A method for optical fiber differential protection of photovoltaic power station transmission lines based on A-CNN

Abstract

Different from the fault current characteristics of synchronous generators, the high controllability and weak feedback characteristics of photovoltaic power stations change the amplitude and phase angle characteristics of the fault current of the photovoltaic power station transmission line, resulting in the performance degradation or even incorrect operation of the traditional optical fiber differential protection designed based on the fault characteristics of synchronous generators. In order to solve this problem, the present invention discloses an optical fiber differential protection algorithm for photovoltaic power station transmission lines based on data-enhanced convolutional neural network (A‑CNN). The algorithm can correctly identify faults and fault phases within the area, and is not affected by factors such as AC system parameters, fault location, fault time, fault type, fault resistance, severe current transformer saturation, current transformer measurement error and abnormal values, and meets the four requirements of relay protection.

Description

Technical Field

[0001] The invention relates to an optical fiber differential protection method for a photovoltaic power station transmission line based on a data-enhanced convolutional neural network (A-CNN), and belongs to the technical field of electrical engineering relay protection. Background Art

[0002] Fiber optic differential protection is the main protection for high-voltage transmission lines of 220kV and above in my country. It can correctly distinguish between internal and external faults, has its own phase selection function, and has good selectivity and sensitivity. However, the access of photovoltaic power stations has changed the fault behavior of the power grid, resulting in the problem of principle incompatibility of traditional fiber optic current differential protection. When an internal fault occurs, the fiber optic differential protection may have a decrease in sensitivity or even refuse to operate. Considering the influence of current transformer saturation, fiber optic differential protection is prone to false operation when an external fault occurs. As the first line of defense of the power system, the incorrect operation of relay protection affects the stability of the power system. Proposing a new relay protection method that can adapt to the access of photovoltaic power stations is of great significance to ensuring the safe operation of the new power system and achieving the national "dual carbon" goals.

[0003] With the further integration of science and technology such as big data, artificial intelligence, 5G communication, Internet of Things, cloud computing and digital twins with power systems, the power grid is transforming and upgrading towards digitalization and intelligence. The new power system must be a multi-dimensional interactive integrated energy system driven by massive data. It acquires massive data through data collection, performs data processing to restore digital information to the operating status of the power grid, and conducts in-depth analysis and model learning to obtain relevant technical indicators. Deep learning is a technology that can be applied to smart grids, and its combination with relay protection has broad application prospects. Summary of the invention

[0004] Aiming at the problem that the action performance of optical fiber differential protection is degraded or even incorrect after the photovoltaic power station is connected, a new improved scheme is proposed. The convolutional neural network is used to identify the difference in the current phase angle difference waveform at both ends of the line, and the function of identifying the fault inside and outside the area, the fault phase and the non-fault phase is realized. Due to the good generalization ability of A-CNN, the performance evaluation shows that the scheme is not affected by factors such as AC system parameters, fault type, fault location, fault resistance, severe current transformer saturation, current transformer measurement error and outliers.

[0005] A method for optical fiber differential protection of photovoltaic power station transmission lines based on A-CNN includes the following four parts:

[0006] Part 1: Current phase angle difference sampling link on both sides of the line;

[0007] The specific steps include:

[0008] Step (1): Sample the three-phase current at both ends of the photovoltaic power station transmission line to obtain the photovoltaic side current phasor and the system side current phasor

[0009] Step (2): Perform discrete Fourier transform on the three-phase current at both ends of the line to calculate the current phase angle on the photovoltaic side and the system side current phase angle x=a, b, c represent phase A, phase B and phase C respectively;

[0010] Step (3): Subtract the phase angles of the currents on both sides of the same phase to obtain the current phase angle difference on both sides of the line:

[0011] Part II: Dataset generation and data preprocessing;

[0012] The specific steps include:

[0013] Step (1): For a 50 Hz power system, select the current sampling values ​​on both sides of the line within a period of 20 ms after the fault occurs, calculate the current phase difference between the photovoltaic side and the system side, and form a waveform image;

[0014] Step (2): forming a large number of data sets using circuit model simulation data and target line historical real data, where the circuit model simulation data and target line historical real data cover various fault conditions, including different AC system parameters, fault locations, fault times, fault types, fault resistances, severe current transformer saturation, current transformer measurement errors, and abnormal values;

[0015] Step (3): Rotate, translate, scale and add Gaussian noise to the data set, unify the image resolution, convert it into a grayscale image, and perform normalization;

[0016] Part 3: A-CNN model learning and training;

[0017] The specific steps include:

[0018] Step (1): Divide the dataset into a training set and a test set in a ratio of 8:2;

[0019] Step (2): Send the processed data set to the A-CNN neural network model, train it through forward propagation and back propagation, and calculate the accuracy and loss function until convergence;

[0020] Part 4: Real-time monitoring and protection links;

[0021] The specific steps include:

[0022] Step (1): Take the calculated value of the current phase angle difference on both sides of the line in the cycle before time t, construct a current phase angle difference waveform image, change the image resolution, and perform normalization processing;

[0023] Step (2): At each moment, the three-phase current phase angle difference waveform generated at that moment is sent to A-CNN for real-time fault identification, distinguishing between internal and external faults, and determining the fault phase.

