A transformer tap changer control method based on artificial intelligence

Abstract

The application relates to the technical field of smart grids, in particular to a transformer tap switch control method based on artificial intelligence, which comprises the following steps: inputting a switch image into a gear position recognition network which has been trained, and acquiring a real-time gear position; in response to the real-time gear position not being equal to a preset gear position in a control instruction, controlling a transformer tap switch according to the rotation angle between the real-time gear position and the preset gear position. Through the technical scheme, the real-time gear position of the transformer tap switch can be accurately acquired, and the control precision of the transformer tap switch is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart grid, and particularly relates to a transformer tap switch control method based on artificial intelligence. BACKGROUND

[0002] The transformer plays a key role in the power system, and the gear adjustment of the transformer tap switch is crucial to ensure the stability and efficiency of power transmission. In practical applications, the tap switch of the transformer needs to be frequently adjusted to adapt to different load conditions, thereby ensuring the safe and reliable operation of the power system and increasing the flexibility of power grid dispatching.

[0003] In the process of smart grid operation, it is often necessary to rotate the transformer tap switch to adjust it to the preset gear. Before each adjustment, the real-time gear of the transformer tap switch is directly related to the accuracy of the rotation angle of the transformer tap switch. Therefore, how to accurately obtain the real-time gear of the transformer tap switch and improve the control accuracy of the transformer tap switch is a problem to be solved. SUMMARY

[0004] In order to solve the technical problem of low control accuracy of the running transformer tap switch, the present application provides a transformer tap switch control method based on artificial intelligence to accurately obtain the real-time gear of the transformer tap switch and improve the control accuracy of the transformer tap switch.

[0005] In the first aspect of the present application, a transformer tap switch control method based on artificial intelligence is provided, which comprises: inputting a switch image into a trained gear recognition network to obtain a real-time gear; in response to the real-time gear not being equal to a preset gear in a control instruction, controlling the transformer tap switch according to the rotation angle between the real-time gear and the preset gear; wherein the gear recognition network comprises a plurality of convolution modules connected in series and a classifier; the convolution module comprises a convolution layer and a screening binary image, the screening binary image is multiplied by the intermediate feature map output by the convolution layer to obtain the feature extraction result of the convolution module; the classification module is used for classifying the feature extraction result mapping of the last convolution module to obtain the real-time gear; the screening binary image is obtained by: using a CAM algorithm to obtain the activation mapping of the intermediate feature map of the convolution layer, and calculating the feature response value of each position point in the intermediate feature map, the feature response value is positively correlated with the contribution degree in the activation mapping, and the Euclidean distance between the feature vector in the intermediate feature map and the average feature vector; performing differential binary operation on the feature response value of each position point in the intermediate feature map to obtain the screening binary image of the convolution layer.

[0006] Preferably, the control of the transformer tap changer according to the rotation angle between the real-time gear and the preset gear comprises: querying the switch angles of the preset gear and the real-time gear, subtracting the switch angle of the real-time gear from the switch angle of the preset gear to obtain the rotation angle between the real-time gear and the preset gear, and realizing the control of the transformer tap changer.

[0007] Preferably, the method for obtaining the activation map of the convolutional layer k comprises: performing global average pooling on the intermediate feature map of the convolutional layer k to obtain normalized weights of each channel; the contribution degree of the position point (i, j) in the activation map of the convolutional layer k is W, H and C are the width, height and channel number of the intermediate feature map respectively, and β c is the normalized weight of the channel c, is the gradient value of the position point (i, j) in the channel c of the intermediate feature map.

[0008] Preferably, the feature response value x i，j of the position point (i, j) in the intermediate feature map of the convolutional layer k is:

[0009] the contribution degree of the position point (i, j) T i，j is the feature vector of the position point (i, j), is the average feature vector of the intermediate feature map, and exp() is the exponential function with e as the base.

[0010] Preferably, the differential binarization operation is:

[0011] α is the amplification factor, is the feature response value of the position point (i, j) in the intermediate feature map of the convolutional layer k, and X k is the adaptive threshold value of the convolutional layer k, is the screening value of the position point (i, j) in the screening binary map of the convolutional layer k.

[0012] Preferably, the training method of the gear recognition network comprises: inputting a training sample into the gear recognition network to obtain an output result, calculating a cross-entropy loss function between the output result and a labeled gear of the training sample; obtaining the screening binary map of each convolutional layer, and calculating the change amount of the screening binary map of each convolutional layer in two adjacent iterations, taking the sum of the change amount and the cross-entropy loss function as the total loss, and iteratively training the gear recognition network until the total loss is less than a preset loss or the iteration number is greater than a preset number, and the training is completed.

