Safety margin active early warning method of power transformer considering bubble generation risk

Abstract

The application discloses a kind of power transformer safety margin active early warning method considering bubble generation risk, comprising the following steps: A, the physical characteristic parameter of initialization transformer, environmental parameter and load parameter;B, the temperature field inside transformer is modeled and the hotspot temperature θ of model is solved h ;C, bubble generation temperature T Is calculated;D, safety margin is calculated and early warning is carried out according to real-time safety margin.The application can solve the deficiency of prior art, improve the temperature field simulation precision, reduce the simulation calculation amount.

Description

Technical Field

[0001] This invention relates to the field of transformer operation and maintenance technology, and in particular to a proactive early warning method for the safety margin of power transformers that takes into account the risk of bubble generation. Background Technology

[0002] In the spot market trading model, the market-driven spatial and temporal distribution of power flow fluctuates dramatically. This, coupled with the mismatch between the original design capacity of transformers and the actual load characteristics under spot trading, leads to severe overloads in power transformers. Power transformers are core equipment in the power system, and their operating status directly affects the safe and stable operation of the power grid. Accurately assessing the operating status of severely overloaded power transformers is crucial for ensuring the safe and stable operation of the power system. Oil-paper insulation is currently the most widely used insulation material in power transformers. The formation of air bubbles from moisture within the oil-paper insulation due to temperature changes is the main cause of air bubble formation in transformer oil-paper insulation during operation. Rapid increases in transformer load cause a rapid rise in the internal temperature of the oil-paper insulation, which, when heated, can lead to air bubble formation. Under the combined effects of high temperature and electric field, the formation of air bubbles significantly increases the risk of transformer insulation failure. One of the key bases for the industry's recommended maximum overload temperature for transformers is the consideration of the air bubble formation temperature within the transformer oil-paper insulation during operation.

[0003] In the prior art, Chinese invention patent application 202310305771.0 discloses a method for constructing a transformer temperature field using the POD-RBFLP order reduction algorithm. However, this method requires high accuracy (ε≥99.9) when constructing the order reduction model, resulting in a large-scale order reduction model under complex operating conditions, which is not convenient for subsequent model fitting. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an active early warning method for the safety margin of power transformers that takes into account the risk of bubble formation. This method can overcome the shortcomings of the prior art, improve the accuracy of temperature field simulation, and reduce the amount of simulation calculation.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows.

[0006] A proactive early warning method for the safety margin of power transformers that considers the risk of bubble formation, characterized by the following steps: A. Initialize the physical characteristics, environmental parameters, and load parameters of the transformer; B. Model the internal temperature field of the transformer and solve for the hot spot temperature θ in the model. h ; C. Calculate the bubble generation temperature T ; D. Calculate the safety margin and issue early warnings based on the real-time safety margin.

[0007] Preferably, the physical characteristic parameters of the transformer include the load-to-no-load loss ratio R, the oil index x, and the oil time constant τ. o Winding time constant τ w The temperature difference Δθ between the hot spot and the top oil layer under rated current. hr Top oil temperature rise Δθ under rated loss steady state or ,pressure P Input the initial moisture content (WCP) of the insulating paper and the initial degree of polymerization (DP) of the cellulose insulation, and environmental parameters including ambient temperature (θ). a The load parameters include the load K.

[0008] As a preferred method, modeling the internal temperature field of a transformer includes the following steps. Set m sampling points, and sample the temperature of each sampling point under n different operating conditions to obtain a sample matrix M (m×n). Perform singular value decomposition on matrix M. Where U and V are orthonormal matrices, and Σ is a singular value matrix; the top d column vectors with the largest corresponding eigenvalues ​​in the singular value matrix U are extracted to construct an orthonormal basis for the sample space. U’ ,satisfy .

