Transformer surge current detection method, device and computer readable storage medium

Abstract

Disclosed are an inrush current detection method, an inrush current detection device, and a computer-readable storage medium for a transformer. The inrush current detection method comprises: sampling at least a portion of a current signal of the transformer to obtain a numerical matrix; providing the numerical matrix as an input to an inrush current detection neural network; and calculating and outputting a label vector corresponding to the numerical matrix by the inrush current detection neural network, wherein the label vector indicates whether the current signal is an inrush current.

Description

Technical Field

[0001] The present disclosure relates to a transformer inrush current detection method and device, and more particularly, to a transformer inrush current detection method, an inrush current detection device, and a computer-readable storage medium. Background Art

[0002] The large current that appears during the transformer power-on process is called inrush current. Inrush current is the current generated when the voltage on the power supply side establishes a magnetic field in the transformer core. This current does not occur on the side without power supply. The maximum peak value of inrush current can reach 6 to 8 times its rated current and can last up to 10 seconds. Relay protection must avoid action at this time, otherwise the transformer will not be able to be successfully powered on, that is, it will not be able to enter the voltage state of connecting the power supply from the no-voltage state of power supply disconnection. This behavior of avoiding relay protection action is called "lockout protection". The commonly used lockout protection algorithm currently uses the phenomenon that the inrush current contains rich second harmonic content to identify the appearance and disappearance of inrush current. When the second harmonic is higher than a certain threshold value, for example, when the ratio between the amplitude of the second harmonic of the current and the amplitude of the fundamental wave is greater than or equal to 25%, the current signal is determined to be an inrush current and "lockout protection" should be adopted. However, the second harmonic method has a slow detection speed and low accuracy. Summary of the invention

[0003] The present disclosure relates to a method capable of quickly and accurately detecting an inrush current of a transformer.

[0004] According to one aspect of the present disclosure, a method for inrush current detection of a transformer is provided. The method comprises: sampling at least a portion of a current signal of the transformer to obtain a numerical matrix; providing the numerical matrix as an input to an inrush current detection neural network; and calculating and outputting a label vector corresponding to the numerical matrix by the inrush current detection neural network, the label vector indicating whether the current signal is an inrush current.

[0005] Optionally, the method further includes: detecting whether the maximum value in the numerical matrix is ​​greater than a preset start threshold; and providing the numerical matrix as input to the surge detection neural network only when the maximum value in the numerical matrix is ​​greater than or equal to the preset start threshold.

[0006] Optionally, at least a portion of the current signal is a portion corresponding to a half cycle of the current signal.

[0007] Optionally, the current signal is a three-phase current signal, and a row or a column of the numerical matrix corresponds to one of the three-phase current signals.

[0008] Optionally, the surge detection neural network is a convolutional neural network.

[0009] Optionally, the surge detection neural network is a trained neural network, and the training includes the following steps: Step 1, obtaining a current signal training sample set for training the surge detection neural network; Step 2, providing the training samples in the current signal training sample set as input to the surge detection neural network; Step 3, calculating a label vector corresponding to the training sample by the surge detection neural network; Step 4, determining the processing loss of the surge detection neural network based on the label vector of the training sample; Step 5: if the processing loss is greater than or equal to a preset processing loss threshold, updating the parameters of the surge detection neural network and performing steps 2 to 5 based on the updated surge detection neural network, and if the processing loss is less than or equal to the preset processing loss threshold, stopping the training.

[0010] Optionally, the surge detection neural network includes at least a first convolutional layer, a second convolutional layer, a fully connected layer, an activation layer and an output layer; and the parameters of the surge detection neural network include at least a weight matrix of a first convolutional kernel included in the first convolutional layer, a weight matrix of a second convolutional kernel included in the second convolutional layer, and a weight matrix of the fully connected layer.

[0011] Optionally, the surge detection neural network also includes a first pooling layer located between the first convolutional layer and the second convolutional layer, and a second pooling layer located between the second convolutional layer and the fully connected layer, and the first pooling layer and / or the second pooling layer adopts maximum pooling method for downsampling.

[0012] Optionally, the first convolution layer adopts a first convolution kernel of size 3*3; the second convolution layer adopts a second convolution kernel of size 3*3; and the activation function is one of a sigmoid function, a relu function, and a tanh function.

[0013] Optionally, one or more samples in the current signal training sample set are generated by an electromagnetic transient simulation program EMTP.

[0014] According to another aspect of the present disclosure, an inrush current detection device for a transformer is provided. The device comprises: a sampling module, sampling at least a portion of the current signal of the transformer to obtain a numerical matrix; an input module, providing the numerical matrix as an input to an inrush current detection neural network; and a classification module, the inrush current detection neural network calculating and outputting a label vector corresponding to the numerical matrix, the label vector indicating whether the current signal is an inrush current.

[0015] Optionally, the input module is also configured to detect whether the maximum value in the numerical matrix is ​​greater than a preset start threshold, and provide the numerical matrix as input to the surge detection neural network only when the maximum value in the numerical matrix is ​​greater than or equal to the preset start threshold.

[0016] According to another aspect of the present disclosure, a surge current detection device for a transformer is provided, which includes: a processor; and a memory, wherein a computer executable program is stored in the memory, and when the program is executed by the processor, the device executes any one of the above methods.

[0017] According to another aspect of the present disclosure, a computer-readable storage medium is provided, on which computer instructions are stored. When the computer instructions are executed by a processor, any of the above methods is implemented.