[0013] Preferably, the total loss L(t) of the tth iteration satisfies: ​

[0014] ut represents the training sample input at iteration t. For the labeled position of the training sample ut, y ut The output of the gear position recognition network, for and y ut The cross-entropy loss function between K and K is the total number of convolutional layers. This is the filtered binary image of convolutional layer k in the t-th iteration. This is a binary image of the selected convolutional layer k in the (t-1)th iteration; ζ k is the adjustment coefficient for convolutional layer k.

[0015] Preferably, the adjustment coefficient ζ of convolutional layer k k Satisfying the relation:

[0016] ζ k =1-exp(-σ k ), σ k The variance of the binary image is used to filter the convolutional layer k in the historical iterations.

[0017] Preferably, the adjustment coefficients of each convolutional layer are preset values.

[0018] The technical solution of this application has the following beneficial technical effects:

[0019] The trained gear position recognition network accurately identifies the real-time gear position in the switch image, and then compares the real-time gear position with the preset gear position in the control command to accurately obtain the rotation angle, thereby achieving precise control of the transformer tap changer.

[0020] Furthermore, the gear position recognition network includes multiple cascaded convolutional modules. Each convolutional module includes a convolutional layer and a filtering binary map. The convolutional layer can extract features from the switch image to obtain an intermediate feature map, while the filtering binary map can filter the feature vectors in the intermediate feature map, removing noise information that is irrelevant to gear position recognition and retaining significant features that can distinguish gear positions, ensuring that the gear position recognition network can output accurate real-time gear positions.

[0021] Furthermore, during the training process of the gear recognition network, the stability of the filtering binary map of the same convolutional layer is used as part of the overall loss. This constrains the filtering binary maps of each convolutional layer to remain consistent in two adjacent iterations, ensuring that the filtering binary map can accurately remove noise information and retain significant features that can distinguish gear categories, thereby accelerating the convergence speed of the network. Attached Figure Description

[0022] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description read in conjunction with the accompanying drawings, in which several embodiments of the present application are shown by way of example, and in which the same reference numerals indicate similar, corresponding, or identical elements throughout the several drawings, wherein:

[0023] Figure 1 is a flowchart of a transformer tap changer control method based on artificial intelligence according to an embodiment of the present application;

[0024] Figure 2 is a structural diagram of a convolution module according to an embodiment of the present application;

[0025] Figure 3 is a flowchart of a training method of a gear recognition network according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0027] It should be understood that when the claims, the specification and the drawings of the present application use the terms "first", "second", etc., they are only used to distinguish different objects, and are not used to describe a specific sequence. The terms "include" and "contain" used in the specification and claims of the present application indicate the presence of the described features, whole, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, whole, steps, operations, elements, components and / or sets thereof.

[0028] According to a first aspect of the present application, the present application provides a transformer tap changer control method based on artificial intelligence. Figure 1 is a flowchart of a transformer tap changer control method based on artificial intelligence according to an embodiment of the present application. As shown in Figure 1 , the transformer tap changer control method based on artificial intelligence includes steps S101 to S102, which will be described in detail below.

[0029] S101, input the switch image into the trained gear recognition network, and obtain the real-time gear.

[0030] In one embodiment, a switch image of the transformer tap changer is acquired by using an image acquisition device, a field of view of the image acquisition device can cover all areas of the transformer tap changer, and a pose of the image acquisition device is fixed, that is, a position of the transformer tap changer is fixed in the switch image.

[0031] The switch image is input into the trained gear recognition network, and a real-time gear of the transformer tap changer can be acquired. The gear recognition network comprises a plurality of convolution modules connected in series and a classifier. The convolution module comprises a convolution layer and a screening binary image. The screening binary image is multiplied by an intermediate feature map output by the convolution layer to obtain a feature extraction result of the convolution module. The classification module is used for classifying the feature extraction result of the last convolution module to obtain the real-time gear.

[0032] The classifier is a Softmax function.

[0033] On the basis of a convolutional neural network feature extraction backbone such as ResNet, VGGNet or ShuffleNet, a screening binary image is added after each convolution layer to obtain the gear recognition network. The convolution layer and the screening binary image of the convolution layer form a convolution module.

