A method, system, electronic device and readable storage medium for aging assessment of a surge arrester

Abstract

This invention relates to a method, system, electronic device, and readable storage medium for surge arrester aging assessment, belonging to the field of surge arrester aging assessment technology. It solves the problems of low assessment accuracy and poor generalization ability in existing technologies. The method includes: collecting time-series characteristic data and corresponding aging index data of multiple surge arresters to construct an initial sample set; performing data augmentation on the initial sample set to obtain a training sample set; constructing a multi-branch time-series fusion network model for aging index prediction; training the constructed multi-branch time-series fusion network model based on the training sample set to obtain a trained multi-branch time-series fusion network model; inputting the time-series characteristic data of the surge arrester to be assessed into the trained multi-branch time-series fusion network model to predict the aging index, and obtaining the aging degree assessment result of the surge arrester to be assessed. This achieves fast and accurate surge arrester aging assessment.

Description

Technical Field

[0001] This invention relates to the field of surge arrester aging assessment technology, and in particular to a surge arrester aging assessment method, system, electronic device and readable storage medium. Background Technology

[0002] Metal oxide surge arresters (MOAs) are core devices in power systems that limit overvoltages and protect electrical equipment. During long-term operation, MOAs age and deteriorate due to exposure to power frequency voltage, inrush current, and environmental factors (such as humidity and temperature). This degradation is mainly manifested in the deterioration of the valve plate's nonlinear characteristics, increased leakage current, and increased power consumption. Accurately assessing the aging degree of MOAs is crucial for ensuring the safe operation of power systems and guiding condition-based maintenance.

[0003] Existing neural network-based aging assessment methods for surge arresters typically use features such as resistive current and power consumption as inputs to construct shallow neural networks (such as backpropagation networks) to predict the degree of aging. While these methods achieve nonlinear mapping to some extent, they suffer from the following drawbacks: the input features are mostly single-point measurements, failing to fully utilize the temporal evolution of the surge arrester's aging process; the network structure is simplistic (mostly fully connected or with a single hidden layer), making it difficult to simultaneously capture short-term fluctuations and long-term trends; and in practical engineering, training samples with aging labels are scarce, leading to model overfitting and poor generalization ability. Existing surge arrester aging assessment methods struggle to fully utilize the temporal evolution patterns and multi-scale time features in multi-period monitoring data, resulting in low assessment accuracy and poor generalization ability. Summary of the Invention

[0004] Based on the above analysis, the embodiments of the present invention aim to provide a surge arrester aging assessment method, system, electronic device, and readable storage medium to solve the problems of low assessment accuracy and poor generalization ability in existing methods.

[0005] On one hand, embodiments of the present invention provide a method for assessing the aging of surge arresters, comprising the following steps: An initial sample set was constructed by collecting time-series characteristic data and corresponding aging index data of multiple surge arresters. Data augmentation is performed on the initial sample set to obtain the training sample set; A multi-branch temporal fusion network model is constructed for aging index prediction; the constructed multi-branch temporal fusion network model is trained based on the training sample set to obtain a trained multi-branch temporal fusion network model. The time-series characteristic data of the surge arrester to be evaluated is input into a trained multi-branch time-series fusion network model to predict the aging index and obtain the aging degree evaluation result of the surge arrester to be evaluated.

[0006] Based on the above method, further improvements are made to the initial sample set, and the training sample set is obtained by data augmentation using the following method: Construct a conditional generative adversarial network, including a generator and a discriminator; The generator is used to generate time-series feature data based on the aging index and the aging sensitivity index; The discriminator is used to determine whether the input time-series feature data and the corresponding aging index data are real data; The conditional generative adversarial network is trained based on the initial sample set to obtain a trained conditional generative adversarial network. Samples are generated based on a pre-trained conditional generative adversarial network, and data augmentation is performed on the initial sample set.

[0007] Based on the above method, a further improvement is made to calculate the training loss of the conditional generative adversarial network using the following formula:

[0008] in, This indicates the generation of adversarial loss. Indicates the time-series consistency loss. This represents the balance coefficient.

[0009] Based on the further improvement of the above method, the timing consistency loss is calculated using the following formula:

[0010] in, Expressing expectations, This represents the k-th difference vector calculated from the real sample; This represents the k-th difference vector calculated from the generated samples. Let T represent the square of the L2 norm of the vector, and let T represent the sequence length of the real sample.