[0018] According to the above aspects of the present disclosure, a neural network is used to detect the inrush current of a transformer, which requires fewer sampling points, has a fast detection speed, and a high accuracy, and can perform detection when inrush current and fault current coexist. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] These and / or other aspects, features and advantages of the present disclosure will become more clear and easily understood through the following description of the present disclosure in conjunction with the accompanying drawings, in which:

[0020] Figure 1 is a diagram illustrating a usage scenario of a surge detection device according to an embodiment of the present disclosure. ;

[0021] Figure 2 is a flow chart of a surge current detection method according to an embodiment of the present disclosure;

[0022] Figure 3 is a flow chart of a surge current detection method according to another embodiment of the present disclosure;

[0023] Figure 4 is a flow chart of a method for training a surge detection neural network according to an embodiment of the present disclosure;

[0024] Figure 5 is a schematic diagram of the structure of a surge detection neural network according to an embodiment of the present disclosure;

[0025] Figure 6 is a test result obtained by testing the inrush current detection neural network trained according to an embodiment of the present disclosure;

[0026] Figure 7 is a hardware block diagram showing a surge current detection device according to an embodiment of the present disclosure;

[0027] Figure 8is a hardware block diagram showing a surge current detection device according to another embodiment of the present disclosure; and

[0028] Fig. 9 is a schematic diagram illustrating a computer-readable storage medium according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0029] The present disclosure will be described in detail below with reference to exemplary embodiments of the present disclosure. However, the present disclosure is not limited to the embodiments described herein, and it can be implemented in many different forms. The described embodiments are only used to make the present disclosure thorough and complete, and to fully convey the concept of the present disclosure to those skilled in the art. The features of the various embodiments described can be combined or replaced with each other, unless explicitly excluded or should be excluded according to the context.

[0030] At present, the main method for inrush current detection is the second harmonic method. Its principle is to use the phenomenon that the inrush current contains rich second harmonic content to identify the appearance and disappearance of the inrush current. When the second harmonic is higher than a certain threshold value, for example, when the ratio between the amplitude of the second harmonic of the current and the amplitude of the fundamental wave is greater than or equal to 25%, the current signal is determined to be an inrush current. However, the second harmonic method has the significant disadvantage of slow detection speed. For example, taking a transformer with an industrial frequency of 50Hz as an example, the period of its current signal is 20ms. In this case, the detection of the second harmonic brings at least 20ms of time delay to the lockout protection action. This is because for any current mutation, the discrete Fourier transform (DFT) will generate higher harmonics including the second harmonic within at least one cycle. These higher harmonics will disappear within 20-30ms, so for some scenarios, it cannot meet the needs of fast lockout protection action. Once the lockout protection mechanism based on the second harmonic method is adopted, the lockout protection action cannot be implemented before the second harmonic disappears. In addition, since almost all current mutations will generate second harmonics, second harmonics may also detect non-inrush current (fault current) as inrush current, so the second harmonic method cannot be used when inrush current and fault current coexist.

[0031] In addition to the second harmonic method, inrush current criteria based on pure voltage are also used, such as the voltage harmonic braking principle and the voltage ratio method; inrush current criteria based on pure current, such as the waveform symmetry principle, the discontinuity angle principle, etc.; criteria based on current and voltage, such as the power differential principle, the flux characteristic principle, the equivalent circuit principle, etc. However, these methods have high usage conditions or are invalid in certain protection scenarios. For example, the discontinuity angle method uses the discontinuity angle that usually appears in the inrush current waveform, while the fault waveform usually does not have a discontinuity angle to perform inrush current detection. However, the discontinuity angle method requires high-resolution sampling, and the number of sampling points per power frequency cycle is at least 36 points. Moreover, since the current of the inrush discontinuity angle is very small and the absolute value is close to zero, and the conversion processing loss of the ADC chip near the zero point is the largest, the discontinuity angle method also requires a high-resolution ADC chip. In addition, in some cases, the current waveform may also be distorted due to high-order harmonics, and the distorted waveform usually does not produce a discontinuity angle, so the discontinuity angle method is no longer valid. For another example, the waveform symmetry method uses the fact that the fault waveform is usually symmetrical about the time axis, while the single-phase inrush current waveform is usually not symmetrical about the time axis to perform inrush current detection. However, in the case of distortion, it is very likely that the originally symmetrical current will become asymmetrical, making this method invalid.

[0032] Therefore, there is a need for a surge detection method that can be used in various protection scenarios and can improve detection speed and accuracy.

[0033] Figure 1 is a diagram illustrating a usage scenario of the surge detection device 100 according to an embodiment of the present disclosure.

[0034] like Figure 1 As shown, the substation supplies power to loads A, B, and C such as business centers and office buildings through transformers. The transformer distributes the voltage to feeders A, B, and C through the busbar to supply power to loads A, B, and C via feeders A, B, and C. There are relays and circuit breakers in each feeder A, B, and C (for simplicity, only feeder A is shown in the figure). When the transformer is in working condition, the voltages on the primary and secondary sides will reach the rated values, such as about 10 kV on the primary side and about 380 V on the secondary side. When the transformer is powered on but has not yet entered the working state, inrush current may occur due to the full voltage charging of the transformer. At this time, the relay is required to take a "locking protection" action instead of a tripping protection action to trip the circuit breaker.

