Machine learning-based ultra-high voltage direct current transmission locking fault diagnosis method and system

Abstract

The invention discloses a machine learning-based ultra-high-voltage direct-current transmission locking fault diagnosis method and system, and the method comprises the steps: obtaining various electrical quantity change data of each network node before and after a fault, screening out feature electrical quantities and feature nodes from the electrical quantity change data, and forming fault feature data; extracting a fault occurrence time point and a frequency component from the fault feature data through wavelet transform, constructing a fault signal model, and judging a fault position by adopting an extreme point detection method; constructing an input sequence according to the extracted fault occurrence time point and frequency component, and constructing a convolutional neural network model; expanding the input sequence into a time feature matrix, redistributing the weight of each feature in the time feature matrix by adopting a channel attention mechanism, and performing weighted average on the input time sequence of each time step length in the time feature matrix by adopting a time attention mechanism to obtain an improved convolutional neural network model; and training the improved convolutional neural network model, and extracting fault features of extra-high voltage direct current locking to complete fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of high-voltage DC transmission system control, and more specifically, relates to a UHV DC transmission blocking fault diagnosis method and system based on machine learning. Background Art

[0002] With the proposal of the "dual carbon" goal and the increase in new energy power generation methods such as wind power and solar energy, the power grid structure has become more complex and diversified. At the same time, due to the smaller loss and higher efficiency of DC transmission in long-distance transmission, it has become an important means for new energy grid connection and cross-regional power allocation. However, with the wide application of DC transmission systems, the incidence rate and impact range of their blocking faults have also increased, bringing new challenges to the stable operation of the power grid. Therefore, it is urgent to optimize the existing blocking fault diagnosis methods.

[0003] Among the current UHV DC fault diagnosis methods, some are based on the fault mechanism of the transmission system and the equivalent of physical models, and some propose methods based on signal decomposition to extract fault features. However, in the face of complex and changeable fault situations, they show problems such as low diagnostic accuracy, poor real-time performance, and insufficient generalization ability. As an advanced intelligent algorithm, deep learning has self-learning, self-adaptive, and powerful feature extraction capabilities, which can effectively improve the accuracy and real-time performance of fault diagnosis.

[0004] The authorized prior art document 1 (CN 115498677 B) discloses a DC blocking detection and control method based on a long short-term memory network (LSTM). The deficiencies of the prior art document 1 are as follows:

[0005] 1. Low fault recognition accuracy: Although LSTM can process time series data, its automatic feature extraction ability is insufficient, and it cannot comprehensively capture the complex and diverse features of DC blocking faults, resulting in a decrease in fault recognition accuracy.

[0006] 2. Limited anti-interference ability: When LSTM processes data containing noise, it may not be able to effectively screen out key information, resulting in a decline in fault diagnosis performance in complex environments.

[0007] 3. Strong dependence on labeled data: The LSTM model usually requires a large amount of labeled data for training to ensure that the model can learn accurate fault features. It may not achieve the ideal training effect when the labeled data is limited, increasing the workload and cost of data collection and label annotation. Summary of the Invention

[0008] In order to solve the deficiencies in the prior art, the present invention provides a UHV DC transmission lockout fault diagnosis method and system based on machine learning, which improves the accuracy and real-time performance of UHV DC transmission lockout fault diagnosis.

[0009] The present invention adopts the following technical solution.

[0010] A first aspect of the present invention provides a method for diagnosing a UHV DC transmission lockout fault based on machine learning, comprising the following steps:

[0011] Step 1, obtaining various electrical quantity change data of each network node before and after the UHV DC blocking fault, and screening out characteristic electrical quantities and characteristic nodes to form fault characteristic data;

[0012] Step 2: extract the time point and frequency component of the fault from the fault feature data through wavelet transform, build a fault signal model, and use the extreme point detection method to determine the fault location;

[0013] Step 3: construct an input sequence based on the extracted time point and frequency component of the fault, and build a convolutional neural network model;

[0014] Step 4: Based on the convolutional neural network model, the input sequence is expanded into a time feature matrix, the channel attention mechanism is used to reallocate the weight of each feature in the time feature matrix, and the time attention mechanism is used to perform weighted averaging on the input time series of each time step in the time feature matrix to obtain an improved convolutional neural network model; the improved convolutional neural network model is trained using fault feature data, and the trained improved convolutional neural network model is used to extract the fault features of the UHV DC lockout to complete fault diagnosis.

[0015] Preferably, in step 1, screening characteristic electrical quantities includes:

[0016] Perform multiple blocking fault simulations on the UHV DC transmission system, and obtain multiple electrical quantity change data before and after each blocking fault simulation, including active power change data, reactive power change data, AC node voltage amplitude change data, and DC node voltage amplitude change data, as a set of simulation data of the fault simulation data set;

[0017] When the proportion of the number of groups of a certain type of electrical quantity change data that is less than the set electrical threshold ε exceeds the set proportion, this type of electrical quantity change data is removed from the fault simulation data set, and the remaining electrical quantity change data is used as the characteristic electrical quantity change data, and the corresponding electrical quantity is used as the characteristic electrical quantity.