[0011] Based on a further improvement of the above method, the multi-branch temporal fusion network model includes: Gated loop unit branches are used to extract long-term temporal dependency features of the input data; Temporal convolutional network branches are used to extract multi-scale local features from the input data; Causal self-attention branch is used to extract global interaction features of the input data; The gated fusion layer is used to adaptively weight and fuse the features extracted from each branch, and to predict the aging index based on the fused features.

[0012] Based on a further improvement of the above method, the gated loop unit branch adopts a bidirectional single-layer gated loop unit; The first-level gated loop unit is used to calculate the first hidden state at each time step based on the input data at that time step and the hidden state output by the first-level gated loop unit at the previous time step. The second-level gated loop unit is used to calculate the second hidden state at each time step based on the first hidden feature output by the first-level gated loop unit at that time step and the second hidden state output by the second-level gated loop unit at the previous time step. The second hidden state at each time step is the temporal long-term dependency feature extracted from the gated recurrent unit branch.

[0013] Based on a further improvement of the above method, the first-layer gated loop unit extracts the first hidden state using the following formula:

[0014]

[0015]

[0016]

[0017] in, This represents the input data at time t. This represents the hidden state of the output of the first-level gated loop unit at time t-1. , , , , and Both represent trainable weight matrices. , and Both represent bias vectors. This represents the Sigmoid function. This represents the tanh activation function. , and Indicates intermediate parameters. This represents the hidden state of the output of the first-level gated loop unit at time t. The table represents element-wise multiplication.

[0018] Based on a further improvement of the above method, the temporal convolutional network branch includes at least three parallel sub-branches and a fusion layer; Each sub-branch independently performs dilated causal convolution to capture local features at different scales; The fusion layer is used to concatenate local features of different scales along the feature dimension and obtain multi-scale local features by pointwise convolution.

[0019] Based on the further improvement of the above method, the training loss of the multi-branch temporal fusion network model is calculated using the following formula:

[0020] in, Indicates the predicted loss. This represents the loss due to physical monotonicity constraints. This represents the trend consistency regularization loss. This represents the trainable parameters of a multi-branch temporal fusion network model. The L2 norm of a vector is the square of its value. , and This represents the weighting coefficient.

[0021] Based on the further improvement of the above method, the physical monotonicity constraint loss is calculated using the following formula:

[0022] in, Represents the set of sample pairs The number of sample pairs in the middle This represents the aging index predicted by the model for the j-th sample at time T. Let represent the aging index predicted by the model for the i-th sample at time T. Represents the marginal parameter. This represents the maximum difference in the aging sensitivity index within the sample set. This represents the aging sensitivity index of the i-th sample. This represents the aging sensitivity index of the j-th sample.

[0023] Based on the further improvement of the above method, the trend consistency regularization loss is calculated using the following formula:

[0024] in, This represents the aging index predicted by the model for the i-th sample at time t. The aging index of the i-th sample at time t+1, as predicted by the model, is represented by T, which represents the sequence length. This indicates the number of samples in the current training batch.

[0025] Based on further improvements to the above method, the timing characteristic data of the surge arrester are obtained in the following manner: At each moment, the voltage signal and total leakage current signal across the surge arrester are collected synchronously; Based on the Discrete Fourier Transform, the voltage amplitude, voltage phase angle, current amplitude, and current phase angle corresponding to each harmonic are extracted from the voltage signal and the total leakage current signal, respectively. Based on the voltage amplitude, voltage phase angle, current amplitude, and current phase angle corresponding to each harmonic, the capacitive current components corresponding to each harmonic are obtained through orthogonal projection calculation and temperature-frequency compensation model. The capacitive current components corresponding to each harmonic are superimposed to obtain the total capacitive current, and the resistive current component of the surge arrester is decoupled from the total leakage current signal based on the total capacitive current. Based on the resistive current component at each moment, the characteristic data of that moment are obtained, and the timing characteristic data of the surge arrester are obtained.

[0026] Based on further improvements to the above method, the characteristic data at this moment include the fundamental amplitude of resistive current, the odd-even harmonic ratio of resistive current, approximate power consumption, and aging sensitivity index.

