Method for judging unbalanced state of transformer load

Abstract

The invention discloses a method and a system for judging a load imbalance state of a transformer. The judgment method comprises the following steps: (1) acquiring infrared temperature image data of a to-be-detected transformer; (2) processing the infrared temperature image data to obtain an intermediate image; and (3) inputting the intermediate image into the trained judgment model to obtain the current unbalanced load operation state of the transformer to be detected. According to the invention, when a common current detection method cannot be carried out, the three-phase unbalanced load operation state of the transformer can be effectively detected; the device does not have electric contact with electrified equipment, wiring or installation of other measuring equipment is not needed in the test process, the test safety and the detection efficiency are higher, the operation is convenient, and the detection cost is lower.

Description

Technical Field

[0001] The invention belongs to the technical field of filter materials, and in particular relates to a method for determining an unbalanced state of transformer load. Background Art

[0002] Transformers are an important part of the power system and are responsible for the transmission of electric energy between power networks of different voltage levels. When the transformer is running under the condition of large unbalanced three-phase load, it will have adverse effects on the power grid and transformer, mainly manifested in: ① When the three-phase load is unbalanced, there are differences in the three-phase currents, and unbalanced currents are generated on the neutral line of the power grid, which on the one hand leads to increased line losses, and on the other hand leads to the drift of the neutral point of the transformer, and the phase voltages of each phase change, which can easily cause the electrical equipment to burn or become unusable; ② Increased transformer power loss; ③ Hysteresis and eddy current losses are generated in steel components, resulting in local overheating of the transformer and aging of insulation; ④ The output capacity of the transformer cannot reach the rated value, the spare capacity is reduced, and the overload capacity is reduced. When the three-phase load is seriously unbalanced, as the operating time increases, the transformer may burn due to excessive current in one phase.

[0003] Therefore, in order to improve the reliability of power supply, the power system generally has strict requirements on the three-phase load imbalance of the transformer. At present, the three-phase load imbalance problem is mainly discovered by checking the transformer gateway meter, but in order to verify whether the meter results are accurate, it is often necessary to manually detect the three-phase load current and neutral line current of the transformer on site. This method mainly has the problems of being time-consuming and labor-intensive, and requiring contact with live equipment. Especially for distribution transformers that are numerous and widely distributed, more personnel and equipment are required for detection. In addition, for distribution transformers in remote locations, there is often a lack of three-phase load imbalance detection means, and the load imbalance phenomenon is often difficult to detect. The testing difficulty is also relatively large, which is not conducive to the safe and reliable operation of the transformer. Summary of the invention

[0004] In view of the above technical problems, the present application proposes a method for determining the unbalanced state of transformer load, and the specific technical solution is as follows:

[0005] A method for determining a transformer load imbalance state, the method comprising the following steps:

[0006] (1) Obtain infrared temperature image data of the transformer to be tested;

[0007] (2) Processing the infrared temperature image data to obtain an intermediate image;

[0008] (3) The intermediate image is input into the trained judgment model to obtain the current unbalanced load operating status of the transformer to be tested.

[0009] Furthermore, a collection time interval and a collection quantity are set, and infrared temperature image data of the transformer to be inspected is acquired according to the collection time interval.

[0010] Furthermore, the infrared temperature image data is processed to obtain an intermediate image, including obtaining a time series fusion image, a temperature gradient image, and a hot spot image respectively through time series image fusion, temperature gradient calculation, and hot spot detection.

[0011] Furthermore, a method for obtaining a time series fused image by fusion of time series images is: averaging the parameter values ​​representing the temperature information on the corresponding pixels of all acquired images to obtain a time series fused image.

[0012] Furthermore, the method for obtaining the temperature gradient image by temperature gradient calculation is:

[0013] (1) According to the time series relationship of image acquisition, the parameter value representing the temperature information of the corresponding pixel on the subsequent image is subtracted from the parameter value representing the temperature information of the corresponding pixel on the previous image to obtain N-1 intermediate images, and the parameter value representing the temperature information of the corresponding pixel on each intermediate image is obtained;

[0014] (2) Add the parameter values ​​of the corresponding pixels on the N-1 intermediate images to obtain an intermediate image 2;

[0015] (3) Use the Sobel operator to calculate the temperature gradient on the intermediate image 2 and obtain a temperature gradient image.