[0035] The surge detection device 100 according to the embodiment of the present disclosure can be integrated into the relay as a part of the relay, or located outside the relay and exchange data or commands with the relay through a wired or wireless network. The surge detection device 100 can sample the current signal I from the transformer, determine whether the current signal I is a surge, and generate a detection result. The relay takes a "lockout protection" action based on the detection result indicating that the current signal I is a surge, otherwise it takes a tripping protection action. The current signal I can be a current signal after conversion by a current transformer, and the function of the current transformer is to convert a larger current value into a smaller current value.

[0036] The inrush current detection device 100 may be configured with an inrush current detection neural network according to an embodiment of the present disclosure, and the inrush current detection neural network may perform feature extraction processing on the current signal I, and generate a detection result indicating whether the current signal I is an inrush current based on the extracted features. The inrush current detection neural network needs to be trained according to the training method of an embodiment of the present disclosure to become a neural network capable of determining whether the current signal I is an inrush current.

[0037] In addition, the surge detection device 100 can also output the surge detection result to an external display device (not shown in the figure). The external display device can communicate with the surge detection device via a wired or wireless network. The display device can display the detection result of the surge detection device 100 in a variety of ways such as text, image, video, etc.

[0038] Since the detection result of the surge current detection device 100 is generated by using a trained surge current detection neural network, based on the significant advantages of neural networks in feature extraction and classification, it overcomes the shortcomings of conventional detection methods such as the slow second harmonic detection speed and the inability to distinguish between surge current and fault current, the high sampling resolution of the intermittent angle method and possible invalidity, and the waveform symmetry method may be invalid, thereby achieving more efficient surge current detection.

[0039] Figure 2 is a flow chart of a surge detection method 200 according to an embodiment of the present disclosure.

[0040] Figure 2 The surge detection method 200 shown in FIG. 2 may be performed by Figure 1 The inrush current detection device 100 is implemented in the inrush current detection device 100, and the inrush current detection neural network trained according to the embodiment of the present disclosure is configured in the inrush current detection device 100. Figure 2 As shown, the surge detection method 200 may include steps S201 - S203 .

[0041] In step S201, at least a portion of the current signal I of the transformer is sampled to obtain a numerical matrix M. The current signal I in this step may be a current signal that has been appropriately preprocessed after being output from the transformer 101, and the preprocessing includes but is not limited to converting the current signal with a larger value directly output from the transformer 101 into a current signal with a smaller value at a certain ratio through a current transformer. In this step, the sampling resolution when sampling the current signal I is k sampling points per power frequency cycle, and the value of k can be selected according to application needs. According to an embodiment of the present invention, the value of k can be smaller, for example, as low as 32, and the sampling range is N cycles, for example, N can be as low as 1 / 2. The numerical matrix M in this step is an arrangement of the sampled values ​​of the current signal I, wherein a row or a column corresponds to one phase in the current signal. For example, in the case where the current signal I is a three-phase current, sampling N cycles in the current signal I will obtain a numerical matrix M of size 3×Nk, and a row in the numerical matrix M corresponds to one phase in the three-phase current signal.

[0042] In step S202, the numerical matrix M is provided as input to the inrush current detection neural network. The inrush current detection neural network is a trained inrush current detection neural network, and its training method will be described in detail below. Figure 5 Give a description.

[0043] In step S203, the inrush current detection neural network calculates and outputs a label vector V corresponding to the numerical matrix M, and the label vector V indicates whether the current signal is an inrush current. In this step, the inrush current detection neural network performs feature extraction processing on the current signal I and generates a detection result based on the extracted features. In other words, the inrush current detection neural network performs a series of operations on the numerical matrix M input in step S203 to obtain a label vector V indicating whether the current signal I is an inrush current. The category corresponding to the maximum value in the label vector V is the prediction result of the inrush current neural network.

[0044] In the embodiments of the present disclosure, an inrush current detection neural network is used to detect whether the current signal of the transformer is an inrush current. Since neural networks have significant advantages in feature extraction and feature-based classification, the fault current can be accurately judged as a non-inrush current signal even when inrush current and fault current coexist, thereby improving the accuracy of inrush current detection.

[0045] In addition, in the embodiment of the present disclosure, the sampling range of the surge detection device 100 can be as low as N = 1 / 2 cycle, that is, when the power frequency is 50Hz, it can obtain the detection result 10ms after the surge occurs. Compared with the second harmonic method which requires at least 20 to 30ms, the detection time is shortened and the detection speed is improved.

[0046] Furthermore, in the embodiments of the present disclosure, the sampling resolution may be as low as k=32, which reduces the requirement on the sampling resolution compared to the existing detection methods such as the discontinuity angle method, which has a sampling resolution of at least 36.

[0047] Figure 3 is a flow chart of a surge detection method 300 according to another embodiment of the present disclosure.

[0048] There are various reasons that cause the current of the transformer to suddenly increase. For example, during the transformer power-on period, such as when the transformer is put into operation without load or the voltage is restored after the external fault is removed, the inrush current is generated because its magnetic flux cannot change suddenly; or during the operation of the transformer, when an external or internal fault occurs, the current suddenly increases due to the short circuit. At this time, the current mutation belongs to the fault current. However, most of the time, the current of the transformer is within the normal range. In order to save computing resources and reduce power consumption, a start threshold can be set for the start of the inrush current judgment.