[0009] As a preferred option, the model hotspot temperature θ h The calculation method is as follows: A neural network model is established, comprising a cascaded physical information neural network and a convolutional neural network. The physical information neural network is used to fit a set of discretized differential equations concerning the real-time hotspot temperature. These discretized equations include the state update relationship of the hotspot temperature evolving over time steps and its constraint parameters. The convolutional neural network receives the state feature matrix corresponding to the discretized differential equations and performs convolution operations and mapping to solve for the model's hotspot temperature θ. h ; A physical information neural network consists of an input layer, hidden layers, fully connected layers, and an output layer. The output layer is connected to the hidden layers via a forget gate. An input gate is placed between the input layer and the hidden layers, and an output gate is placed between the hidden layers and the fully connected layers. The expression for the forget gate is... The expression for the input gate is The expression for the output gate is: Where σ is the activation function, X t For the output of the t-th iteration, H t-1 Let b be the hidden state at time t-1, and b be the bias term. x W is the weight matrix input to the hidden state. h The weight matrix between hidden states; The loss functions of a physical information neural network include the data bias loss function, the temporal bias loss function, and the physical consistency loss function. The data bias loss function... , used to describe the deviation between predicted temperature and actual temperature, where T i For the actual temperature, u i For temperature prediction, the time-series bias loss function This is used to describe the consistency deviation of the same sampling point in different sampling periods, where ReLU is the activation function and the physical consistency loss function is... , used to describe the consistency deviation between the prediction process and the preset physical laws, where H i The residual function is a function that assumes the physical laws of the problem. A convolutional neural network consists of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer; the loss function is set as follows: Where τ is the weighting coefficient, L ce To calculate the category cross-entropy loss of the measured and predicted temperatures discretized according to a preset temperature range, L de The maximum difference between the predicted temperature and the measured temperature in each round is represented by the softmax function, which is the activation function of the convolutional layer.

[0010] Preferably, when training the neural network model, separate learning rates are set for the physical information neural network and the convolutional neural network, and the learning rate is related to... Proportional, of which This corresponds to the gradient of the loss function value in the neural network. This is the regularization term for the corresponding neural network.

[0011] As a preferred method, the Kendall correlation coefficient of the sample set is calculated, sample data with a Kendall correlation coefficient lower than a preset threshold are deleted, and the remaining data are input into the neural network model.

[0012] Preferably, the bubble generation temperature T The calculation method is as follows: ,in a, c, n, α β is an empirical constant.

[0013] As a preferred option, safety margin ;when S When >0, the transformer is in a safe operating state; when S When the value is ≤0, it indicates the critical moment when the bubble risk is triggered under this working condition, and an active warning is issued.

[0014] The beneficial effects of adopting the above technical solution are as follows: This invention proposes a temperature field fitting method using POD-PINN-CNN for the first time, solving the requirement for feature extraction accuracy in the traditional POD-RBFLP reduction algorithm. This not only reduces the computational load of temperature field fitting but also improves fitting accuracy. Furthermore, regarding the fitting process of POD-PINN-CNN, this invention improves the architecture of the physical information neural network and convolutional neural network to better adapt to the requirements of temperature field fitting. By setting a dynamic safety margin assessment mechanism, it can provide a basis for operational status assessment, load adjustment, and early warning decisions, guiding operation and maintenance and dispatching, and improving the stability of power system operation. Attached Figure Description

[0015] Figure 1 This is a flowchart of the present invention.

[0016] Figure 2 This is a simulation experiment result diagram of the present invention.

[0017] Figure 3 The results show the experimental comparison of the errors in fitting temperature fields between the traditional POD-RBFLP algorithm and the POD-PINN-CNN algorithm proposed in this invention. Detailed Implementation

[0018] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0019] Reference Figure 1 The first step is to initialize the transformer's physical characteristic parameters, environmental parameters, and load parameters. The transformer's physical characteristic parameters include the load-to-no-load loss ratio R, the oil index x, and the oil time constant τ. o Winding time constant τ w The temperature difference Δθ between the hot spot and the top oil layer under rated current. hr Top oil temperature rise Δθ under rated loss steady state or ,pressure P Input the initial moisture content (WCP) of the insulating paper and the initial degree of polymerization (DP) of the cellulose insulation, and environmental parameters including ambient temperature (θ). a The load parameters include the load K.

[0020] The second step involves setting up m sampling points and sampling the temperature at each point under n different operating conditions to obtain a sample matrix M (m×n). Singular value decomposition is then performed on matrix M. Where U and V are orthonormal matrices, and Σ is a singular value matrix; the top d column vectors with the largest corresponding eigenvalues ​​in the singular value matrix U are extracted to construct an orthonormal basis for the sample space. U’ ,satisfy This invention relaxes the condition for ε to 95%, which can significantly reduce the size of the POD reduction model. This is because, unlike traditional algorithms that rely on the accuracy of the reduced model to ensure the accuracy of subsequent fitting, this invention relies on the deep learning capabilities of the subsequent neural network model to guarantee fitting accuracy.

[0021] The third step is to establish a neural network model, which includes a cascaded physical information neural network and a convolutional neural network.

