing the parameter update in 4.7, Adam proceeds to the next time step and repeats the above steps. Each iteration computes new gradients, updates the first- and second-order moment estimates, applies bias corrections, and updates the parameters based on the new estimates. As training progresses, the parameter updates gradually approach the minimum value of the loss function.

[0054] 4.8 After the optimization is completed, the network model with the optimal weight parameters is obtained.

[0055] (5) The transformer load current, active power, top oil temperature, ambient temperature, and oil flow velocity within a period of time after the input data are selected as the input of the VSN-TKAN-GRN deep learning network parallel computing model to predict the hot spot temperature of the transformer winding; (6) Calculate the average relative error percentage between the transformer winding hot spot temperature prediction results and the measured results The winding status is determined based on the change in the average relative error percentage. If the average relative error percentage is less than 3%, the transformer winding is considered to be operating normally. If the average relative error percentage is less than or greater than 3%, the transformer winding is considered to have changed, and there may be a turn-to-turn short circuit fault. The calculation formula for the average relative error percentage is:

[0056] Where, Indicates the predicted value of the hot spot temperature of the transformer winding; Indicates the measured value of the hot spot temperature of the transformer winding; The number of data points representing the hot spot temperature of the winding.

[0057] If the actual measured value is ℃, the predicted value is ℃, then MAPE is ; Here, the average relative error percentage between the transformer winding hotspot temperature prediction results and the measured results is 2.65%, indicating that the transformer winding status is normal.

Claims

1. A method for detecting transformer winding status based on temperature characteristics, characterized in that the method include: Step 1: Collect transformer data in operation. The sampling frequency is , the length is ; Step 2: Preprocess the collected transformer data; Step 3: Use the preprocessed transformer data as input data and the normalized winding hotspot temperature data as output data to construct a data set; Step 4. Establish a VSN-TKAN-GRN deep learning network parallel computing model for transformer winding hot spot temperature considering the multi-head attention mechanism, and use the Adam optimizer to optimize the weight parameters until the optimal parameter combination of the network is found; Step 5: Select the transformer load current, active power, top oil temperature, ambient temperature, and oil flow rate within a period of time after the input data as the input of the VSN-TKAN-GRN deep learning network parallel computing model to predict the transformer winding hotspot temperature; Step 6. Calculate the average relative error percentage between the transformer winding hot spot temperature prediction results and the measured results , judge the winding status according to the change of the average relative error percentage. If the average relative error percentage is less than 3%, it is judged that the transformer winding operation status is normal; If the average relative error percentage is less than or greater than 3%, it is determined that the transformer winding has changed and there may be a turn-to-turn short circuit fault.

2. The method for detecting transformer winding status based on temperature characteristics according to claim 1, characterized in that: The Step 1 specifically includes: The voltage signal, current signal, top oil temperature, ambient temperature, oil flow rate and winding hot spot temperature of the transformer in operation are collected. The voltage signal and current signal are obtained through the voltage transformer and current transformer respectively. The sampling frequency is , the length is The top oil temperature and ambient temperature are obtained by temperature sensors; the oil flow velocity and winding hot spot temperature are obtained by optical fiber oil flow sensor and optical fiber temperature sensor respectively.

3. The method for detecting transformer winding status based on temperature characteristics according to claim 2, characterized in that: The Step 2 specifically includes: The collected transformer voltage signal, current signal, top oil temperature, ambient temperature, oil flow velocity and winding hot spot temperature data are preprocessed, including the calculation of the effective value of the transformer voltage signal, current signal, and active power, as well as the calculation of the moving average and moving standard deviation of the top oil temperature, ambient temperature, oil flow velocity and winding hot spot temperature. The transformer's load current, active power, top oil temperature after removing abnormal values, ambient temperature, oil flow velocity and winding hot spot temperature are normalized.