[0016] Furthermore, the method of obtaining a hotspot image through hotspot detection is:

[0017] (1) Setting a parameter value threshold so that all areas above the threshold on the original image are marked as hot spots, and the unmarked areas are set to default values ​​for the parameter value representing the temperature information on the pixels;

[0018] (2) All labeled images are fused to obtain a hotspot image.

[0019] Furthermore, the parameter value is the brightness or hue of the pixel.

[0020] Furthermore, data enhancement processing is performed on the acquired time series fusion images, temperature gradient images, and hot spot images.

[0021] Furthermore, the judgment model adopts a convolutional neural network model, and the structure of the convolutional neural network model is:

[0022] (1) Input layer (440, 440, 3);

[0023] (2) Convolutional layer

[0024] The first convolutional layer

[0025] Number of filters: 64

[0026] Kernel size: 3x3

[0027] Step length: 1

[0028] Filling: same

[0029] Activation function: ReLU; Second convolutional layer Number of filters: 64 Kernel size: 3x3 Step size: 1 Padding: same Activation function: ReLU; Third convolutional layer Number of filters: 128 Kernel size: 3x3 Step size: 1 Padding: same Activation function: ReLU; (3) Pooling layer

[0030] The first pooling layer has a pooling size of 2x2 and a stride of 2.

[0031] The second pooling layer

[0032] Pooling size: 2x2 stride: 2;

[0033] (4) Flattening layer

[0034] (5) Fully connected layer Number of neurons in the first fully connected layer: 512

[0035] Activation function: ReLU;

[0036] (6) Output layer

[0037] Number of neurons: 1

[0038] Activation function: Sigmoid;

[0039] Training parameters

[0040] Optimizer: Adam

[0041] Learning rate: 0.001

[0042] Loss Function: Binary Crossentropy

[0043] Evaluation index: Accuracy

[0044] Batch size: 32

[0045] Number of training rounds: 10

[0046] Validation set ratio: 20%

[0047] Threshold: 0.5, an alarm is triggered when the model prediction probability is greater than 0.5.

[0048] The beneficial effects of the present invention are as follows: when the commonly used current detection method cannot be carried out, the present invention can effectively detect the three-phase unbalanced load operating state of the transformer; and the present invention has no electrical contact with the live equipment, and the test process does not require wiring or installation of other measuring equipment, the test is safer, the detection efficiency is higher, the operation is convenient, and the detection cost is low. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Shown is a flowchart of the determination method. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments.

[0051] The method for determining the unbalanced load state of a transformer of the present application aims to collect relevant parameters during the operation of the transformer by a non-invasive method, and after processing the relevant parameters, use a pre-configured and trained judgment model to input the transformer-related parameters acquired at any time into the judgment model, and output the judgment result, so as to improve the detection efficiency, convenience and safety of the unbalanced load state of the transformer.

[0052] In the implementation scheme, the method for determining the unbalanced state of the transformer load includes the following steps:

[0053] (1) Obtain infrared temperature image data of the transformer to be tested;

[0054] (2) Processing the infrared temperature image data to obtain an intermediate image;

[0055] (3) The intermediate image is input into the trained judgment model to obtain the current unbalanced load operating status of the transformer to be tested.

[0056] The method of this embodiment can effectively detect the three-phase unbalanced load problem of the transformer without current detection, and has the advantages of no need to contact the live parts of the equipment and high detection safety and efficiency.

[0057] In the implementation scheme, the acquisition time interval and the number of acquisitions are set to obtain the infrared temperature image data of the transformer to be inspected. For example, the acquisition time interval is selected as a 1-minute time interval, because the temperature change of the transformer may be more significant in a shorter period of time; for example, 5 images are collected each time to ensure that there is enough data to capture the dynamic characteristics of the temperature change.