[0034] In one embodiment, one convolution layer corresponds to one screening binary image. Please refer to Figure 2 is a structural diagram of the convolution module according to an embodiment of the present application. The method for obtaining the screening binary image in the convolution module comprises: acquiring an activation mapping image of an intermediate feature map in the convolution layer by using a CAM algorithm, and calculating a feature response value of each position point in the intermediate feature map. The feature response value is positively correlated with a contribution degree in the activation mapping image, and a Euclidean distance between a feature vector in the intermediate feature map and an average feature vector. A differential binary operation is performed on the feature response value of each position point in the intermediate feature map to obtain the screening binary image of the convolution layer.

[0035] The output result of the CAM algorithm (Class Activation Mapping, class activation mapping algorithm) is an activation mapping image. The activation mapping image comprises a contribution degree of each position point in the intermediate feature map. The greater the contribution degree is, the more effective the feature vector of the corresponding position point is for gear recognition. Specifically, the method for acquiring the activation mapping image of the convolution layer k comprises: performing global average pooling on the intermediate feature map of the convolution layer k to obtain normalized weights of each channel. The contribution degree of the position point (i, j) in the activation mapping image of the convolution layer k is W, H and C are width, height and channel number of the intermediate feature map respectively, and β c is the normalized weight of the channel c, ​is a gradient value of a position point (i, j) in a channel c of the intermediate feature map.

[0036] After obtaining the activation map of the intermediate feature map, a feature response value of each position point in the intermediate feature map can be calculated; the greater the contribution degree of the position point (i, j) in the activation map, the more effective the feature vector of the position point (i, j) is in distinguishing the gear categories, and the greater the feature response value of the position point (i, j) is assigned. Further, an average feature vector of the intermediate feature map is calculated, and if the Euclidean distance between the feature vector of the position point (i, j) and the average feature vector is large, it indicates that the position point (i, j) is a salient feature in the intermediate feature map, and the position point (i, j) is also assigned a greater feature response value. Therefore, the greater the feature response value is, the more effective the feature vector of the corresponding position point in the intermediate feature map is in distinguishing the gear categories, and the stronger the feature saliency is. Specifically, the feature response value x i，j of the position point (i, j) in the intermediate feature map of the convolutional layer k is:

[0037] is the contribution degree of the position point (i, j) T i，j is the feature vector of the position point (i, j), is an average feature vector of the intermediate feature map, and exp() is an exponential function with e as the base.

[0038] wherein the differential binarization operation is: α is an amplification factor, is a feature response value of a position point (i, j) in an intermediate feature map of a convolutional layer k, and X k is an adaptive threshold value of the convolutional layer k, is a screening value of the position point (i, j) in a screening binary map of the convolutional layer k.

[0039] wherein the value of the amplification factor α is 50, and the value of the screening value is 0 or 1.

[0040] It should be noted that one convolutional layer corresponds to one adaptive threshold value, and the adaptive threshold value is a trainable network parameter, and the adaptive threshold values of the convolutional layers can be continuously updated in the subsequent training process of the gear recognition network.

[0041] The adaptive threshold value of each convolutional layer is determined by using the differential binarization operation, if the feature response value of the position point is greater than the adaptive threshold value, the screening value of the position point is set to 1, and if the feature response value of the position point is not greater than the adaptive threshold value, the screening value of the position point is set to 0, thereby obtaining the screening binary map of each convolutional layer.

[0042] The size of the intermediate feature map is WxHxC, W and H are the width and height of the intermediate feature map, and C is the number of channels of the intermediate feature map. If the value of the position point (i, j) in the screening binary image is 1, it means that the feature vector of the position point (i, j) in the intermediate feature map can provide effective information for gear identification, and the feature vector of the position point (i, j) in the intermediate feature map is retained. On the contrary, if the value of the position point (i, j) is 0, it means that the feature vector of the position point (i, j) in the intermediate feature map cannot provide effective information for gear identification, and the feature vector of the position point (i, j) in the intermediate feature map can be regarded as noise information. At this time, the feature vector of the position point (i, j) in the intermediate feature map is deleted to improve the accuracy of gear identification. The size of the feature vector is 1x1xC.

[0043] In this way, the gear identification network can obtain multiple intermediate feature maps of the switch image, and adaptively generate a screening binary image for each intermediate feature map to screen the intermediate feature map, delete noise information irrelevant to gear identification, retain significant features capable of distinguishing gear categories, and ensure that the gear identification network can output accurate real-time gears.

[0044] S102, in response to the real-time gear being different from the preset gear in the control instruction, controlling the transformer tap switch according to the rotation angle between the real-time gear and the preset gear.