[0049] like Figure 3 As shown, the surge detection method 300 may include steps S301-S304. Step S301 in the method 300 is Figure 2 The method is the same as step S201 in the embodiment, except that a start threshold determination operation is added at step 302.

[0050] In step S302, it is determined whether the maximum value in the numerical matrix M obtained in step S301 is greater than a preset start threshold value, where the maximum value corresponds to the maximum current value in the collected current signal. enable is the threshold value for starting the surging neural network for feature extraction and classification. enable It can be set according to actual needs, for example, it can be set to a multiple of the rated current, such as 5 times, 6 times, etc. And the preset start threshold T enable It can be changed artificially, for example, by the user adjusting it according to the power environment in which the transformer is located during use of the transformer.

[0051] If the result of step S302 is that the maximum value in the value matrix M is greater than or equal to the preset start threshold T enable , the method 300 proceeds to step S303. Similar to step S202 of method 200, in step S303, the numerical matrix M is provided as input to the inrush current detection neural network. Similar to step S203 of method 200, in step S304, the inrush current detection neural network calculates and outputs a label vector corresponding to the numerical matrix M, which indicates whether the current signal is an inrush current. If the result of the determination in step S302 is that the maximum value in the numerical matrix M is less than the preset start threshold T enable , then wait for the next numerical matrix M obtained in step 301. It should be understood that Figure 3 Step S301 in the process can be continuously executed, that is, the current signal I of the transformer is continuously collected (for example, each collection range is half a cycle of the current signal I), therefore, steps S302 to S304 are repeated for each numerical matrix obtained in step S301.

[0052] In the embodiments of the present disclosure, by setting the inrush current detection neural network to be started only when a preset start threshold is met, computing resources are saved.

[0053] Figure 4 is a flow chart of a method 400 for training a surge detection neural network according to an embodiment of the present disclosure.

[0054] like Figure 4 As shown, the training method 400 for the inrush current detection neural network includes steps S401-S407.

[0055] In step S401, a current signal training sample set for training the inrush current detection neural network is obtained. In an embodiment of the present disclosure, each sample in the current signal training sample set may be a current signal collected during the actual use of the transformer or a current signal generated by an electromagnetic transient simulation program EMTP or other software program according to a specific algorithm. These current signals may include, for example, signals that have been determined to be inrush current, fault current, normal current, etc. For example, when the current signal that occurs during the actual use of the transformer is used as a sample, it is impossible to obtain the sample quantity required for training the inrush current detection neural network, and the training samples generated by EMTP can be used as a supplement. For each sample in the current signal training sample set, regardless of which of the two aforementioned cases it comes from, the true label vector of the sample is known, and the detection result of the inrush current detection neural network, that is, the label vector output by the inrush current neural network, should be consistent with the true label vector of the sample. Therefore, the goal of training the inrush current detection neural network is to minimize the difference between the label vector calculated by the inrush current detection neural network and its true label vector for each sample in the current signal training sample set.

[0056] In step S402, the training samples in the current signal training sample set are provided as input to the inrush current detection neural network. The training samples input in this step can be all samples, a part of samples or one sample of the training sample set, which can be selected according to the sample size of the sample set, the size of computing resources, the requirements of network prediction accuracy, etc.

[0057] In step S403, the inrush current detection neural network calculates a label vector corresponding to the training sample. As described above, the inrush current detection neural network extracts features from the training sample input in step S402, and classifies the extracted features to obtain a label vector V indicating whether the current signal corresponding to the training sample is an inrush current.

[0058] In step S404, the processing loss of the surge detection neural network is determined based on the label vector of the training sample. For the training sample input in step S402, the label vector output by the surge detection neural network and its known true label vector can be compared. For example, it is known that for a training sample that is actually a surge, its true label vector is [1,0], indicating that the probability of the training sample being a surge is 1, and the probability of being a non-surge is 0; for a training sample that is actually a non-surge, its true label vector is [0,1], indicating that the probability of the training sample being a surge is 0, and the probability of being a non-surge is 1. When the training of the inrush flow detection neural network is not completed, the output training sample label vector may be significantly different from the true label vector of the training sample. For a training sample known to be inrush flow, for example, the output label vector may be [0.3, 0.7], indicating that the neural network believes that the probability of the sample being inrush flow is 0.3 and the probability of being non-inrush flow is 0.7. This detection result is completely wrong. For another example, the output label vector may be [0.6, 0.4], indicating that the neural network believes that the probability of the sample being inrush flow is 0.6 and the probability of being non-inrush flow is 0.4. Although this detection result shows that the probability of inrush flow is greater than that of non-inrush flow, the difference between it and the true label vector is still relatively large, and the training goal of minimizing the difference between the label vector output by the neural network and the true label vector has not yet been achieved. The difference between the label vector output by the inrush flow neural network and the true label vector can be measured by the processing loss L. The determination of the processing loss L is related to the loss function selected for the inrush flow detection neural network. The loss function can be one of the mean square error function, the cross entropy loss function, the L1 loss function, the L2 loss function, etc. The training samples input in step 402 are all samples, a part of samples or one sample of the training sample set, respectively, and the processing loss L can be determined according to batch gradient descent (BGD), mini-batch gradient descent (SGD) or stochastic gradient descent (MBGD).