[0058] In the implementation scheme, the purpose of processing the infrared temperature image data to obtain the intermediate image is to enhance the information of the time dimension and enhance the correlation between the temperature image and the transformer load imbalance by fusing multiple images.

[0059] The infrared temperature image data is processed to obtain an intermediate image, including obtaining image data of three different channels, namely, a time series fusion image, a temperature gradient image, and a hotspot image, through time series image fusion, temperature gradient calculation, and hotspot detection.

[0060] In the implementation scheme, the method for obtaining a time series fusion image by fusion of time series images is: averaging the parameter values ​​representing the temperature information on the corresponding pixels of all the acquired images, and obtaining the parameter value representing the temperature information on each pixel on the time series fusion image. Where N is the number of original images, I i It is the parameter value that represents the temperature information on the pixel of the image.

[0061] In some embodiments, I i It can be the brightness on the pixel. The infrared temperature image data shows the difference in color according to the difference in temperature. The higher the temperature, the darker the color and the greater the brightness. Therefore, the brightness can distinguish and characterize the temperature information.

[0062] In some embodiments, I i It can be the hue of the pixel. Based on the unique degree of the position corresponding to different colors on the hue wheel, different temperatures will result in different colors, and the corresponding degrees on the hue wheel will also be different. Therefore, the hue can distinguish and characterize the temperature information.

[0063] In the implementation scheme, the method for obtaining the temperature gradient image by temperature gradient calculation is:

[0064] (1) According to the time series relationship of image acquisition, the parameter value representing the temperature information of the corresponding pixel in the next image is subtracted from the parameter value representing the temperature information of the corresponding pixel in the previous image to obtain N-1 intermediate images, and the parameter value representing the temperature information of the corresponding pixel in each intermediate image is obtained. i -I i-1 , ΔI may be positive, 0, or negative;

[0065] (2) Add ΔI on corresponding pixels in N-1 intermediate images to obtain an intermediate image 2;

[0066] (3) Use the Sobel operator to calculate the temperature gradient on the intermediate image 2 and obtain a temperature gradient image.

[0067] In the implementation scheme, the method for obtaining a hotspot image through hotspot detection is:

[0068] (1) Set a threshold I threshold, so that all areas above the threshold on the original image are marked as hot spots, the formula is H = (I>I threshold ), the unmarked area will set the parameter value representing the temperature information on the pixel to the default value. The default value is determined according to the parameter value selection. For example, when the parameter value is selected as hue, according to the color distribution on the color wheel, the high temperature area (red, orange and other warm colors) corresponds to between 0 and 60°, and the low temperature area (blue, green and other cold colors) corresponds to about 240°, so the default value is set to 240° accordingly;

[0069] (2) All the marked images are subjected to maximum fusion to obtain a hotspot image. Maximum fusion means comparing the parameter values ​​of all images at corresponding pixels and taking the maximum value.

[0070] In some embodiments, data enhancement processing is also performed on the acquired time series fusion images, temperature gradient images, and hotspot images to increase data diversity and improve the generalization ability of the model. Data enhancement processing includes:

[0071] Rotate: Randomly rotate the image by an angle in the range [-10°, 10°].

[0072] Flip: Flip the image horizontally or vertically.

[0073] Scaling: Randomly scale the image with a scaling factor in the range [0.9, 1.1].

[0074] Brightness adjustment: Adjust the brightness and contrast of the image. The brightness adjustment range is [0.8, 1.2], and the contrast adjustment range is [0.8, 1.2].

[0075] In the implementation scheme, the judgment model adopts a convolutional neural network model, and the setting of the convolutional neural network model is as follows:

[0076] 1. Input layer

[0077] Input shape: (440,440,3);

[0078] 440x440: the height and width of the image;

[0079] 3: Number of channels, including original temperature image, temperature gradient image, and hotspot image.