[0045] In one embodiment, the control instruction includes a preset gear. When the real-time gear is equal to the preset gear in the control instruction, the transformer tap switch does not need to be adjusted. When the real-time gear is not equal to the preset gear in the control instruction, the transformer tap switch needs to be adjusted at this time.

[0046] Specifically, controlling the transformer tap switch according to the rotation angle between the real-time gear and the preset gear includes: querying the switch angle of the preset gear and the real-time gear, subtracting the switch angle of the real-time gear from the switch angle of the preset gear to obtain the rotation angle between the real-time gear and the preset gear, and realizing the control of the transformer tap switch.

[0047] Wherein, the switch angle corresponding to each gear category is related to the model of the transformer tap switch, and the implementer can set it according to the actual situation.

[0048] In this way, on the basis of accurately obtaining the real-time gear of the transformer tap switch, the precise control of the transformer tap switch is realized.

[0049] In one embodiment, in order to enable the gear identification network to output accurate real-time gears, the gear identification network needs to be trained. Please refer to Figure 3 is a flowchart of a training method of the gear identification network according to an embodiment of the present application. As shown in Figure 3As shown, the training method of the gear recognition network includes steps S201 to S202, which are described in detail below.

[0050] S201, input the training sample into the gear recognition network to obtain an output result, and calculate a cross-entropy loss function between the output result and the labeled gear of the training sample.

[0051] The training sample is a switch image collected in a historical time, and the gear category of the switch image is manually labeled, that is, the labeled gear of the training sample is obtained.

[0052] The cross-entropy loss function is used to judge the consistency between the output result and the labeled gear of the training sample. When the output result is equal to the labeled gear of the training sample, the cross-entropy loss function is equal to 0, otherwise, the cross-entropy loss function is a larger value.

[0053] S202, obtain the screening binary image of each convolutional layer, and calculate the change amount of the screening binary image of each convolutional layer in adjacent two iterations, and take the sum of the change amount and the cross-entropy loss function as the total loss. The gear recognition network is iteratively trained until the total loss is less than a preset loss or the iteration number is greater than a preset number, and the training is completed.

[0054] The preset loss is 0.01; and the preset number is 500.

[0055] The gradient descent method is used to iteratively train the gear recognition network, and the network parameters in the gear recognition network are updated each time. The total loss L(t) of the tth iteration satisfies:

[0056] ut is the training sample input at the tth iteration, is the labeled gear of the training sample ut, y ut is the output result of the gear recognition network, is the cross-entropy loss function between y and y ut , K is the total number of convolutional layers, is the screening binary image of the convolutional layer k in the tth iteration, is the screening binary image of the convolutional layer k in the (t-1)th iteration; ζ k is the adjustment coefficient of the convolutional layer k. The adjustment coefficients of each convolutional layer are preset values, and the preset value is 0.5.

[0057] It can be understood that the first part

[0058] The output result of the gear recognition network is constrained to be consistent with the labeled gear; in addition, a screening binary image of a convolution layer is used to screen the intermediate feature map output by the convolution layer. Since the pose of the image acquisition device is fixed and unchanged, the position of the transformer tap switch in the switch image is also fixed and unchanged. Therefore, in an ideal state, the position points in the intermediate feature map of each convolution layer that can provide effective information for gear recognition are also fixed and unchanged. In other words, the screening binary image of the same convolution layer in different iteration processes should be fixed and unchanged. Therefore, the second part of the total loss L(t) is the change amount of the screening binary image of each convolution layer in adjacent iterations, and is used to constrain the screening binary image of each convolution layer to remain consistent.

[0059] In other embodiments, the variance of the screening binary image of any convolution layer in the historical iteration is calculated. If the variance of the convolution layer k is large, it indicates that the screening binary image of the convolution layer k has large fluctuations. At this time, the adjustment coefficient of the convolution layer k should be increased. Therefore, the adjustment coefficient ζ k of the convolution layer k satisfies the relationship:

[0060] ζ k =1-exp(-σ k ), σ k is the variance of the screening binary image of the convolution layer k in the historical iteration.

[0061] In this way, the trained gear recognition network is obtained, and the trained gear recognition network can accurately extract the image features related to gear recognition and output accurate real-time gears.