[0059] In step S405, it is determined whether the processing loss L is greater than or equal to the preset processing loss threshold T L . Preset processing loss threshold T LThe value can be selected in combination with the accuracy of the inrush current detection neural network required by the power environment in which the transformer is located, for example, the value can be 0.01, 0.02, etc. If it is determined that the processing loss L is greater than or equal to the preset processing loss threshold T L , it is considered that the inrush current detection neural network training has not been completed, and the process proceeds to step S406. On the contrary, if it is determined that the processing loss L is less than the preset processing loss threshold T L , it is considered that the inrush current detection neural network training has been completed, and the process proceeds to step S407.

[0060] In step S406, the parameters of the surge detection neural network are updated. This step can be performed using a back propagation algorithm, and one or more parameters in the surge detection neural network are updated so that the processing loss L obtained in the next training is smaller. These parameters mainly include weight parameters of each layer of the neural network, for example, the weight matrix of one or more convolution kernels in the convolution layer, the weight matrix of the fully connected layer, etc. After the update is completed, steps S402 to S405 are repeated until it is determined in step S405 that the processing loss L is less than the preset processing loss threshold T. L , then proceed to step S407.

[0061] In step S407, the training of the inrush current detection neural network is stopped.

[0062] Figure 5 is a schematic diagram of the structure of a surge detection neural network according to an embodiment of the present disclosure.

[0063] like Figure 5 As shown, the inrush current detection neural network may be a convolutional neural network including a first convolutional layer, a second convolutional layer, a fully connected layer, an activation layer, and an output layer. The first convolutional layer may include n convolutional kernels: convolutional kernel 11, convolutional kernel 12, ..., convolutional kernel 1 n The second convolution layer may include m convolution kernels: convolution kernel 21, convolution kernel 22, ..., convolution kernel 2 m The numerical matrix input to the surge detection neural network is M.

[0064] After the numerical matrix M is input into the surge detection neural network, it is first processed by the first convolution layer. The first convolution layer includes the first convolution kernel 11 to 1 n Each of the convolution kernels is convolved with the numerical matrix M to obtain the convolution result of the first convolution layer, that is, the numerical matrix C1. Each row in the numerical matrix C1 corresponds to the convolution result of each convolution kernel and the numerical matrix M.

[0065] When the numerical matrix C1 is large in size, in order to reduce the amount of data to be processed, the numerical matrix C1 can be pooled by the first pooling layer. The pooling method of the first pooling layer can be an average pooling method or a maximum pooling method. The pooling result of the first pooling layer is recorded as the numerical matrix P1.

[0066] Next, the numerical matrix P1 is output to the second convolution layer. The second convolution layer includes the second convolution kernels 21 to 2 m Each of them is convolved with the numerical matrix P1 to obtain the convolution result of the second convolution layer, that is, the numerical matrix C2. Each row in the numerical matrix C2 corresponds to the convolution result of every two convolution kernels and the numerical matrix P1.

[0067] Similarly, when the size of the numerical matrix C2 is large, in order to further reduce the amount of data to be processed, the numerical matrix C2 can be pooled by a second pooling layer. The pooling method of the second pooling layer can also be an average pooling method or a maximum pooling method. The pooling result of the second pooling layer is recorded as a numerical matrix P2.

[0068] Next, the numerical matrix P2 is expanded into a row or a column, and the expanded result is the fully connected layer, which is recorded as the numerical matrix U.

[0069] Next, the numerical matrix U is multiplied by the fully connected layer weight matrix W to obtain the weighted sum vector S.

[0070] Finally, the weighted sum vector S is converted into an output layer label vector V through the activation layer. The category corresponding to the maximum value in the label vector V is the prediction result of the influx neural network. The activation function G(x) of the activation layer may include but is not limited to one of the sigmoid function, the relu function, and the tanh function.

[0071] It should be noted that Figure 5 The specific structure of the surge detection neural network shown is only an example listed for the convenience of explanation. In a specific environment, changes and modifications can be made to the structure without departing from the scope of protection of the present disclosure. For example, when the computing resources are powerful enough or the amount of numerical data to be processed is small, it is not necessary to include the first and / or second pooling layer; the number of convolutional layers, pooling layers, activation layers, etc. can be appropriately modified according to the training of the surge detection neural network; the number and size of the convolution kernels included in each convolutional layer can also be modified; the size of the convolution kernels included in each convolutional layer can also be different, etc.

[0072] For ease of understanding, the following is a detailed example Figure 5 The surge neural network shown is based on Figure 4 The training process of the training method shown in the figure, and the process of using the trained inrush current neural network to detect whether the current signal I is an inrush current.

[0073] For the convenience of explanation, the following settings are made for this example: the sampling resolution is k=32, the sampling range for each time is N=1 / 2 cycle, the number of convolution kernels contained in the first convolution layer and the second convolution layer of the inrush detection neural network are both 6, that is, n=m=6, the size of the first convolution kernel 11 to 16 and the size of the second convolution kernel 21 to 26 are both 3×3, the first and second pooling layers use the maximum pooling method for downsampling, and the processing loss function uses the cross entropy loss function Processing loss threshold T L is 0.01, and the activation layer uses the activation function σ(z)=1 / (1+e -z ).

[0074] Taking this example, Figure 5 The surge neural network shown is based on Figure 4 The training process of the training method shown.