[0027] Based on a further improvement to the above method, characteristic data for each moment is obtained based on the resistive current component, including: Based on the discrete Fourier transform, the resistive current amplitude and resistive current phase angle corresponding to each harmonic are extracted from the decoupled resistive current components. The fundamental amplitude of the resistive current and the odd-even harmonic ratio of the resistive current are extracted based on the resistive current amplitude corresponding to each harmonic. The approximate power consumption of the surge arrester is calculated based on the resistive current amplitude, resistive current phase angle, voltage amplitude, and voltage phase angle corresponding to each harmonic.

[0028] Based on further improvements to the above method, the aging sensitivity index is obtained in the following way:

[0029] in, This represents the aging sensitivity index at time t. This represents the fundamental amplitude of the resistive current at time t. This represents the odd-even harmonic ratio of the resistive current at time t. This represents the approximate power consumption at time t. This indicates the fundamental resistive current of the surge arrester at the time of manufacture. This indicates the approximate power consumption of the surge arrester at the time of manufacture.

[0030] On the other hand, embodiments of the present invention provide a surge arrester aging assessment system, comprising: The sample set construction module is used to collect time-series characteristic data and corresponding aging index data of multiple surge arresters to construct an initial sample set. The data augmentation module is used to augment the initial sample set to obtain the training sample set; The model training module is used to construct a multi-branch temporal fusion network model for aging index prediction; the constructed multi-branch temporal fusion network model is trained based on the training sample set to obtain a trained multi-branch temporal fusion network model. The evaluation module is used to input the time-series characteristic data of the surge arrester to be evaluated into the trained multi-branch time-series fusion network model to predict the aging index and obtain the aging degree evaluation result of the surge arrester to be evaluated.

[0031] On the other hand, embodiments of the present invention also provide an electronic device, the electronic device including a processor and a memory; wherein, the memory stores a computer program, and when the computer program is executed by the processor, it implements the steps of the metal oxide surge arrester leakage current decoupling method as described above.

[0032] On the other hand, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the surge arrester aging assessment method as described above.

[0033] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. Augment the initial sample set to expand the training set and improve the model's generalization ability under small sample conditions.

[0034] 2. By using a multi-branch temporal fusion network model to simultaneously capture different types of change features during the aging process, the model's ability to fit complex aging trajectories is improved.

[0035] 3. By constructing a multi-period (e.g., weekly or monthly) time series feature sequence, the model can learn the dynamic evolution trend of surge arrester aging, effectively suppressing the interference of random noise in a single measurement, and making the evaluation results more stable and reliable.

[0036] 4. Output a continuous aging index in the range of 0 to 1. Compared with the traditional discrete level, it can reflect the gradual process of aging more precisely, which is conducive to the formulation of differentiated maintenance strategies (such as early warning, trend analysis, etc.).

[0037] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0038] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a flowchart of the surge arrester aging assessment method according to an embodiment of the present invention; Figure 2 A simplified equivalent model diagram of a metal oxide surge arrester; Figure 3 This is a schematic diagram showing the phasor relationship between voltage and current in a surge arrester. Figure 4 This is a schematic diagram comparing the decoupling effect of a preferred embodiment of the present invention with that of the prior art. Detailed Implementation

[0039] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0040] A specific embodiment of the present invention discloses a method for evaluating the aging of surge arresters, such as... Figure 1 As shown, it includes the following steps: S1. Collect time-series characteristic data and corresponding aging index data of multiple surge arresters to construct an initial sample set; S2. Perform data augmentation on the initial sample set to obtain the training sample set; S3. Construct a multi-branch temporal fusion network model for aging index prediction; train the constructed multi-branch temporal fusion network model based on the training sample set to obtain a trained multi-branch temporal fusion network model. S4. Input the time-series characteristic data of the surge arrester to be evaluated into the trained multi-branch time-series fusion network model to predict the aging index and obtain the aging degree evaluation result of the surge arrester to be evaluated.

[0041] Compared with existing technologies, the surge arrester aging assessment method provided in this embodiment constructs an initial sample set by collecting time-series feature data and corresponding aging index data of multiple surge arresters. The initial sample set is then augmented to expand the training set, improving the generalization ability of the model under small sample conditions. By constructing a multi-branch time-series fusion network model for aging index prediction, it can simultaneously capture multiple time-series features of the aging process, enabling the model to learn the dynamic evolution trend of surge arrester aging and significantly improving the prediction accuracy of the multi-branch time-series fusion network model. By inputting the time-series feature data of the surge arrester to be evaluated into the trained multi-branch time-series fusion network model, the aging degree assessment result of the surge arrester to be evaluated can be obtained quickly and accurately.