[0022] A physical information neural network is used to fit a system of discrete differential equations relating to real-time hotspot temperatures. Traditional physical information neural networks employ a feedforward fully connected neural network structure to achieve general function approximation capabilities. However, feedforward fully connected neural networks are prone to getting trapped in local optima and overfitting, and their deep learning capabilities are insufficient. Therefore, this invention modifies the structure of the physical information neural network.

[0023] A physical information neural network consists of an input layer, hidden layers, fully connected layers, and an output layer. The output layer is connected to the hidden layers via a forget gate. An input gate is placed between the input layer and the hidden layers, and an output gate is placed between the hidden layers and the fully connected layers. The expression for the forget gate is... The expression for the input gate is The expression for the output gate is: Where σ is the activation function, X t For the output of the t-th iteration, H t-1 Let b be the hidden state at time t-1, and b be the bias term. x W is the weight matrix input to the hidden state. h This is the weight matrix between the hidden states.

[0024] The loss functions of a physical information neural network include the data bias loss function, the temporal bias loss function, and the physical consistency loss function. The data bias loss function... , used to describe the deviation between predicted temperature and actual temperature, where T i For the actual temperature, u i For temperature prediction, the time-series bias loss function This is used to describe the consistency deviation of the same sampling point in different sampling periods, where ReLU is the activation function and the physical consistency loss function is... , used to describe the consistency deviation between the prediction process and the preset physical laws, where H i The residual function is a function that assumes the physical laws.

[0025] The improved physical information neural network greatly enhances its ability to learn from historical data by introducing a feedback loop with forgetting functionality. At the same time, this invention uses a set of discrete differential equations for real-time hotspot temperatures as the output of the physical information neural network, rather than temperature field data, thereby simplifying the hidden layer structure and reducing the training difficulty of the physical information neural network.

[0026] A convolutional neural network is used to receive the state feature matrix corresponding to the discretized differential equation system, and to perform convolution operations and mapping solutions on it to obtain the model hotspot temperature θ. h A convolutional neural network consists of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer; the loss function is set as follows: Where τ is the weighting coefficient, L ce To calculate the category cross-entropy loss of the measured and predicted temperatures discretized according to a preset temperature range, L de The maximum difference between the predicted temperature and the measured temperature in each round is represented by the softmax function, which is the activation function of the convolutional layer.

[0027] This invention utilizes a convolutional neural network to perform convolution operations and mapping solutions on the state feature matrices corresponding to a set of discretized differential equations. Compared with directly calculating the set of differential equations, it has strong robustness to noise and sparse data, thus making up for the shortcomings of physical information neural network fitting results.

[0028] When training the neural network model, separate learning rates are set for the physical information neural network and the convolutional neural network, respectively. The learning rate is related to... Proportional, of which This corresponds to the gradient of the loss function value in the neural network. This refers to the regularization term for the corresponding neural network. Furthermore, this invention redesigns the activation function of the convolutional neural network, which can effectively accelerate model training efficiency.

[0029] Before inputting the sample set into the neural network model, the Kendall correlation coefficient of the sample set is calculated, and sample data with a Kendall correlation coefficient lower than a preset threshold are deleted. The remaining data is then input into the neural network model. By preprocessing the sample set data, the effective data utilization rate can be improved.

[0030] The fourth step is to calculate the bubble formation temperature. T。 ,in a， c, n α and β are empirical constants in transformer insulating oil. a The value is 8812.73. c The value is 1. n The value of is 2.5915, and the value of α is 3.7 × 10. -5The value of β is 324.4.

[0031] Fifth step, calculate the safety margin. ;when S When >0, the transformer is in a safe operating state; when S When the value is ≤0, it indicates the critical moment when the bubble risk is triggered under this working condition, and an active warning is issued.

[0032] Reference Figure 2 Simulation example: In the example, the DP of paper insulation is set to 1000, the ambient temperature is constant at 20℃, and the pressure value is set to 766 torr. The load is increased to 1.505 per unit from the initial load of 1.089 per unit (hot spot temperature 110℃).

[0033] Finally, the temperature field data predicted by this invention and the temperature field data predicted by the traditional POD-RBFLP order reduction algorithm are compared with the measured data, and referenced... Figure 3 The results show that the temperature field predicted by the present invention is more accurate.