[0075] Corresponding to step S401, EMTP is used to generate 5000 numerical matrices representing sampled current signals as a training sample set, the training sample set includes 2500 numerical matrices corresponding to inrush current signals and 2500 numerical matrices corresponding to non-inrush current signals, the non-inrush current signals may include CT saturation current signals, fault current signals, or current signals of inrush current and CT saturation coexisting, etc. These 5000 numerical matrices are randomly sorted and numbered M1 to M2 in sequence. 5000 For the surge signal sample, its true label vector is [1,0], which means that the probability of the sample being a surge signal is 1, and the probability of being a non-surge signal is 0; for the non-surge signal sample, its true label vector should be [0,1], which means that the probability of the sample being a surge signal is 0, and the probability of being a non-surge signal is 1.

[0076] Corresponding to step S402, the training sample M1 is provided to the inrush current detection neural network. It is known that M1 is an inrush current signal sample, and its true label vector is [1,0].

[0077] Corresponding to step S403, the surge detection neural network outputs a label vector y1 corresponding to M1, for example, y1 = [0.4, 0.6], that is, the surge detection neural network at this time believes that the probability that sample M1 belongs to a surge signal is 0.4, and the probability that it belongs to a non-surge signal is 0.6. This means that the classification of sample M1 by the surge detection neural network at this time is completely wrong.

[0078] Corresponding to step S404, based on the loss function The surge detection neural network can determine that its processing loss L1 = -(1×log0.4+0×log0.6) = 0.91.

[0079] Corresponding to step S405, the surge detection neural network determines that its processing loss L1=0.91≥T L =0.01.

[0080] Corresponding to step S406, the surge detection neural network updates its parameters according to the back propagation algorithm, which include a 3×3 weight matrix of the first convolution kernels 11 to 16, a 3×3 weight matrix of the second convolution kernels 21 to 26, and a 2×12 weight matrix W of the fully connected layer.

[0081] After the update is completed, steps S402 to S406 are repeated, that is, the training sample M2 is provided to the inrush current detection neural network, and then feature extraction, classification, and processing loss calculation of the training sample M2 are performed. Until the processing loss L is determined in step S405 <T L , the training ends.

[0082] After the training is completed, the determined 6×3×3 weight matrix corresponding to the 6 first convolution kernels of the first convolution layer, the determined 6×3×3 weight matrix corresponding to the 6 second convolution kernels of the second convolution layer, and the determined 2×12 weight matrix corresponding to the weight matrix W of the fully connected layer will be obtained. These parameters together with the previously set structure, loss function, activation function, etc. of the surge detection neural network constitute the trained surge detection neural network.

[0083] It should be noted that the above simple examples are only for the convenience of explanation and do not limit the present disclosure. Various modifications and changes can be made to them according to actual needs. For example, in step S402, multiple samples can be output each time, and the processing loss L calculated in step S404 can be the average value of the processing losses of these samples; for example, more complexly, the inrush current detection neural network can also be trained to perform multi-classification neural networks, such as enabling it to distinguish between four current signals: inrush current, CT saturation current, fault current, and current in which inrush current and CT saturation coexist. In this case, the training samples are divided into 4 categories, and the corresponding real label vectors are [1,0,0,0], [0,1,0,0], [0,0,1,0], and [0,0,0,1]. The structure of the inrush current detection neural network, the loss function, etc. can also be adjusted accordingly. The present disclosure will not repeat other modifications and changes.

[0084] The following continues to use this example to illustrate the process of using the trained inrush current detection neural network to detect whether the current signal I is an inrush current.

[0085] refer to Figure 1 and Figure 5When the inrush current detection device detects the three-phase current signal I of the transformer converted by the current transformer (the waveform of each phase is shown as Iph1, Iph2, and Iph3 respectively), it samples N=1 / 2 cycles of the current signal according to the sampling resolution k=32 to obtain a numerical matrix M of size 3×16, in which one row of M represents the sampling value of one phase.

[0086] After M is input into the surge detection neural network, the six first convolution kernels 11 to 16 contained in the first convolution layer are convolved with it respectively. The convolution result of each convolution kernel and M is a numerical vector including 16 numerical values. Therefore, the convolution result of the six first convolution kernels and M is a numerical matrix C1 of size 6×16.

[0087] To reduce the amount of numerical data to be processed, the numerical matrix C1 is pooled through the first pooling layer. In this example, the maximum pooling method is used to downsample the numerical matrix C1 to obtain a numerical matrix P1 of size 3×8.

[0088] Next, each of the six second convolution kernels 21 to 26 included in the second convolution layer is convolved with the numerical matrix P1. Similar to the first convolution layer, the convolution result of the second convolution layer is a numerical matrix C2 of size 6×8.

[0089] To further reduce the amount of numerical data to be processed, the numerical matrix C2 is pooled through the second pooling layer. In this example, the maximum pooling method is used to downsample the numerical matrix C2 to obtain a 3×4 numerical matrix P2.

[0090] Next, the numerical matrix P2 is expanded into a row or a column, that is, the numerical values ​​in the matrix P2 are arranged into a row or a column. In this example, the matrix P2 is expanded into a column to obtain a numerical matrix U of size 12×1.

[0091] Next, the fully connected layer weight matrix W of size 2×12 is multiplied by the 12×1 numerical matrix U to obtain a weighted sum vector S of size 2×1.