[0042] During implementation, time-series characteristic data and corresponding aging index data of multiple surge arresters are collected to construct an initial sample set.

[0043] Assuming the time window length of the timing characteristic data is T, and the length of one time step is, for example, one week, then one moment corresponds to one week of characteristic data. For example, T=12 corresponds to three months of weekly measurement data. The collected timing characteristic data of the surge arrester at each moment includes the fundamental amplitude of the resistive current, the odd-even harmonic ratio of the resistive current, the approximate power consumption, and the aging sensitivity index.

[0044] During implementation, the timing characteristic data of the surge arrester is obtained through the following methods: S11. At each moment, synchronously collect the voltage signal and total leakage current signal across the surge arrester; S12. Based on the discrete Fourier transform, extract the voltage amplitude, voltage phase angle, current amplitude, and current phase angle corresponding to each harmonic from the voltage signal and the total leakage current signal, respectively. S13. Based on the voltage amplitude, voltage phase angle, current amplitude and current phase angle corresponding to each harmonic, the capacitive current component corresponding to each harmonic is obtained by orthogonal projection calculation and temperature frequency compensation model. S14. The capacitive current components corresponding to each harmonic are superimposed to obtain the total capacitive current, and the resistive current component of the surge arrester is decoupled from the total leakage current signal based on the total capacitive current. S15. Based on the resistive current component at each moment, obtain the characteristic data of that moment, and obtain the timing characteristic data of the surge arrester.

[0045] Under normal operating conditions, a small leakage current path exists in the surge arrester. When operating in the low current region, methods such as... Figure 2 The simplified model shown provides an electrical representation of the surge arrester. In this simplified model, the surge arrester is represented by a capacitor (C) connected in parallel with a nonlinear resistor (R). Figure 2 V is the voltage applied by the surge arrester, and I is... T It is the total leakage current, I R It is the resistive current component, I C It is the capacitive current component. Among them, the resistive current component and its harmonics are key indicators for assessing the aging of surge arresters. Extracting the resistive current component from the total leakage current is very important for surge arrester condition monitoring.

[0046] The decoupling of the resistive current components depends primarily on the orthogonality between the two current components. Figure 2 In the simplified equivalent model, the total leakage current (I T ) is a nonlinear resistive current (I R ) and capacitive current (IC The vector sum of the components can be expressed as follows: Figure 3 The vector image shown. Figure 3 In, θ v It is the phase angle of the applied voltage, which lies between the voltage and the reference vector; θ i It is the phase angle of the total leakage current, which lies between the leakage current and the reference vector; the resistive current component is in phase with the applied voltage, but the capacitive current component is perpendicular to the applied voltage. According to Figure 3 The phase angle between the total leakage current and the resistive current is (θ). i -θ v The resistive current component can be calculated from the total leakage current.

[0047] However, as is well known, any voltage source contains harmonics, so the actual applied voltage also includes multiple harmonic components. Existing technology extracts the resistive current component from the total leakage current based on a simplified equivalent model and using only a pure sine wave applied voltage, which significantly reduces the accuracy of the data. Therefore, to accurately obtain the resistive component from the total leakage current, the errors caused by voltage harmonics must be accurately eliminated.

[0048] During implementation, the voltage signal across the surge arrester is collected simultaneously. and total leakage current signal . Indicates the first The voltage at each sampling point 。 indicates the first The total leakage current at each sampling point. Anti-aliasing filtering and synchronous sampling are performed on the signal. To ensure accurate separation of voltage harmonics in subsequent signal processing, the sampling frequency is... f s It must satisfy Nyquist's theorem, that is f s ≥2 N f 0, where f 0 represents the fundamental frequency of the voltage. N This represents the highest harmonic order for analysis. For power grid signals within my country, the fundamental voltage frequency... f 0 is typically 50Hz. Furthermore, based on experience, the amplitude of voltage harmonic components above the 7th order is negligible. Therefore, it is preferable to analyze only the influence of voltage harmonics up to the 7th order, i.e., the highest harmonic order to analyze. N The preferred value is 7.

[0049] Subsequently, the Discrete Fourier Transform (DFT) was used to extract the voltage amplitude, voltage phase angle, current amplitude, and current phase angle corresponding to each harmonic from the voltage signal and the total leakage current signal, respectively.