[0024] The beneficial effects of the present invention are:

[0025] 1. Based on the traditional optical fiber differential protection scheme, the present invention adopts the current phase difference waveform at both ends of the line and combines it with A-CNN model training. The method is simple and easy to implement.

[0026] 2. After performance evaluation, the protection method proposed in the present invention can realize the correct positioning of intra-zone faults and extra-zone faults, fault phases and non-fault phases, and is not affected by factors such as AC system parameters, fault location, fault time, fault type, fault resistance, severe current transformer saturation, current transformer measurement errors and abnormal values, and meets the four requirements of relay protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 The overall schematic diagram of the optical fiber differential protection method for the outgoing line of a photovoltaic power station based on A-CNN;

[0028] Figure 2 This is a flow chart of the optical fiber differential protection method for photovoltaic power station transmission lines based on A-CNN;

[0029] Figure 3 Schematic diagram of data window fault identification based on dynamic coverage rules. DETAILED DESCRIPTION

[0030] The present invention will be further described below in conjunction with the accompanying drawings.

[0031] A method for optical fiber differential protection of photovoltaic power station transmission lines based on A-CNN includes the following four parts:

[0032] Part 1: Current phase angle difference sampling link on both sides of the line;

[0033] The specific steps include:

[0034] Step (1): Sample the three-phase current at both ends of the photovoltaic power station transmission line to obtain the photovoltaic side current phasor and the system side current phasor

[0035] Step (2): Figure 1 and Figure 2, perform discrete Fourier transform on the three-phase current at both ends of the line, and calculate the current phase angle φpx on the photovoltaic side and the current phase angle on the system side x=a, b, c represent phase A, phase B and phase C respectively;

[0036]

[0037] In the formula, arg() represents the phase angle of the phase variable. is the current phase angle on the photovoltaic side, is the current phase angle on the system side, x=a, b, c represent phase A, phase B and phase C respectively, N is the number of sampling points in one power frequency cycle, i px (k) is the kth sampling current on the photovoltaic side, i gx (k) is the kth sampling current on the system side;

[0038] Step (3): If Figure 1 and Figure 2 , the phase angle difference of the current on both sides of the line is obtained by subtracting the phase angle of the current on both sides of the same phase:

[0039] Part II: Dataset generation and data preprocessing;

[0040] The specific steps include:

[0041] Step (1): Figure 1 and Figure 2 ,For 50Hz power system, the current sampling values ​​on both sides of the line within 20ms after the fault occurs are selected, and the current phase angle difference between the photovoltaic side and the system side is calculated to form a waveform image;

[0042] Step (2): Figure 1 and Figure 2 , a large number of data sets are formed by collecting data from circuit model simulation and historical real data of target lines. The data collected from circuit model simulation and historical real data of target lines cover various fault conditions, including different AC system parameters, fault locations, fault times, fault types, fault resistances, severe current transformer saturation, current transformer measurement errors and abnormal values;

[0043] Step (3): If Figure 1 and Figure 2 , the data set is processed by rotating, translating, scaling and adding Gaussian noise, the image resolution is unified, converted into grayscale images, and normalized;

[0044] Part 3: A-CNN model learning and training;

[0045] The specific steps include:

[0046] Step (1): Figure 1 and Figure 2 , the data set is divided into training set and test set in a ratio of 8:2;

[0047] Step (2): Figure 1 and Figure 2 , the processed data set is sent to the A-CNN neural network model, trained through forward propagation and back propagation, and the accuracy and loss function are calculated until convergence;

[0048] The convolutional neural network is used for model training. Its main structure is as follows: the convolution layer contains 6 5×5 convolution kernels, the convolution step is 1, all zero padding and batch normalization are used, and the Relu activation function is used; the pooling layer uses a 2×2 pooling kernel, the pooling step is 2, all zero padding and maximum pooling mode are used, and Dropout is used to alleviate overfitting, with a ratio of 0.2. The output layer contains 2 neurons and uses Softmax activation;

[0049] Among them, the Relu activation function expression is:

[0050] f(x)=max(0,x) (2)

[0051] In the formula, x and f(x) represent input and output respectively;

[0052] The Softmax function expression is:

[0053]

[0054] Where exp(x) represents e x The Softmax function means that if there are n neurons in the output layer, the output y of the kth neuron is calculated. k ; The numerator of the Softmax function is the input signal a k The exponential function of , the denominator is the sum of the exponential functions of all input signals;

[0055] Part 4: Real-time monitoring and protection links;

[0056] The specific steps include:

[0057] Step (1): Figure 1 , Figure 2 and Figure 3 , take the calculated value of the current phase angle difference on both sides of the line in the cycle before time t, construct the current phase angle difference waveform image, change the image resolution, and perform normalization processing;

[0058] The essence of continuous recording is composed of dot matrix with a certain sampling frequency. Assuming the sampling frequency is 2kHz, 40 data points are sampled in one power frequency cycle T, that is, sampling is performed every 500μs. The data window recording with a width of 20ms (one cycle) moves backward every 500μs to construct the current phase angle difference waveform image. After image preprocessing, it is sent to the trained A-CNN model for fault identification.