[0092] Finally, the weighted sum vector S is converted into a label vector V through the activation layer, so that the value in the label vector V is in the range of 0 to 1. The category corresponding to the maximum value in the label vector V is the prediction result of the inrush neural network. For example, if the output label vector V is [0.9956, 0.0044], it means that the probability that the inrush neural network believes that the detected current signal I is an inrush is 0.9956, and the probability that the detected current signal I is a non-inrush is 0.0044. The category corresponding to the maximum value 0.9956 in the vector V is inrush, indicating that the prediction result of the inrush neural network is inrush.

[0093] In addition, Figure 4After the training process shown is completed, the trained inrush current detection neural network can be tested, for example, using current signal samples that do not participate in the training process as test samples to test the detection effect of the inrush current detection neural network, and the true label vector of each test sample is known.

[0094] Figure 6 FIG. 2 shows the test results of the inrush current detection neural network according to the embodiment of the present disclosure. Figure 6 The left side of the figure shows a surge signal sample in the test sample set, where the horizontal axis is the index of the sampling point and the vertical axis is the current value in A. Figure 4 and Figure 5 Similar to the example described above, the inrush signal sample is sampled for half a cycle at a sampling resolution of k=32 to obtain a numerical matrix of size 3×16, which is input into the inrush detection neural network. After calculation, the inrush detection neural network outputs a label vector of [0.9964, 0.0036], and the prediction result is "Pred: inrush" (i.e., inrush). Figure 6 The right side of the figure shows a non-inrush signal in the test sample set, where the horizontal axis is the index of the sampling point and the vertical axis is the current value in A. Similarly, the inrush signal is sampled for half a cycle at a sampling resolution of k=32 to obtain a numerical matrix of size 3×16, which is input into the inrush detection neural network. After calculation, the inrush detection neural network outputs a label vector of [0.0052, 0.9948], and displays the final detection result as "Pred: Non_I" (i.e., non-inrush).

[0095] The present disclosure has demonstrated through the test of the above exemplary convolutional neural network that the inrush current detection method according to the embodiment of the present disclosure can achieve accurate detection through a sampling range as low as N=1 / 2 cycle and a resolution as low as 32 per cycle. Obviously, those skilled in the art will understand that the inrush current detection method of the present disclosure is not limited to the use of the above exemplary convolutional neural network, and other suitable neural networks may also be used.

[0096] Figure 7 is a module block diagram showing a surge detection device 700 according to an embodiment of the present disclosure.

[0097] like Figure 7 As shown, the surge detection device 700 includes a sampling module 701, an input module 702, and a classification module 703. These three modules can be configured in the same physical location, for example, they are all integrated in Figure 1Alternatively, without considering the time delay caused by data communication, one or two of the three modules can also be configured to be not in the same physical location as the other two or one, and connected through a wired or wireless communication network to transmit data or commands between each other.

[0098] The sampling module 701 is configured to sample at least a portion of the transformer current signal I to obtain a numerical matrix M. The sampling resolution is k sample points per power frequency cycle, and the sampling range of each time may be N cycles of the current signal I, N≥1 / 2.

[0099] The input module 702 provides the numerical matrix M obtained by the sampling module 701 as input to the inrush current detection neural network. In addition, in order to save computing resources, the input module 702 can also detect whether the maximum value in the numerical matrix M is greater than a preset start threshold, and only when the maximum value in the numerical matrix is ​​greater than or equal to the preset start threshold, the numerical matrix is ​​provided as input to the inrush current detection neural network.

[0100] The classification module 703 is configured with a surge current detection neural network according to an embodiment of the present disclosure, and the surge current detection neural network outputs a label vector corresponding to the numerical matrix M, where the label vector indicates whether the current signal is a surge current.

[0101] Obviously, the above description on the surge current detection method is also applicable to the surge current detection device 700 and will not be repeated here.

[0102] Figure 8 is a hardware block diagram showing a surge detection device 800 according to another embodiment of the present disclosure.

[0103] The electronic device 800 according to the embodiment of the present disclosure includes at least a memory for storing computer program instructions and a processor for executing the computer program instructions. When the computer program instructions are loaded and run by the processor, the processor executes the training method of the inrush current detection neural network and the inrush current detection method.

[0104] Figure 8 The surge detection device 800 shown may specifically include: a central processing unit (CPU) 801, a graphics processing unit (GPU) 802, and a main memory 803. These units are interconnected via a bus 804. The central processing unit (CPU) 801 and / or the graphics processing unit (GPU) 802 may be used as the above-mentioned processor, and the main memory 803 may be used as the above-mentioned memory for storing computer program instructions. In addition, the electronic device 800 may also include a communication unit 805, a storage unit 806, an output unit 807, an input unit 808, and an external device 809, which are also connected to the bus 804.

[0105] Fig. 9 is a schematic diagram showing a computer-readable storage medium according to an embodiment of the present disclosure. Fig. 9 As shown, a computer readable storage medium 900 according to an embodiment of the present disclosure stores computer program instructions 901. When the computer program instructions 901 are executed by a processor, the training method of the surge detection neural network and the surge detection method according to the embodiment of the present disclosure described with reference to the above figures are executed.

[0106] The computer readable storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, optical disk, magnetic disk, etc.

[0107] In the above, the inrush current detection method and device based on a neural network according to the embodiment of the present disclosure are described with reference to the accompanying drawings. Based on the powerful feature extraction and classification functions of the neural network, the inrush current detection method, inrush current detection device and computer-readable storage medium using the neural network can quickly and accurately determine whether the current signal of the transformer is an inrush current. And the inrush current detection result can be provided to the relay for it to make a decision on whether to take "lockout protection".