[0018] Preferably, in step 1, screening characteristic nodes includes:

[0019] The importance ratio of the characteristic electrical quantity change data is calculated by the trimmed mean method, which is expressed by the following formula:

[0020]

[0021] In the formula: ΔD is the change data of any type of characteristic electrical quantity, and ΔD|i m is the change data of the characteristic electrical quantity in the i m -th simulation; m * is the number of groups of simulation data, and ΔD max is the maximum value of the change data of this characteristic electrical quantity in m * groups of simulation data, and ΔD min is the minimum value of the change data of this characteristic electrical quantity in m * groups of simulation data, and ΔD a0 is the trimmed mean of the electrical quantity change data, D before is the characteristic electrical quantity data before the fault, and k′ is the number of branch lines; is the characteristic importance ratio;

[0022] Sort the characteristic importance ratios from largest to smallest to obtain the degree of influence of the node on the fault. Set the importance threshold ε′ according to the degree of influence of the node on the fault. If the characteristic importance ratios of the change data of various electrical quantities of any node are not less than the corresponding importance threshold ε′, then this node is regarded as a characteristic node, and the electrical quantity data and the electrical quantity change data before and after the fault of this node are retained.

[0023] Preferably, step 2 includes:

[0024] Select a wavelet basis function Perform a Fourier transform on the wavelet basis function where the Fourier transform satisfies the following conditions:

[0025]

[0026] In the formula: represents the modulus of the Fourier transform, and |w| is the modulus of the frequency threshold;

[0027] Translate and scale the wavelet basis function to obtain a continuous wavelet basis function which is expressed by the following formula:

[0028]

[0029] In the formula: x is the scaling factor of the continuous wavelet basis function, y is the translation factor of the continuous wavelet basis function, and s is the time point when the fault occurs;

[0030] Using continuous wavelet basis functions For any I 2 Transform the signal function w(s) in the (Q) space, where I 2 (Q) space is a function space composed of all square-integrable functions defined on the given set Q, and the transformation result satisfies the following conditions:

[0031]

[0032] In the formula: M w (x, y) is the frequency component output by the wavelet transform, and w is the frequency threshold.

[0033] Preferably, in step 2, the fault signal model is represented by the following formula:

[0034]

[0035] In the formula: w(s) is the signal function, X is the sampling signal, m′ is the angular frequency, β is the signal phase shift, X j sin(jm′s + β j ) is the harmonic component, s is the time point when the fault occurs, Xe -k is the DC attenuation component, k is the amplitude of the harmonic component, T is the sampling length, X j is the j-th component of the sampling signal, β j is the j-th component of the signal phase shift.

[0036] Preferably, step 3 includes:

[0037] Construct an input sequence, which is represented by the following formula:

[0038] A = [a1, a2,..., a s ,..., a t s

[0039] In the formula: A ∈ P t*c is the input parameter of the model, P represents the two-dimensional space, c is the number of eigenvalues, a s is the actual output sequence of the model for the s-th category, a t is the actual output sequence of the model for the t-th category;

[0040] Construct a convolutional layer to extract features from the input sequence through convolutional operations, which is represented by the following formula:

[0041]

[0042] In the formula: is the convolutional kernel, n′ is the time window width, d′ is the number of convolutional kernels, y is the bias, ​The i-th feature generated for the convolutional kernel, f p is the activation function;

[0043] Construct a pooling layer to compress and abstract the extracted features through sub-region sampling, which is expressed by the following formula:

[0044]

[0045] In the formula: pool represents the pooling operation, is the i-th feature output by the pooling layer;

[0046] Construct a fully connected layer to connect the extracted features with the decision output, and abstract and combine the global time features, which is expressed by the following formula:

[0047] xfd = fp(xp - lastWfd + y)

[0048] In the formula: x fd is the output of the fully connected layer, W fd is the weight matrix of the fully connected layer, x p-last represents the output feature sequence of the neurons in the previous layer;

[0049] Construct an output layer, and the classification result output is expressed by the following formula:

[0050]

[0051] In the formula: is the classification result, is the sigmoid activation function.

[0052] Preferably, the expression of the loss function adopted in the model parameter update process is as follows:

[0053] T l = ∑b t log(q t )

[0054]

[0055] In the formula: log is the natural logarithm, q t is the intermediate calculation amount, u is the number of classification tasks, a t is the actual output sequence of the model for the t-th category, b t is the label output sequence of the model for the t-th category.

[0056] Preferably, in step 4, the channel attention mechanism includes:

[0057] Expand the input sequence into a time feature matrix, which is expressed by the following formula:

[0058]

[0059] In the formula: n is the number of features, and T is the sampling length; represents the original feature sequence containing n features at time τ;

[0060] represents the original time series of the m-th feature at T time steps;

[0061] The normalized channel attention mechanism weights are calculated by the following formula:

[0062]

[0063] In the formula: is the channel attention mechanism weight, is the normalized channel attention mechanism weight, W e and U e are the first weight matrix, the second weight matrix, and the third weight matrix of the channel attention mechanism respectively, and b e is the bias term of the channel attention mechanism, h τ-1 represents the hidden layer state at time τ - 1, and s′ τ-1 represents the state variable at time τ - 1.