4. The method for detecting transformer winding status based on temperature characteristics according to claim 3, characterized in that: The Step 3 specifically includes: The preprocessed transformer load current, active power, top oil temperature, ambient temperature and oil flow velocity are used as input data, and the normalized winding hot spot temperature data are used as output data to construct the data set.

5. The method for detecting transformer winding status based on temperature characteristics according to claim 4, characterized in that: The deep learning network model in the Step 4 is composed of an input layer, a VSN module, a TKAN module, a GRN module, a self-attention layer, a fully connected layer and an output layer, wherein the VSN module is used to perform nonlinear processing on the input data and combine variables according to weights; the TKAN module is used to capture the dependency of the data processed by the VSN module, and at the same time uses the output of the encoder as the input of the decoder to generate a predicted time series based on the GRN module; the self-attention layer uses a multi-head attention mechanism to capture the relationship between different time steps, and is linked to the fully connected layer to transform the output of the self-attention layer into the final output result.

6. The method for detecting transformer winding status based on temperature characteristics according to claim 5, characterized in that: The preprocessing in Step 2 specifically includes: Step 2.

1. Calculate the effective value of the transformer voltage signal, current signal, and active power respectively. The corresponding calculation formula is: ； ； ； Where: and Represent the effective values of voltage signal and current signal respectively; Indicates power factor; Indicates the phase angle of voltage and current signals; Step 2.

2. Calculate the moving average and moving standard deviation of the top oil temperature, ambient temperature, oil flow velocity, and winding hotspot temperature, respectively. Set thresholds to identify abnormal values in the top oil temperature, ambient temperature, and oil flow velocity. The calculation formulas for the moving average and moving standard deviation of the top oil temperature are: ； Where, Indicates time Moving average of Indicates time The moving standard deviation of It's time The data value of Indicates the number of data points; Represents the transformer top oil temperature data; the moving average and moving standard deviation of the ambient temperature, oil flow velocity and winding hot spot temperature are calculated according to the formula of the top oil temperature; Represents the normalized top oil temperature value, indicating the difference between the current top oil temperature value and the moving average; the formula is normalized by subtracting the moving average and dividing by the moving standard deviation , so that data at different time points can be compared on the same scale; by calculating Values can identify abnormal data points that deviate significantly from the normal range; Step 2.

3. Normalize the transformer's load current, active power, top oil temperature after removing abnormal values, ambient temperature, oil flow velocity, and winding hot spot temperature. The normalized calculation formula for the transformer's load current is: ； Where: is the normalized effective value of the transformer load current; and They represent the minimum and maximum effective values of the transformer load current respectively; the active power, top oil temperature after removing abnormal values, ambient temperature, oil flow velocity and normalized winding hot spot temperature are calculated according to the load current formula.

7. The method for detecting transformer winding status based on temperature characteristics according to claim 6, characterized in that: The specific steps of using the Adam optimizer to optimize the weight parameters in Step 4 include: Step 4.1, Initialize parameters; Step 4.

2. Calculate the gradient of the loss function with respect to the parameters. Step 4.3, update the first-order moment estimate; Step 4.4, update the second-order moment estimate; Step 4.5, correct the first-order moment estimate and the second-order moment estimate; Step 4.6, Update parameters: Update parameters based on the revised estimate; Step 4.7: After completing the parameter update, Adam enters the next time step and repeats the above steps. Each iteration calculates new gradients, updates the first- and second-order moment estimates, performs bias correction, and updates the parameters based on the new estimates. As training progresses, the parameter updates gradually approach the minimum value of the loss function. Step 4.8: After the optimization is completed, the network model with the optimal weight parameters is obtained.

8. The method for detecting transformer winding status based on temperature characteristics according to claim 7, characterized in that: The calculation formula for the average relative error percentage in Step 6 is: ； Where, Indicates the predicted value of the hot spot temperature of the transformer winding; Indicates the measured value of the hot spot temperature of the transformer winding; The number of data points representing the hot spot temperature of the winding.

Images