[0080] 2. Convolutional Layer

[0081] The convolution layer is the core part of CNN and is used to extract local features. The main parameters of each convolution layer include:

[0082] The first convolutional layer

[0083] Number of filters: 64

[0084] Function: 64 filters can capture 64 different local features.

[0085] Kernel size: 3x3

[0086] Function: The 3x3 convolution kernel can capture the features of adjacent pixels in a local area.

[0087] Step length: 1

[0088] Function: A step size of 1 means that the convolution kernel moves one pixel at a time, which can retain more local information.

[0089] Filling: same

[0090] Function: By adding zero padding to the input edge, the output size is made consistent with the input size, keeping the image size unchanged.

[0091] Activation function: ReLU

[0092] Function: The ReLU function can convert negative values ​​to 0 and retain positive values, which helps the model learn complex features.

[0093] The second convolutional layer

[0094] Number of filters: 64

[0095] Function: Continue to capture more complex local features.

[0096] Kernel size: 3x3

[0097] Function: The 3x3 convolution kernel continues to capture features in the local area.

[0098] Step length: 1

[0099] Function: A step size of 1 means that the convolution kernel moves one pixel at a time, retaining more local information.

[0100] Filling: same

[0101] Function: Keep the output size consistent with the input size.

[0102] Activation function: ReLU

[0103] Function: Introducing nonlinearity to help the model learn complex features.

[0104] The third convolutional layer

[0105] Number of filters: 128

[0106] Function: Increase the number of filters to capture more complex features.

[0107] Kernel size: 3x3

[0108] Function: The 3x3 convolution kernel continues to capture features in the local area.

[0109] Step length: 1

[0110] Function: A step size of 1 means that the convolution kernel moves one pixel at a time, retaining more local information.

[0111] Filling: same

[0112] Function: Keep the output size consistent with the input size.

[0113] Activation function: ReLU

[0114] Function: Introducing nonlinearity to help the model learn complex features.

[0115] 3. Pooling layer

[0116] The pooling layer is used to reduce feature dimensions, reduce computational effort, and prevent overfitting.

[0117] The first pooling layer

[0118] Pooling size: 2x2

[0119] Function: The 2x2 pooling window can spatially merge the features of adjacent 2x2 regions into 1, thereby halving the width and height of the image.

[0120] Step length: 2

[0121] Effect: A step size of 2 means that the pooling window moves 2 pixels at a time, reducing the size of the feature map.

[0122] The second pooling layer

[0123] Pooling size: 2x2

[0124] Effect: The 2x2 pooling window continues to spatially merge the features of adjacent 2x2 regions into 1, further halving the width and height of the image.

[0125] Step length: 2

[0126] Effect: A step size of 2 means that the pooling window moves 2 pixels at a time, reducing the size of the feature map.

[0127] 4. Flattening Layer

[0128] The flattening layer flattens the multi-dimensional feature map into a one-dimensional vector for input into the fully connected layer.

[0129] Function: Flatten the multi-dimensional feature map output by the convolution layer and the pooling layer into a one-dimensional vector.

[0130] 5. Fully connected layer

[0131] The fully connected layer is used to perform high-level abstraction on the extracted features and further learn the complex relationships of the data.

[0132] The first fully connected layer

[0133] Number of neurons: 512

[0134] Role: 512 neurons can learn more complex feature combinations.

[0135] Activation function: ReLU

[0136] Function: Introducing nonlinearity to help the model learn complex features.

[0137] 6. Output Layer

[0138] The output layer is used to output the final prediction results.

[0139] Number of neurons: 1

[0140] Function: One neuron outputs a probability value between 0 and 1, indicating whether the transformer is in a load imbalance state.

[0141] Activation function: Sigmoid

[0142] Function: The Sigmoid function compresses the output to between 0 and 1, which is convenient for binary classification tasks.