[0064] Preferably, in step 4, the time attention mechanism includes:

[0065] Using the normalized channel attention weight as the weighting coefficient, calculate the adaptive weighted input to replace the original feature sequence x τ as the input of the time attention mechanism, as shown in the following formula:

[0066]

[0067] According to the adaptive weighted input update the hidden layer state h at time τ τ , which is expressed by the following formula:

[0068]

[0069] In the formula: f1 is the gated recurrent unit;

[0070] The time attention mechanism weights are calculated by the following formula:

[0071]

[0072] In the formula: is the source hidden state of the encoder of the time attention mechanism layer, T represents the sampling length, score is the performance evaluation function, c τ represents the adaptive time main sequence output;

[0073] Fuse the adaptive time main sequence output c τ with the hidden layer state h τ as the input of the decoder of the time attention mechanism layer, and calculate the fault diagnosis result, which is represented by the following formula:

[0074]

[0075] In the formula: W c and b c are the weight and bias of the fused input respectively; tanh is the hyperbolic tangent function; is the fault diagnosis result output by the decoder of the time attention mechanism layer.

[0076] The second aspect of the present invention provides a UHV DC transmission blocking fault diagnosis system based on machine learning, which runs the above-mentioned UHV DC transmission blocking fault diagnosis method based on machine learning, including:

[0077] A data extraction module for acquiring and screening fault feature data;

[0078] An information extraction module for extracting the time point and frequency components of the fault occurrence from the fault feature data, constructing a fault signal model and an input sequence, and discriminating the fault location;

[0079] A network improvement module for improving the convolutional neural network model using the channel attention mechanism and the time attention mechanism;

[0080] A diagnosis output module for performing fault diagnosis using the trained improved convolutional neural network model.

[0081] Compared with the prior art, the beneficial effects of the present invention at least include:

[0082] 1. Higher fault recognition accuracy: Traditional fault diagnosis methods often rely on manual feature extraction, which may not be able to fully cover the complexity and diversity of faults. The CNN based on the attention mechanism can automatically learn the deep features of DC blocking faults and highlight key information through the attention mechanism, thus significantly improving the accuracy of fault recognition.

[0083] 2. Stronger anti-interference ability: In the diagnosis of UHV DC blocking faults, the attention mechanism can help the network focus on fault features, reduce the interference of noise and irrelevant factors, and enable the model to maintain high diagnostic performance in complex environments.

[0084] 3. Reduced dependence on labeled data: Traditional fault diagnosis methods often require a large amount of accurately labeled data for training. The introduction of the attention mechanism enables the CNN to utilize data more effectively, achieving good training results even when the labeled data is limited. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 The flowchart of the UHVDC transmission blocking fault diagnosis method based on machine learning provided by the embodiment of the present invention;

[0086] Figure 2 The schematic diagram of the basic structure of the CNN in the embodiment of the present invention;

[0087] Figure 3 The schematic diagram of the multi-level attention mechanism model in the embodiment of the present invention;

[0088] Figure 4 The schematic diagram of the UHVDC transmission model in the application example of the present invention;

[0089] Figure 5 The schematic diagram for comparing the diagnosis accuracy of the UHVDC monopole blocking fault in the application example of the present invention;

[0090] Figure 6 The schematic diagram for comparing the diagnosis accuracy of the UHVDC bipolar blocking fault in the application example of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0091] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. The embodiments described in this application are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the spirit of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0092] The present invention provides a UHVDC transmission blocking fault diagnosis method and system based on machine learning, analyzes the electrical quantity characteristics under blocking faults, uses traveling wave theory and wavelet transform to decompose fault information to achieve fault location and form a feature matrix; takes the decomposed feature matrix as an input sequence, combines a multi-level attention mechanism to build and train a convolutional neural network model, further extracts the fault characteristics of UHVDC blocking, and realizes fast fault diagnosis.

[0093] As Figure 1 shown, Embodiment 1 of the present invention provides a UHVDC transmission blocking fault diagnosis method based on machine learning, including the following steps:

[0094] Step 1: Obtain the change data of various electrical quantities before and after the UHVDC blocking fault for each network node, and screen out the characteristic electrical quantities and characteristic nodes from them to form the fault characteristic data.

[0095] The UHVDC blocking fault will cause drastic changes in electrical quantities, which will in turn affect the stable operation of the power grid. After the blocking fault occurs, parameters such as current, voltage, active power, and reactive power will all fluctuate. The changes in electrical quantities are numerous and complex, posing great challenges to fault analysis. To accurately determine the cause of the fault, it is necessary to screen these electrical quantity change data and extract key information.

[0096] It should be noted that the following takes the power change data and voltage change data as examples for illustration, but the obtained change data is not limited to the change data of electrical quantities such as current, voltage, power, power factor, phase angle, and frequency.

[0097] In the preferred but non-limiting embodiment of the present invention, Step 1 specifically includes:

[0098] Step 1.1: Conduct m * times of blocking fault simulations on the UHVDC transmission system, where the number of network nodes is n * , and according to the real-time power flow, obtain m * groups of data sections before and after the fault, and calculate the active power change data ΔP ij and reactive power change data ΔQ ij of the ijth line before and after the DC blocking fault, which are expressed by the following formula:

[0099]

[0100] In the formula: and respectively represent the active and reactive power transmission values of the ijth branch before and after the fault.

[0101] Step 1.2: Calculate the voltage amplitude change data ΔU i of the ith AC node and the voltage amplitude change data ΔU k of the kth DC node before and after the DC blocking fault, which are expressed by the following formula:

[0102]

[0103] In the formula: and respectively represent the voltage amplitudes of the AC node and the DC node before and after the fault.

[0104] All the electrical quantity change data of each node obtained from the above m * times of blocking fault simulations constitute the fault simulation data set.