[0062] The technical principles and implementation details of the transformer tap switch control method based on artificial intelligence are introduced above through specific embodiments. The trained gear recognition network accurately identifies the real-time gear in the switch image, and then compares the real-time gear with the preset gear in the control instruction to accurately obtain the rotation angle and realize precise control of the transformer tap switch. Further, the gear recognition network includes a plurality of convolution modules connected in series, the convolution module includes a convolution layer and a screening binary image, the convolution layer can extract features from the switch image to obtain an intermediate feature map, and the screening binary image can screen the feature vectors in the intermediate feature map, delete noise information irrelevant to gear recognition, retain significant features capable of distinguishing gear categories, and ensure that the gear recognition network can output accurate real-time gears.

[0063] The technical features of the above-described embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described, but as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present disclosure.

[0064] The above embodiments only express several implementation ways of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation to the patent scope of the application. It should be pointed out that for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, which all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A transformer tap changer control method based on artificial intelligence, characterized in that, The control method includes: Input the switch image into the trained gear recognition network to obtain the real-time gear; In response to the real-time gear position not being equal to the preset gear position in the control command, the transformer tap changer is controlled based on the rotation angle between the real-time gear position and the preset gear position. The gear position recognition network includes multiple cascaded convolutional modules and a classifier. Each convolutional module includes a convolutional layer and a filtered binary map. The filtered binary map is multiplied by the intermediate feature map output by the convolutional layer to obtain the feature extraction result of the convolutional module. The classifier is used to classify the feature extraction result mapping of the last convolutional module to obtain the real-time gear position. The method for obtaining the filtered binary image includes: using the CAM algorithm to obtain the activation map of the intermediate feature map of the convolutional layer, and calculating the feature response value of each position point in the intermediate feature map. The feature response value is positively correlated with the contribution of the activation map and the Euclidean distance between the feature vector and the average feature vector in the intermediate feature map; performing differential binarization on the feature response value of each position point in the intermediate feature map to obtain the filtered binary image of the convolutional layer. Convolutional layer Methods for obtaining activation maps include: for convolutional layers The intermediate feature maps are subjected to global average pooling to obtain normalized weights for each channel; convolutional layers Location points in the activation map Contribution for: , , and These represent the width, height, and number of channels of the intermediate feature map, respectively. For channel Normalized weights, Channels of the intermediate feature map Midpoint The gradient value.

2. The transformer tap changer control method based on artificial intelligence according to claim 1, characterized in that, Controlling the transformer tap changer based on the rotation angle between the real-time and preset tap positions includes: querying the switching angles of the preset and real-time tap positions, subtracting the switching angle of the real-time tap position from the switching angle of the preset tap position to obtain the rotation angle between the real-time and preset tap positions, thereby controlling the transformer tap changer.

3. The transformer tap changer control method based on artificial intelligence according to claim 1, characterized in that, Convolutional layer Location points in the intermediate feature map Characteristic response value for: , Location point Contribution Location point eigenvectors, The average eigenvector of the intermediate feature map. It is an exponential function with base e.

4. The transformer tap changer control method based on artificial intelligence according to claim 1, characterized in that, The differential binarization operation is as follows: , As the amplification factor, Convolutional layer Location points in the intermediate feature map Characteristic response value, Convolutional layer Adaptive threshold, Convolutional layer Filtering location points in a binary image The filter value.

5. The transformer tap changer control method based on artificial intelligence according to claim 1, characterized in that, The training method for the gear position recognition network includes: Input the training samples into the gear recognition network to obtain the output results, and calculate the cross-entropy loss function between the output results and the gears labeled in the training samples. Obtain the filtering binary image of each convolutional layer, and calculate the change in the filtering binary image of each convolutional layer in two adjacent iterations. Use the sum of the change and the cross-entropy loss function as the overall loss, and iteratively train the gear recognition network until the overall loss is less than the preset loss or the number of iterations is greater than the preset number, and then the training is completed.

6. The transformer tap changer control method based on artificial intelligence according to claim 5, characterized in that, No. Total loss of the next iteration satisfy: , For the first The training samples input during iteration, For training samples The marked gear position, The output of the gear position recognition network, for and The cross-entropy loss function between them This represents the total number of convolutional layers. For the first In the next iteration, the convolutional layer Filtering binary images, For the first In the next iteration, the convolutional layer The selection of binary images; Convolutional layer The adjustment coefficient.

7. The transformer tap changer control method based on artificial intelligence according to claim 6, characterized in that, Convolutional layer adjustment coefficient Satisfying the relation: , Convolutional layers in historical iterations Filter the variance of binary plots.

8. The transformer tap changer control method based on artificial intelligence according to claim 6, characterized in that, The adjustment coefficients for each convolutional layer are all preset values.

Images