[0050] Specifically, the Discrete Fourier Transform (DFT) is a mathematical tool for transforming a finite-length discrete signal from the time domain to the frequency domain. It discretizes the spectrum by sampling the signal in the frequency domain, thus facilitating computer processing. A typical definition of the DFT is: ; in, Let represent the complex spectrum at the nth frequency point in the frequency domain. This represents the k-th sampling point of the input signal, where K represents the number of sampling points. For example, for a voltage signal, V [1] can represent the fundamental voltage component. V [ 3 This can represent the third voltage harmonic component; for the total leakage current signal, I [ 5 This can represent the total leakage current component corresponding to the 5th voltage harmonic. I [ 7 This can represent the total leakage current component corresponding to the 7th voltage harmonic; Represents the first discrete-time signal k Each sampled value contains harmonic amplitude (mode) and phase (argument) information; it refers to the voltage of the surge arrester. or total leakage current A sequentially arranged sampling sequence; k Indicates the index of the time-domain sampling point (0≤ k ≤ K 1) Corresponding sampling time = k / f s Or rather, the sampling time Corresponding sampling point index k = ·f s Therefore, it can be seen that Voltage collected at all times or total leakage current The corresponding sampling sequence V( to DFT) k )or middle; m Represents the frequency point index in the frequency domain (1≤ n ≤ N As mentioned above, the highest harmonic order analyzed in this embodiment of the invention is...N The preferred value is 7; K This indicates the total length of the sampling sequence, i.e., the number of samplings during synchronous signal acquisition; j It is the imaginary unit; 2π / K This represents the radian interval corresponding to the frequency resolution.

[0051] Therefore, by applying Discrete Fourier Transform (DFT) to the synchronously acquired voltage signal and total leakage current signal to obtain the corresponding harmonic components, and then calculating the amplitude and phase angle according to the polar coordinate representation corresponding to the complex form of each harmonic component, the amplitude of the voltage harmonic components from the fundamental wave to the 7th harmonic can be obtained. (1≤ n ≤ 7 and phase angle ; and the amplitudes of the current harmonic components corresponding to the fundamental frequency up to the 7th harmonic. and phase angle .

[0052] Specifically, the plural Represented in polar coordinates, let

[0053] in, Indicates the real part, Indicates the imaginary part. Amplitude Phase angle: .

[0054] Substituting the harmonic components of the voltage signal into the above calculation formula yields the amplitude and phase angle of each harmonic component. Similarly, substituting the harmonic components of the total leakage current signal into the above calculation formula yields the amplitude and phase angle of each harmonic component.

[0055] according to Figure 3 As illustrated, to accurately determine the resistive current component, the capacitive current component needs to be subtracted from the total leakage current. In the embodiments of the present invention, accurate calculation of the capacitive current component is achieved, wherein the theoretical capacitive current is first calculated through dynamic compensation of temperature-frequency coupling capacitor parameters.

[0056] In a preferred embodiment of the present invention, an equivalent capacitance of the surge arrester is established. With harmonic order n and ambient temperature A variable nonlinear model (temperature-frequency compensation model): ; Among them, temperature is acquired in real time through a temperature sensor. , Indicates reference temperature The fundamental capacitance value below, α For temperature coefficient, β These are frequency characteristic coefficients. The two coefficients are determined through a standardized offline calibration process for characteristic parameters of a specific type of surge arrester.

[0057] The specific implementation process is as follows: In the laboratory, a sample of the same model of surge arrester was placed in a temperature-controlled chamber and connected to a high-precision impedance analyzer and a programmable voltage source. Multiple temperature points were selected within the typical operating temperature range of the surge arrester. (e.g., -10°C, 10°C, 30°C, 50°C, 70°C, etc.). At each temperature point, a frequency covering the fundamental to the 7th harmonic is applied. f n Sinusoidal voltage signals at (50Hz, 100Hz, 150Hz, 200Hz, 250Hz, 300Hz, 350Hz).

[0058] In each (temperature) ,frequency f n Under the combined conditions, a high-precision impedance analyzer is used to directly measure the complex admittance of the surge arrester sample and accurately calculate its equivalent capacitance value. .

[0059] Specify a reference temperature (Typically 25°C), and the fundamental (50Hz) capacitance measured at this temperature is used as the reference capacitance. C 0.