[0105] Step 1.3, among the simulation results of m * groups, determine whether the frequency of the power change data of the AC-DC power grid in Step 1.1 or the voltage amplitude change data in Step 1.2 being less than the set electrical threshold ε = (ε1, ε2, ε3, ε4) exceeds 2m * / 3 times. If so, it is determined that the corresponding electrical quantity change data ΔP ij , ΔQ ij , ΔU i or ΔU k cannot directly reflect the fault characteristics, and this type of electrical quantity change data is excluded from the fault simulation data set. The remaining electrical quantity change data is still divided into m * groups as the characteristic electrical quantity change data, and the corresponding electrical quantities are used as the characteristic electrical quantities. In this embodiment, it is assumed that ΔP ij , ΔQ ij , ΔU i and ΔU k are all characteristic electrical quantity change data.

[0106] It should be noted that different frequencies can also be set according to actual needs to determine whether the electrical quantity change data can directly reflect the fault characteristics. Based on the spirit of the present invention, they all fall within the protection scope of the present invention.

[0107] Step 1.4, calculate the importance degree of the characteristic electrical quantity change data.

[0108] Further preferably, to weaken the influence of extreme data, the trimmed mean is used to measure the importance of each electrical quantity. Taking ΔP ij as an example, the calculation methods of the other electrical quantity change data are the same, and its calculation expression is:

[0109]

[0110] In the formula: ΔP ij |i m is the active power change data of the i m th simulation, and k is the number of branch lines.

[0111] Step 1.5, considering the characteristics of the AC-DC interconnected power grid, for the nodes and branches closer to the DC fault point, their electrical information can better reflect the fault characteristics. Sort and from large to small respectively. The larger the ratio, the greater the influence of the fault. If the and of any node are not less than the importance threshold ε′ = ε′1, ε′2, ε′3, ε′4), where ε′1, ε′2, ε′3, ε′4 are respectively and If it corresponds to an importance threshold, the original data is retained. and and the corresponding electrical quantity change data ΔP ij , △Q ij , △U i , ΔU k .

[0112] Step 2: Extract the time point and frequency components of the fault occurrence from the fault feature data through wavelet transform, construct a fault signal model, and use the extreme point detection method to identify the fault location.

[0113] Since the actually collected fault recording signals contain rich fault information, wavelet transform can accurately capture the time point of the fault occurrence and its frequency components through time-frequency localization characteristics, providing fine time-frequency information for fault diagnosis. In addition, its multi-scale analysis ability enables it to identify fault features at different frequency levels, thus better understanding the hierarchical structure of the fault. In a complex noise environment, the anti-interference ability of wavelet transform enables it to effectively extract useful signals and ensure the accuracy of fault location.

[0114] In a preferred but non-limiting embodiment of the present invention, Step 2 specifically includes:

[0115] Step 2.1: Select the wavelet basis function as For the wavelet basis function Perform Fourier transform where the Fourier transform Satisfies the following conditions:

[0116]

[0117] In the formula: Represents the modulus of the Fourier transform, and |w| is the modulus of the frequency threshold;

[0118] Step 2.2: Translate and scale the wavelet basis function To obtain the continuous wavelet basis function Which is expressed by the following formula:

[0119]

[0120] In the formula: Is a wavelet basis function that depends on the scaling factor x and the translation factor y. Since x and y are continuous, Is called a continuous wavelet basis; s is the time point of the fault occurrence.

[0121] Step 2.3: Use the continuous wavelet basis function For any I 2(Q) Transform the signal function w(s) in space, where I 2 (Q) The space is a function space composed of all square-integrable functions defined on a given set Q, and the transformation result satisfies the following conditions:

[0122]

[0123] In the formula: w(s) represents the signal function, usually a time-varying signal; M w (x, y) is the frequency component output by the wavelet transform, and w is the frequency threshold.

[0124] Step 2.4, construct a fault signal model, detect the extreme points according to the wavelet transform result, and preliminarily identify the fault location.

[0125] When a fault occurs in the power system, the components of its fault signal are very complex and contain rich information. Due to the very rapid occurrence and development process of power system faults, once a fault occurs, parameters such as current and voltage in the system will change violently instantaneously. The fault signal contains multiple frequency components, covering from low frequency to high frequency, making the fault signal present a complex waveform. In addition, the fault signal contains rich harmonic information, which may be caused by nonlinear elements in the power system, further increasing the difficulty of signal analysis.

[0126] To accurately obtain the fault location information in complex signals, construct a fault signal model, which is expressed by the following formula:

[0127]

[0128] In the formula: X j sin(jm′s + β j ) is the harmonic component; Xe -k is the DC decay component; X is the sampling signal, m′ is the angular frequency, β represents the phase shift of the signal; k represents the amplitude of the harmonic component, T is the sampling length, X j is the j-th component of the sampling signal, and βj is the j-th component of the signal phase shift.

[0129] Step 2.5, since the occurrence of a fault will cause the derivative at the corresponding point to not exist, and the wavelet transform has an extreme value here. Therefore, the preliminary identification of the fault location is achieved by detecting the extreme points, and at the same time, the electrical quantity data after the fault is processed by the wavelet transform to be used as the input sequence for subsequent feature extraction and fault diagnosis.

[0130] Step 3, construct an input sequence according to the extracted time point and frequency component of the fault occurrence, and construct a convolutional neural network model.