[0034] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A proactive early warning method for the safety margin of power transformers considering the risk of bubble formation, characterized in that... Includes the following steps: A. Initialize the physical characteristics, environmental parameters, and load parameters of the transformer; B. Model the internal temperature field of the transformer and solve for the hot spot temperature θ in the model. h ; C. Calculate the bubble generation temperature T ; D. Calculate the safety margin and issue early warnings based on the real-time safety margin.

2. The active early warning method for safety margin of power transformers considering the risk of bubble formation according to claim 1, characterized in that: The physical characteristic parameters of a transformer include the load-to-no-load loss ratio R, the oil index x, and the oil time constant τ. o Winding time constant τ w The temperature difference Δθ between the hot spot and the top oil layer under rated current. hr Top oil temperature rise Δθ under rated loss steady state or ,pressure P Input the initial moisture content (WCP) of the insulating paper and the initial degree of polymerization (DP) of the cellulose insulation, and environmental parameters including ambient temperature (θ). a The load parameters include the load K.

3. The active early warning method for safety margin of power transformers considering the risk of bubble formation according to claim 2, characterized in that: Modeling the internal temperature field of a transformer includes the following steps. Set m sampling points, and sample the temperature of each sampling point under n different operating conditions to obtain a sample matrix M (m×n). Perform singular value decomposition on matrix M. Where U and V are orthonormal matrices, and Σ is a singular value matrix; the top d column vectors with the largest corresponding eigenvalues ​​in the singular value matrix U are extracted to construct an orthonormal basis for the sample space. U’ ,satisfy .

4. The active early warning method for safety margin of power transformers considering the risk of bubble formation according to claim 3, characterized in that: Model hotspot temperature θ h The calculation method is as follows: A neural network model is established, comprising a cascaded physical information neural network and a convolutional neural network. The physical information neural network is used to fit a set of discretized differential equations concerning the real-time hotspot temperature. These discretized equations include the state update relationship of the hotspot temperature evolving over time steps and its constraint parameters. The convolutional neural network receives the state feature matrix corresponding to the discretized differential equations and performs convolution operations and mapping to solve for the model's hotspot temperature θ. h ; A physical information neural network consists of an input layer, hidden layers, fully connected layers, and an output layer. The output layer is connected to the hidden layers via a forget gate. An input gate is placed between the input layer and the hidden layers, and an output gate is placed between the hidden layers and the fully connected layers. The expression for the forget gate is... The expression for the input gate is The expression for the output gate is: Where σ is the activation function, X t For the output of the t-th iteration, H t-1 Let b be the hidden state at time t-1, and b be the bias term. x W is the weight matrix input to the hidden state. h The weight matrix between hidden states; The loss functions of a physical information neural network include the data bias loss function, the temporal bias loss function, and the physical consistency loss function. The data bias loss function... , used to describe the deviation between predicted temperature and actual temperature, where T i For the actual temperature, u i For temperature prediction, the time-series bias loss function This is used to describe the consistency deviation of the same sampling point in different sampling periods, where ReLU is the activation function and the physical consistency loss function is... , used to describe the consistency deviation between the prediction process and the preset physical laws, where H i The residual function is a function that assumes the physical laws of the problem. A convolutional neural network consists of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer; the loss function is set as follows: Where τ is the weighting coefficient, L ce To calculate the category cross-entropy loss of the measured and predicted temperatures discretized according to a preset temperature range, L de The maximum difference between the predicted temperature and the measured temperature in each round is represented by the softmax function, which is the activation function of the convolutional layer.

5. The active early warning method for safety margin of power transformers considering the risk of bubble formation according to claim 4, characterized in that: When training the neural network model, separate learning rates are set for the physical information neural network and the convolutional neural network, respectively. The learning rate is related to... Proportional, of which This corresponds to the gradient of the loss function value in the neural network. This is the regularization term for the corresponding neural network.

6. The active early warning method for safety margin of power transformers considering the risk of bubble formation according to claim 5, characterized in that: Calculate the Kendall correlation coefficient of the sample set, delete sample data with a Kendall correlation coefficient lower than a preset threshold, and input the remaining data into the neural network model.

7. The active early warning method for safety margin of power transformers considering the risk of bubble formation according to claim 2, characterized in that: Bubble generation temperature T The calculation method is as follows: ,in a, c, n, α β is an empirical constant.

8. The active early warning method for safety margin of power transformers considering the risk of bubble formation according to claim 1, characterized in that: Safety margin ;when S When > 0, the transformer is in a safe operating state; when S When the value is ≤0, it indicates the critical moment when the bubble risk is triggered under this working condition, and an active warning is issued.

Images