[0143] Training parameters

[0144] Optimizer: Adam

[0145] Learning rate: 0.001

[0146] Loss Function: Binary Crossentropy

[0147] Evaluation index: Accuracy

[0148] Batch size: 32

[0149] Number of training rounds: 10

[0150] Validation set ratio: 20%

[0151] Threshold: 0.5, an alarm is triggered when the model prediction probability is greater than 0.5.

[0152] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them.

Claims

1. A method for determining a transformer load imbalance state, characterized in that: The determination method comprises the following steps: (1) Obtain infrared temperature image data of the transformer to be tested; (2) Processing the infrared temperature image data to obtain an intermediate image; (3) The intermediate image is input into the trained judgment model to obtain the current unbalanced load operating status of the transformer to be tested.

2. The method for determining the unbalanced state of transformer load according to claim 1, characterized in that: Set the collection time interval and collection quantity, and obtain the infrared temperature image data of the transformer to be tested according to the collection time interval.

3. The method for determining the unbalanced state of transformer load according to claim 2, characterized in that: The infrared temperature image data is processed to obtain an intermediate image, including obtaining a time series fusion image, a temperature gradient image, and a hot spot image respectively through time series image fusion, temperature gradient calculation, and hot spot detection.

4. The method for determining the unbalanced state of transformer load according to claim 3, characterized in that: The method for obtaining a time series fusion image by time series image fusion is: averaging the parameter values ​​representing the temperature information on the corresponding pixels of all acquired images to obtain a time series fusion image.

5. The method for determining the unbalanced state of transformer load according to claim 3, characterized in that: The method for obtaining a temperature gradient image by temperature gradient calculation is: (1) According to the time series relationship of image acquisition, the parameter value representing the temperature information of the corresponding pixel on the subsequent image is subtracted from the parameter value representing the temperature information of the corresponding pixel on the previous image to obtain N-1 intermediate images, and the parameter value representing the temperature information of the corresponding pixel on each intermediate image is obtained; (2) Add the parameter values ​​of the corresponding pixels on the N-1 intermediate images to obtain an intermediate image 2; (3) Use the Sobel operator to calculate the temperature gradient on the intermediate image 2 and obtain a temperature gradient image.

6. The method for determining the unbalanced state of transformer load according to claim 3, characterized in that: The method of obtaining hotspot images through hotspot detection is: (1) Setting a parameter value threshold so that all areas above the threshold on the original image are marked as hot spots, and the unmarked areas are set to default values ​​for the parameter value representing the temperature information on the pixels; (2) All labeled images are fused to obtain a hotspot image.

7. The method for determining the unbalanced state of transformer load according to any one of claims 4 to 6, characterized in that: The parameter value is the brightness or hue of the pixel.

8. The method for determining the unbalanced state of transformer load according to claim 3, characterized in that: Data enhancement processing is performed on the acquired time series fusion images, temperature gradient images, and hot spot images.

9. The method for determining the unbalanced state of transformer load according to claim 1, characterized in that: The judgment model adopts a convolutional neural network model, and the structure of the convolutional neural network model is: (1) Input layer (440, 440, 3); (2) Convolutional layer The first convolutional layer Number of filters: 64 Kernel size: 3x3 Step size: 1 Padding: same Activation function: ReLU; Second convolutional layer Number of filters: 64 Kernel size: 3x3 Step size: 1 Padding: same Activation function: ReLU; Third convolutional layer Number of filters: 128 Kernel size: 3x3 Step size: 1 Padding: same Activation function: ReLU; (3) Pooling layer: Pooling size of the first pooling layer: 2x2, step size: 2; The second pooling layer pooling size: 2x2 Step length: 2; (4) Flattening layer (5) Fully connected layer The first fully connected layer Number of neurons: 512 Activation function: ReLU; (6) Output layer Number of neurons: 1 Activation function: Sigmoid; Training parameters Optimizer: Adam Learning rate: 0.001 Loss function: Binary Crossentropy Evaluation indicator: Accuracy Batch size: 32 Number of training rounds: 10 Validation set ratio: 20% Threshold: 0.5, an alarm is triggered when the model prediction probability is greater than 0.5.

Images