[0131] CNN (Convolutional Neural Networks, convolutional neural network) is a widely used intelligent algorithm in deep learning, with strong feature extraction capabilities. Its composition structure is as Figure 2 shown. The basic structure of CNN includes an input layer, a convolutional layer, an activation layer, a pooling layer, and a fully connected layer. The input layer receives the original data sequence; the convolutional layer extracts features through multiple convolutional kernels; subsequently, the activation layer introduces a non-linear function, enabling the network to learn and simulate more complex features; the pooling layer is responsible for reducing the feature dimension, while retaining important information and reducing the computational amount; the fully connected layer completes the output of the model through classification or regression tasks.

[0132] In a preferred but non-limiting embodiment of the present invention, step 3 specifically includes:

[0133] Step 3.1, according to the wavelet transform result of step 2, construct the input sequence of the model to form the input layer, as shown in the following formula:

[0134] A = [a1, a2,..., a s ,..., a t s

[0135] In the formula: A ∈ P t*c is the input parameter of the model, P represents the two-dimensional space, c is the number of eigenvalue, a s is the actual output sequence of the model for the s-th category, a t is the actual output sequence of the model for the t-th category.

[0136] Step 3.2, construct the convolutional layer, and extract features from the input sequence through convolutional operations. Its expression is:

[0137]

[0138] In the formula: is the convolutional kernel, representing the weight matrix; n' is the time window width, d' is the number of convolutional kernels, y is the bias, is the i-th feature generated by the convolutional kernel; f p is the activation function.

[0139] Step 3.3, construct a pooling layer after the convolutional layer in step 3.2, which is used to compress and abstract features through sub-region sampling, while retaining important information of the input feature map. Its expression is:

[0140]

[0141] In the formula: pool represents the pooling operation, usually divided into max pooling and global max pooling, is the i-th feature output by the pooling layer.​

[0142] Step 3.4: Construct a fully connected layer. The structure of the fully connected layer is the same as that of a traditional neural network. Connect the features extracted in Steps 3.2 - 3.3 with the final decision output, which can realize the conversion from raw data to high-level abstraction and then to specific decisions.

[0143] Further preferably, in the fully connected layer, the global time features are further abstracted and combined in the full connection. The final output is expressed by the following formula:

[0144] x fd = f p [(x p-last W fd + y)

[0145] where: x fd is the output of the fully connected layer, W fd is the weight matrix of the fully connected layer, and x p-last represents the output feature sequence of the neurons in the previous layer.

[0146] Step 3.5: Construct an output layer. For a binary classification problem, the number of output units is 1. The classification result through the full connection is:

[0147]

[0148] where: is the classification result, is the sigmoid activation function.

[0149] Step 3.6: Update the model parameters using the following loss function:

[0150]

[0151] where:

[0152]

[0153] where: log is the natural logarithm, q t is the intermediate calculation quantity, u is the number of classification tasks; a t is the actual output sequence of the model for the t-th category, and b t is the label output sequence of the model for the t-th category.

[0154] Step 4: Based on the convolutional neural network model, expand the input sequence into a time feature matrix, use the channel attention mechanism to reassign the weights of each feature in the time feature matrix, and use the time attention mechanism to perform weighted averaging on the input time series of each time step in the time feature matrix to obtain an improved convolutional neural network model; use the fault feature data to train the improved convolutional neural network model, and use the trained improved convolutional neural network model to extract the fault features of UHVDC blocking to complete the fault diagnosis.

[0155] Figure 3 It is a schematic diagram of an improved CNN model with an attention mechanism added. The model includes four layers: a CNN layer, a fully connected layer, a channel attention mechanism layer, and a time attention mechanism layer. During the data input process, the attention mechanism is used to weight each input feature, thereby strengthening the features that have a significant impact on the result. At the output end, the time attention mechanism is adopted to improve the accuracy of the model by utilizing the correlation of data on the time scale.

[0156] In a further preferred embodiment of the present invention, step 4 specifically includes:

[0157] Step 4.1: Construct the input sequence in two ways, namely time and feature, which is expressed by the following formula:

[0158] x = [x1, x2,..., x T = [x (1) , x (2) ,..., x (n) T

[0159] In the formula: n is the number of features, and T is the sampling length.

[0160] Step 4.2: Expand the input sequence in step 4.1 in matrix form, which is expressed by the following formula:

[0161]

[0162] In the formula: represents the original feature sequence containing n features at time τ; represents the original time series of the mth feature at T time steps.

[0163] Step 4.3: In order to obtain the association between each feature variable and the current time series, that is, for the current time series, the importance of its corresponding feature quantity, use feature attention for calculation. By associating the feature variable at time τ with the hidden layer state h τ-1 at time τ - 1 and the state variable s' τ-1 , calculate the channel attention mechanism weight corresponding to the feature quantity at the current moment ​As shown in the following formula:

[0164]

[0165] In the formula: W e 、U e are respectively the first weight matrix, the second weight matrix, and the third weight matrix of the channel attention mechanism; b e is the bias term of the channel attention mechanism.

[0166] Step 4.4: Normalize the weights of the channel attention mechanism obtained in Step 4.3 so that the sum of the weight values corresponding to time τ is 1, as shown in the following formula:

[0167]

[0168] Step 4.5: Use the normalized weights of the channel attention mechanism as the weighting coefficients, multiply them by the input feature values, and obtain the adaptive weighted input to replace the original feature sequence x τ and feed it into the subsequent model, and update the state h t of the hidden layer at the next moment.