[0059] Step (2): At each moment, the three-phase current phase angle difference waveform generated at that moment is sent to A-CNN for real-time fault identification, distinguishing between internal and external faults, and determining the fault phase.

[0060] A-CNN has good generalization ability and fault feature learning ability. It can accurately identify faults within the zone and determine the fault phase under various conditions such as different AC system parameters, different fault conditions, current transformer saturation, current transformer measurement errors and outliers, meeting the reliability requirements of relay protection.

[0061] The data set uses a complete waveform diagram of one cycle (20ms) after the fault occurs. Due to the good generalization ability of A-CNN, fault identification does not require a complete 20ms fault waveform after the fault occurs, which can meet the speed requirement.

[0062] The above description is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements, substitutions, modifications and embellishments can be made without departing from the principle and purpose of the present invention. These improvements, substitutions, modifications and embellishments should also be regarded as the scope of protection of the present invention.

[0063] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. A method for optical fiber differential protection of photovoltaic power station transmission lines based on A-CNN, characterized in that: It includes the following four parts: Part 1: Current phase angle difference sampling link on both sides of the line; The specific steps include: Step (1): Sample the three-phase current at both ends of the photovoltaic power station transmission line to obtain the photovoltaic side current phasor and the system side current phasor Step (2): Perform discrete Fourier transform on the three-phase current at both ends of the line to calculate the current phase angle on the photovoltaic side and the system side current phase angle x=a, b, c represent phase A, phase B and phase C respectively; Step (3): Subtract the phase angles of the currents on both sides of the same phase to obtain the current phase difference on both sides of the line: Part II: Dataset generation and data preprocessing; The specific steps include: Step (1): For a 50 Hz power system, select the current sampling values ​​on both sides of the line within a period of 20 ms after the fault occurs, calculate the current phase difference between the photovoltaic side and the system side, and form a waveform image; Step (2): forming a large number of data sets using circuit model simulation data and target line historical real data, where the circuit model simulation data and target line historical real data cover various fault conditions, including different AC system parameters, fault locations, fault times, fault types, fault resistances, severe current transformer saturation, current transformer measurement errors, and abnormal values; Step (3): Rotate, translate, scale and add Gaussian noise to the data set, unify the image resolution, convert it into a grayscale image, and perform normalization; Part 3: A-CNN model learning and training; The specific steps include: Step (1): Divide the dataset into a training set and a test set in a ratio of 8:2; Step (2): Send the processed data set to the A-CNN neural network model, train it through forward propagation and back propagation, and calculate the accuracy and loss function until convergence; Part 4: Real-time monitoring and protection links; The specific steps include: Step (1): Take the calculated value of the current phase angle difference on both sides of the line in the cycle before time t, construct a current phase angle difference waveform image, change the image resolution, and perform normalization processing; Step (2): At each moment, the three-phase current phase angle difference waveform generated at that moment is sent to A-CNN for real-time fault identification, distinguishing between internal and external faults, and determining the fault phase.

2. According to claim 1, a photovoltaic power station transmission line optical fiber differential protection method based on A-CNN is characterized in that: The current phase angle difference sampling link on both sides of the line described in the first part, the current phase angle difference The calculation formula is: In the formula, arg() represents the phase angle of the phase variable. is the current phase angle on the photovoltaic side, is the current phase angle on the system side, x=a, b, c represent phase A, phase B and phase C respectively, N is the number of sampling points in one power frequency cycle, i px (k) is the kth sampling current on the photovoltaic side, i gx (k) is the kth sampling current on the system side.

3. According to claim 1, a method for optical fiber differential protection of photovoltaic power station transmission lines based on A-CNN is characterized in that: The specific network structure of the A-CNN model learning and training phase described in the third part is as follows: The convolution layer contains 6 5×5 convolution kernels, the convolution step size is 1, all zero padding and batch normalization are used, and the Relu activation function is used; the pooling layer uses a 2×2 pooling kernel, the pooling step size is 2, all zero padding and maximum pooling mode are used, and Dropout is used to alleviate overfitting, the ratio is 0.2, and the output layer contains 2 neurons and uses Softmax activation; Among them, the Relu activation function expression is: f(x)=max(0,x) (3) In the formula, x and f(x) represent input and output respectively; The Softmax function expression is: Where exp(x) represents e x The Softmax function means that if there are n neurons in the output layer, the output y of the kth neuron is calculated. k ; The numerator of the Softmax function is the input signal a k The denominator is the sum of the exponential functions of all input signals.

4. According to claim 1, a photovoltaic power station transmission line optical fiber differential protection method based on A-CNN is characterized in that: The specific meaning of the real-time monitoring and protection link described in Part 4 is: The essence of continuous recording is composed of a dot matrix with a certain sampling frequency. Assuming the sampling frequency is 2kHz, 40 data points are sampled within one power frequency cycle T, that is, sampling is performed every 500μs. The data window recording with a width of 20ms (one cycle) moves backward every 500μs to construct the current phase difference waveform image and send it to the trained A-CNN model for fault identification.

Images