[0060] All measurements in the entire test matrix and C Substituting 0 into the above nonlinear model yields an overdetermined system of equations.

[0061] A nonlinear least squares method is used to perform global curve fitting on this system of equations. The goal of the fitting algorithm is to find an optimal set of curves. α and β The parameter values ​​are such that all capacitance values ​​calculated by the nonlinear model are... Compared with measured values The overall error between them is the smallest.

[0062] Substituting the optimal parameters into the nonlinear model yields a stable model output, i.e., the output is determined by any given harmonic order. n and ambient temperature The equivalent capacitance value of this type of surge arrester can be obtained directly. .

[0063] Based on the equivalent capacitance value obtained from this model, the theoretical amplitude of the capacitive current under each harmonic can be calculated. This theoretical value will be used for subsequent verification and correction. ; Where n is the harmonic order; ω refers to the angular frequency of the fundamental wave, ω=2. πf 0, f 0 represents the fundamental frequency of the power system; V max,n It refers to the amplitude (peak value) of the nth harmonic voltage.

[0064] The actual values ​​of the capacitive current components under each harmonic are calculated and extracted based on orthogonal projection, and then corrected using the theoretical values ​​obtained from the aforementioned temperature-frequency compensation. Specifically, the difference in phase angle between the current and voltage signals of each harmonic component is calculated based on the results obtained after the aforementioned discrete Fourier transform. θ in - θ vn ); θ in This indicates the phase of the nth harmonic current. θ vn This represents the phase of the nth harmonic voltage.

[0065] Based on the difference in phase angles, the geometric projection method is applied (see principle). Figure 3 (Illustration), from which the resistive current component under the nth harmonic can be obtained. capacitive current component and total leakage current The relationship between them: ; ; From the above relationship, the maximum projected value of the capacitive current component under the nth harmonic can be calculated. : .

[0066] The theoretical value was then calculated using the aforementioned temperature frequency compensation model. For the orthogonal decomposition results Verification and correction: When the system is running normally and the data is highly reliable, the two values ​​should be within the allowable error range (e.g., the relative deviation between the two values ​​is 1% to 5%), proving that the projection result is accurate. The orthogonal projection value can be directly used as the actual value of the capacitive current component under the current nth harmonic. If the deviation is large, it indicates that under the current temperature and harmonic environment, the pure geometric projection result may produce errors due to model mismatch. In this case, it is preferable to use the theoretical calculation value for correction, that is, to use the theoretical calculation value as the actual value of the capacitive current component under the current nth harmonic. Alternatively, a calibration alarm may be issued for recalculation.

[0067] Furthermore, based on the waveform correspondence between capacitance and voltage (the phase always leads the voltage by π / 2), the capacitive current component under the nth harmonic is obtained relative to time. Waveform function: ; Here, n and ω have the same meaning as before; n is the harmonic order; ω refers to the angular frequency of the fundamental wave, ω=2. πf 0, f 0 represents the fundamental frequency of the power system; θ vn Indicates the phase of the nth harmonic voltage; The total capacitive current is obtained by superimposing the capacitive current components under each harmonic: ; N The highest harmonic order for analysis is preferably 7.

[0068] Finally, subtracting the total capacitive current from the total leakage current yields the highly accurate resistive current component: .

[0069] Through the above method, the embodiments of the present invention achieve high-precision decoupling extraction of the resistive component in the total leakage current, which can be used to accurately monitor and evaluate the operating status of the surge arrester. Figure 4 The comparative effects of the methods in the embodiments of the present invention are shown, from... Figure 4 As can be seen from the decoupling effect comparison diagram, the resistive current obtained by the phasor analysis method proposed in this invention (black solid line in the third figure) closely matches the actual resistive current (gray dashed line in the third figure), accurately restoring the third harmonic characteristics. In contrast, the traditional fixed capacitance method (black dotted line in the third figure) exhibits significant deviations due to neglecting harmonics and frequency characteristics. Furthermore, the capacitive current decoupled by the phasor method of this invention (black dotted line in the second figure) also perfectly exhibits the orthogonal characteristics of a 90° phase shift, demonstrating that the method of this embodiment maintains superiority in harmonic processing accuracy and phase.

[0070] After obtaining the resistive current component at each moment, the characteristic data of that moment is obtained based on the resistive current component at each moment, thus obtaining the timing characteristic data of the surge arrester.