[0169] More preferably, the adaptive weighted input is represented by the following formula:

[0170]

[0171] More preferably, the update of the hidden layer state is represented by the following formula:

[0172]

[0173] In the formula: f1 is a GRU network unit (Gated Recurrent Unit).

[0174] It can be understood that by adjusting the weighted input and the hidden layer state in real time through the attention mechanism, the correlation between the feature values and the corresponding time series can be dynamically extracted.

[0175] Step 4.6: After calculating the adaptive channel attention result weighted input and the hidden layer state in Step 4.5, use the output result obtained by the channel attention mechanism as the input of the time attention mechanism in the next stage. The time attention mechanism focuses on the input sequence and obtains the adaptive time main sequence output by weighted averaging the input time series. Figure 3 Illustrates the principle of the time attention mechanism implementation. The weights of the time attention mechanism are calculated as follows:

[0176]

[0177] In the formula: is the source hidden state of the time attention mechanism layer encoder, T represents the sampling length, score is the performance evaluation function, and c τ represents the adaptive time main sequence output.

[0178] The core idea of the above formula is to reorganize the vector c τ multiple times to make it maintain dynamic characteristics. The vector c τ in the traditional CNN model selects the output of the last time step as the final output. Since the data can be regarded as a process of gradually extracting features after being processed by GRU, the output of the last time step usually contains important information about the states of past time steps.

[0179] However, for data with long time steps and strong information correlation, the traditional CNN model will cause important information of past time steps to be ignored or the information closely related to the prediction result to be not prominent. Through the time attention mechanism of the present invention, the vector c τ is dynamized, and c τ is no longer just the output of the last time step, but a dynamic combination of individual time steps. Different weight coefficients are given for different prediction contents, so as to give different c τ and finally realize the function of extracting necessary information in the time series.

[0180] The context vector c τ updated in real time is fused with the hidden layer state h τ as the input of the time attention mechanism layer decoder:

[0181]

[0182] In the formula: W c and b c are the weights and biases for fusing the input respectively; tanh is the hyperbolic tangent function; is the fault diagnosis result output by the time attention mechanism layer encoder.

[0183] It can be understood that the present invention improves the traditional CNN model through the attention mechanism, and constitutes an overall deep learning model. The input of the model is the power data preliminarily screened in step 1, from which the characteristic information of the fault is extracted, and the DC lockout fault diagnosis result is output. Specifically, by introducing the channel attention mechanism, each channel represents a feature learned by the network, which enables the improved model of the present invention to automatically learn the importance of different channels (or features), and adjust the characteristic response of each channel accordingly, strengthen important features, and suppress unimportant features, thereby improving the performance of the model; by introducing the time attention mechanism of sequence data, the model can pay more attention to certain key time points in the UHV DC transmission lockout fault data sequence, rather than the entire sequence, so as to handle sequence dependencies and capture sequence dynamic changes.

[0184] Embodiment 2 of the present invention provides a UHV DC transmission lockout fault diagnosis system based on a machine learning algorithm, and runs the UHV DC transmission lockout fault diagnosis method based on a machine learning algorithm described in Embodiment 1, including:

[0185] Data extraction module, used to obtain and filter fault characteristic data;

[0186] The information extraction module is used to extract the time point and frequency component of the fault from the fault feature data, build the fault signal model and input sequence, and identify the fault location;

[0187] Network improvement module, which is used to improve the convolutional neural network model using channel attention mechanism and temporal attention mechanism;

[0188] The diagnosis output module is used to perform fault diagnosis using the trained improved convolutional neural network model.

[0189] In order to verify the effect of the present invention, the advancement of the proposed UHV DC transmission lockout fault diagnosis method based on machine learning algorithm is illustrated through the following application examples.

[0190] The simulation model of UHV DC transmission is established on the hardware platform AMD(R) Ryzen 7 5800H CPU and NVIDA(R) RTX3070 GPU. Figure 4 The CNN structure used is "6C-1S-12C-2S", where C represents the convolution layer and S represents the downsampling layer. It can be seen that the CNN structure used includes 6 convolution layers and 12 pooling layers, and the specific parameter settings are shown in Table 1. The training cycle refers to the number of times the model updates its parameters on the training data; the batch size refers to the size of the sample batch used each time the model is updated.

[0191] Table 1 Network indicators of CNN

[0192]

[0193] To verify the superiority of the designed fault diagnosis strategy, it is compared with the multi-layer perceptron (MLP) and convolutional neural network (CNN) strategies. Taking the UHV monopolar / bipolar DC blocking fault as the simulation scenario, the fault diagnosis accuracy of different deep learning algorithms is analyzed, as Figure 5 and Figure 6 shown.

[0194] From Figure 5 it can be seen that the accuracy of the three deep learning algorithms all increases with the increase of the training cycle. Among them, the accuracy of the improved CNN strategy proposed in the present invention for DC blocking fault diagnosis is the highest. When the training cycle reaches 7 times, the accuracy will stabilize at 100%. Further increasing the training time will not significantly improve the accuracy. At this time, the model has obtained sufficient information for accurate fault diagnosis. The accuracy of DC blocking fault diagnosis based on MLP is the lowest, and finally stabilizes at 88.9%. The accuracy of blocking fault diagnosis based on the traditional CNN is between MLP and the improved CNN, and stabilizes at 92.4%.