[0108] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0109] The basic principles of the present disclosure are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in the present disclosure are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are required by each embodiment of the present disclosure. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, and are not limitations. The above details do not limit the present disclosure to the necessity of adopting the above specific details to be implemented.

[0110] The block diagrams of the devices, apparatuses, equipment, and systems involved in this disclosure are only illustrative examples and are not intended to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagrams. As will be appreciated by those skilled in the art, these devices, apparatuses, equipment, and systems can be connected, arranged, and configured in any manner. Words such as "including," "comprising," "having," and the like are open words, referring to "including but not limited to," and can be used interchangeably therewith. The words "or" and "and" used herein refer to the words "and / or," and can be used interchangeably therewith, unless the context clearly indicates otherwise. The word "such as" used herein refers to the phrase "such as but not limited to," and can be used interchangeably therewith.

[0111] Additionally, as used herein, "or" used in a list of items beginning with "at least one" indicates a separate list, so that, for example, a list of "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not mean that the example described is preferred or better than other examples.

[0112] It should also be noted that in the system and method of the present disclosure, each component or each step can be decomposed and / or recombined. Such decomposition and / or recombination should be regarded as equivalent solutions of the present disclosure.

[0113] Various changes, substitutions, and modifications of the techniques described herein may be made without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of the present disclosure is not limited to the specific aspects of the processes, machines, manufactures, compositions of events, means, methods, and actions described above. Currently existing or later to be developed processes, machines, manufactures, compositions of events, means, methods, or actions that perform substantially the same functions or achieve substantially the same results as the corresponding aspects described herein may be utilized. Thus, the appended claims include such processes, machines, manufactures, compositions of events, means, methods, or actions within their scope.

[0114] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present disclosure. Therefore, the present disclosure is not intended to be limited to the aspects shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

[0115] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present disclosure to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A method for detecting inrush current of a transformer, comprising: sampling at least a portion of the current signal of the transformer to obtain a numerical matrix; Detecting whether the maximum value in the numerical matrix is ​​greater than a preset start threshold; Only when the maximum value in the numerical matrix is ​​greater than or equal to the preset start threshold, providing the numerical matrix as input to the inrush current detection neural network; as well as The inrush current detection neural network calculates and outputs a label vector corresponding to the numerical matrix, wherein the label vector indicates whether the current signal is an inrush current.

2. The method of claim 1, wherein At least a portion of the current signal is a portion corresponding to a half cycle of the current signal.

3. The method of claim 1, wherein The current signal is a three-phase current signal, and one row or one column of the numerical matrix corresponds to one of the three-phase current signals.

4. The method of claim 1, wherein The inrush current detection neural network is a convolutional neural network.

5. The method of claim 1, wherein The inrush current detection neural network is a trained neural network, and the training comprises the following steps: Step 1, obtaining a current signal training sample set for training the inrush current detection neural network; Step 2, providing the training samples in the current signal training sample set as input to the inrush current detection neural network; Step 3, calculating a label vector corresponding to the training sample by the inrush current detection neural network; Step 4, determining the processing loss of the inrush current detection neural network based on the label vector of the training sample; Step 5: If the processing loss is greater than or equal to a preset processing loss threshold, update the parameters of the surge detection neural network and perform steps 2 to 5 based on the updated surge detection neural network, and if the processing loss is less than or equal to the preset processing loss threshold, stop the training.

6. The method of claim 5, wherein The inrush current detection neural network comprises at least a first convolutional layer, a second convolutional layer, a fully connected layer, an activation layer, and an output layer; and The parameters of the surge detection neural network include at least a weight matrix of a first convolution kernel included in the first convolution layer, a weight matrix of a second convolution kernel included in the second convolution layer, and a weight matrix of the fully connected layer.

7. The method of claim 6, wherein The surge detection neural network further includes a first pooling layer located between the first convolutional layer and the second convolutional layer, and a second pooling layer located between the second convolutional layer and the fully connected layer. The first pooling layer and / or the second pooling layer performs downsampling using a maximum pooling method.

8. The method of claim 6, wherein The first convolution layer uses a first convolution kernel of size 3*3; The second convolution layer uses a second convolution kernel of size 3*3; and The activation function of the activation layer is one of a sigmoid function, a relu function, and a tanh function.

9. The method of claim 5, wherein One or more samples in the current signal training sample set are generated by an electromagnetic transient simulation program EMTP.

10. A surge current detection device for a transformer, comprising: A sampling module, sampling at least a portion of the current signal of the transformer to obtain a numerical matrix; An input module detects whether a maximum value in the numerical matrix is ​​greater than a preset start threshold, and only when the maximum value in the numerical matrix is ​​greater than or equal to the preset start threshold, provides the numerical matrix as an input to the inrush current detection neural network; as well as A classification module is configured with the inrush current detection neural network, wherein the inrush current detection neural network calculates and outputs a label vector corresponding to the numerical matrix, wherein the label vector indicates whether the current signal is an inrush current.

11. A surge current detection device for a transformer, comprising: processor; as well as A memory, wherein a computer executable program is stored in the memory, and when the program is executed by the processor, the device executes the method according to any one of claims 1 to 9.

12. A computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions, when executed by a processor, implement the method according to any one of claims 1 to 9.

Images