[0195] From Figure 6 it can be seen that, similarly, the accuracy of the three deep learning algorithms all increases with the increase of the training cycle. Among them, the accuracy of the improved CNN strategy proposed in the present invention for DC blocking fault diagnosis is the highest. When the training cycle reaches 7 times, the accuracy will stabilize at 98%. In addition, different from monopolar blocking, the diagnosis rate of bipolar blocking decreases by an average of 2%, mainly due to the complexity of its fault manifestation and the diversity of system responses. When a bipolar blocking fault occurs, since the faults of the two poles may affect each other, the fault characteristics may be masked or confused, making it more difficult to capture the clues for fault diagnosis. Therefore, the improved CNN strategy based on the attention mechanism proposed in the present invention shows the highest accuracy rate and realizes the accurate diagnosis of UHV DC blocking faults.

[0196] Compared with the prior art, the beneficial effects of the present invention at least include:

[0197] 1. Higher fault recognition accuracy: Traditional fault diagnosis methods often rely on manual feature extraction and may not be able to fully cover the complexity and diversity of faults. The CNN based on the attention mechanism can automatically learn the deep features of DC blocking faults and highlight key information through the attention mechanism, thus significantly improving the accuracy of fault recognition.

[0198] 2. Stronger anti-interference ability: In the diagnosis of UHV DC blocking faults, the attention mechanism can help the network focus on fault features, reduce the interference of noise and irrelevant factors, and enable the model to maintain a high diagnostic performance in a complex environment.

[0199] 3. Reduced dependence on labeled data: Traditional fault diagnosis methods often require a large amount of accurately labeled data for training. The introduction of the attention mechanism enables the CNN to utilize data more effectively and achieve good training results even when the labeled data is limited.

[0200] The present disclosure may be a system, a method, and / or a computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions thereon for causing a processor to implement various aspects of the present disclosure.

[0201] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent replacements can still be made to the specific embodiments of the present invention, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the protection scope of the claims of the present invention.

Claims

1. A method for diagnosing blocking faults in UHVDC transmission based on machine learning, characterized in that The following steps are involved: Step 1, obtaining various electrical quantity change data of each network node before and after the UHV DC blocking fault, and screening out characteristic electrical quantities and characteristic nodes to form fault characteristic data; Step 2: extract the time point and frequency component of the fault from the fault feature data through wavelet transform, build a fault signal model, and use the extreme point detection method to determine the fault location; Step 3: construct an input sequence based on the extracted time point and frequency component of the fault, and build a convolutional neural network model; Step 4: Based on the convolutional neural network model, the input sequence is expanded into a time feature matrix, the channel attention mechanism is used to redistribute the weight of each feature in the time feature matrix, and the time attention mechanism is used to perform weighted averaging on the input time series of each time step in the time feature matrix to obtain an improved convolutional neural network model. The fault feature data is used to train an improved convolutional neural network model, and the trained improved convolutional neural network model is used to extract the fault features of UHV DC interlocking to complete fault diagnosis.

2. The method for diagnosing UHV DC transmission lockout faults based on machine learning according to claim 1, characterized in that: In step 1, the characteristic electrical quantities to be screened include: Perform multiple blocking fault simulations on the UHV DC transmission system, and obtain multiple electrical quantity change data before and after each blocking fault simulation, including active power change data, reactive power change data, AC node voltage amplitude change data, and DC node voltage amplitude change data, as a set of simulation data of the fault simulation data set; When the proportion of the number of groups of a certain type of electrical quantity change data that is less than the set electrical threshold ε exceeds the set proportion, this type of electrical quantity change data is removed from the fault simulation data set, and the remaining electrical quantity change data is used as the characteristic electrical quantity change data, and the corresponding electrical quantity is used as the characteristic electrical quantity.

3. The method for diagnosing UHV DC transmission lockout faults based on machine learning according to claim 2 is characterized in that: In step 1, the screening feature nodes include: The importance ratio of characteristic electrical quantity change data is calculated using the truncated mean method, which is expressed as follows: Where: ΔD is the change data of any type of characteristic electrical quantity, and ΔD|i m is the change data of the characteristic electrical quantity in the i m -th simulation; m * is the number of groups of simulation data, and ΔD max is the maximum value of the change data of this characteristic electrical quantity in m * groups of simulation data, and ΔD min is the minimum value of the change data of this characteristic electrical quantity in m * groups of simulation data, and ΔD a0 is the trimmed mean of the change data of the electrical quantity, D before is the characteristic electrical quantity data before the fault, and k′ is the number of branch lines; is the ratio of characteristic importance; The characteristic importance ratios are sorted from large to small to obtain the degree of influence of the node by the fault. The importance threshold ε′ is set according to the degree of influence of the node by the fault. If the characteristic importance ratios of various types of electrical quantity change data of any node are not less than the corresponding importance threshold ε′, the node is taken as a characteristic node, and the electrical quantity data and electrical quantity change data before and after the fault of the node are retained.

4. The method for diagnosing UHV DC transmission lockout faults based on machine learning according to claim 1, characterized in that: Step 2 includes: Select wavelet basis function For the wavelet basis function Perform Fourier transform Among them, the Fourier transform Satisfies the following conditions: In the formula: represents the modulus of the Fourier transform, and |w| is the modulus of the frequency threshold; Translation and Scaling Wavelet Basis Functions Obtain the continuous wavelet basis function Expressed by the following formula: Where: x is the scaling factor of the continuous wavelet basis function, y is the translation factor of the continuous wavelet basis function, and s is the time point when the fault occurs; Using continuous wavelet basis functions For any I 2 transform the signal function w(s) in the (Q) space, where I 2 (Q) space is a function space composed of all square-integrable functions defined on a given set Q, and the transformation result satisfies the following conditions: Where: M w (x, y) is the frequency component output by the wavelet transform, and w is the frequency threshold.

5. The method for diagnosing UHV DC transmission lockout faults based on machine learning according to claim 1, characterized in that: In step 2, the fault signal model is expressed as follows: Where: w(s) is the signal function, X is the sampling signal, m′ is the angular frequency, β is the signal phase shift, X j sin(jm′s + β j ) is the harmonic component, s is the time point when the fault occurs, Xe -k is the DC decay component, k is the amplitude of the harmonic component, T is the sampling length, X j is the j-th component of the sampling signal, β j is the j-th component of the signal phase shift.

6. The method for diagnosing blocking faults in UHV DC transmission based on machine learning according to claim 1 is characterized in that: Step 3 includes: Construct an input sequence, which is expressed by the following formula: A = [a1, a2, …, a s , …, a t s ​ Where: A ∈ P t*c is the input parameter of the model, P represents a two-dimensional space, c is the number of eigenvalues, a s is the actual output sequence of the model for the s-th category, a t is the actual output sequence of the model for the t-th category; Construct a convolutional layer, and perform feature extraction on the input sequence through convolutional operations, which is expressed by the following formula: Wherein: is the convolution kernel, n′ is the time window width, d′ is the number of convolution kernels, y is the bias, is the i-th feature generated by the convolution kernel, f p is the activation function; Construct a pooling layer, and compress and abstract the extracted features through sub-region sampling, which is expressed by the following formula: where: pool represents the pooling operation, is the i-th feature output by the pooling layer; Construct a fully connected layer, connect the extracted features with the decision output, and abstract and combine the global time features, which is expressed by the following formula: x fd = f p (x p-last W fd + y) where: x fd is the output of the fully connected layer, and W fd is the weight matrix of the fully connected layer, and x p-last represents the output feature sequence of the neurons in the previous layer; Construct an output layer, and the classification result of the output is expressed by the following formula: Wherein: is the classification result, and f θ is the sigmoid activation function.

7. The method for diagnosing blocking faults in UHV DC transmission based on machine learning according to claim 6 is characterized in that: The expression of the loss function adopted in the model parameter update process is as follows: τ l = ∑b t log(q t ) where: log is the natural logarithm, q t is an intermediate calculation quantity, u is the number of classification tasks, a t is the actual output sequence of the model for the t-th category, b t is the label output sequence of the model for the t-th category.

8. The method for diagnosing blocking faults in UHV DC transmission based on machine learning according to claim 1 is characterized in that: In step 4, the channel attention mechanism includes: Expand the input sequence into a time feature matrix, which is expressed by the following formula: where: n is the number of features, and T is the sampling length; represents the original feature sequence containing n features at time τ; Denote the original time series of the m-th feature at T time steps; Calculate the normalized channel attention mechanism weights through the following formula: Wherein: is the weight of the channel attention mechanism, is the normalized weight of the channel attention mechanism, W e and U e are the first weight matrix, the second weight matrix, and the third weight matrix of the channel attention mechanism respectively, b e is the bias term of the channel attention mechanism, h τ-1 represents the hidden layer state at time τ - 1, s′ τ-1 represents the state variable at time τ - 1.

9. The method for diagnosing blocking faults in UHV DC transmission based on machine learning according to claim 8 is characterized in that: In step 4, the time attention mechanism includes: Using the normalized channel attention weights as the weighting coefficients to calculate the adaptive weighted input to replace the original feature sequence x τ as the input of the temporal attention mechanism, as shown in the following formula: According to the adaptive weighted input Update the hidden layer state h at time τ τ , which is expressed by the following formula: Where: f1 is a gated recurrent unit; Calculate the time attention mechanism weights through the following formula: Wherein: is the source hidden state of the time attention mechanism layer encoder, T represents the sampling length, score is the performance evaluation function, and c τ represents the adaptive time main sequence output; Output the adaptive time main sequence c τ Fuse with the hidden layer state h τ The input of the time attention mechanism layer decoder is fused to calculate the fault diagnosis result, which is expressed by the following formula: Where: W c and b c are the weights and biases of the fusion input respectively; tanh is the hyperbolic tangent function; is the fault diagnosis result output by the decoder of the time attention mechanism layer.

10. A UHVDC transmission blocking fault diagnosis system based on machine learning, which operates according to the UHVDC transmission blocking fault diagnosis method based on machine learning described in any one of claims 1-9, characterized in that, It includes: A data extraction module, which is used to obtain and screen fault feature data; An information extraction module, which is used to extract the time points and frequency components when faults occur from the fault feature data, construct a fault signal model and an input sequence, and discriminate the fault location; A network improvement module, which is used to improve the convolutional neural network model using the channel attention mechanism and the time attention mechanism; A diagnosis output module, which is used to perform fault diagnosis using the trained improved convolutional neural network model.

Images