[0071] The fundamental amplitude of resistive current is the most direct and crucial indicator for assessing the aging condition of an MOA (Metal Oxide Arrangement). An increase in its value directly indicates deterioration in the insulation performance of the valve plate and an increase in leakage current.

[0072] During the aging process, the odd harmonics (especially the 3rd, 5th, and 7th harmonics) of resistive current typically increase at a higher rate than even harmonics. The odd-even harmonic ratio of resistive current can effectively reflect this trend and avoid random interference that may occur to a single harmonic.

[0073] Increased power consumption is an important indicator of device overheating and aging.

[0074] To physically approximate the increase in energy loss due to aging, the feature data also includes an aging sensitivity index.

[0075] Therefore, the characteristic data for each moment include the fundamental amplitude of the resistive current, the odd-even harmonic ratio of the resistive current, the approximate power consumption, and the aging sensitivity index.

[0076] The characteristic data for each moment are obtained based on the resistive current component, including: Based on the discrete Fourier transform, the resistive current amplitude and resistive current phase angle corresponding to each harmonic are extracted from the decoupled resistive current components. The fundamental amplitude of the resistive current and the odd-even harmonic ratio of the resistive current are extracted based on the resistive current amplitude corresponding to each harmonic. The approximate power consumption of the surge arrester is calculated based on the resistive current amplitude, resistive current phase angle, voltage amplitude, and voltage phase angle corresponding to each harmonic.

[0077] In practice, similarly, based on the Discrete Fourier Transform (refer to the aforementioned DFT explanation), the complex spectrum of each frequency point of the resistive current component is obtained. The amplitude is calculated using polar coordinates. and phase angle Extracting the fundamental amplitude of resistive current. .

[0078] resistive current odd-even harmonic ratio .

[0079] Adding a minimum value σ to the denominator avoids division by zero errors and ensures that the ratio remains meaningful when even harmonics are close to 0.

[0080] The power consumption of the MOA is approximately calculated by summing the active power generated by each harmonic. Increased power consumption is an important indicator of equipment heating and aging.

[0081]

[0082] in, It is the effective value of the nth harmonic voltage obtained earlier. It is its phase angle. It is the power factor of the nth harmonic; for the resistive component, this value should be close to 1.

[0083] After obtaining the fundamental amplitude of the resistive current and the odd-even harmonic ratio of the resistive current at time t, and approximating the power consumption, the aging sensitivity index at time t is obtained using the following method:

[0084] in, This represents the aging sensitivity index at time t. This represents the fundamental amplitude of the resistive current at time t. This represents the odd-even harmonic ratio of the resistive current at time t. This represents the approximate power consumption at time t. This indicates the fundamental resistive current of the surge arrester at the time of manufacture. This indicates the approximate power consumption of the surge arrester at the time of manufacture.

[0085] In practice, the aging index of the surge arrester at each moment can be quantified by the dielectric loss area. Dielectric loss area ( This method belongs to a different diagnostic dimension than the existing resistive current method. It is obtained by analyzing the frequency domain dielectric spectrum (FDS), which reflects the overall dielectric response characteristics of the surge arrester varistor over a wide frequency range, and thus can quantify its inherent physical aging degree.

[0086] During implementation, the dielectric loss area of ​​the surge arrester will be measured ( The aging index of the surge arrester is obtained by normalization, ranging from 0 to 1. Time-series characteristic data of multiple surge arresters are also included. and corresponding aging index data This constitutes the initial sample set.

[0087] Feature data at time t Including the fundamental amplitude of resistive current resistive current odd-even harmonic ratio Approximate power consumption and aging sensitivity index .

[0088] In practice, it is difficult to obtain a large number of surge arrester samples with real aging labels (especially severely aged samples), resulting in sparse data and a tendency for the model to overfit. Therefore, data augmentation is performed on the initial sample set to obtain a training sample set.

[0089] Specifically, the initial sample set is augmented using the following methods to obtain the training sample set: S21. Construct a conditional generative adversarial network, including a generator and a discriminator; The generator is used to generate time-series feature data based on the aging index and the aging sensitivity index; The discriminator is used to determine whether the input time-series feature data and the corresponding aging index data are real data; S22. Train the conditional generative adversarial network based on the initial sample set to obtain a trained conditional generative adversarial network; S23. Generate samples based on the trained conditional generative adversarial network and perform data augmentation on the initial sample set.

[0090] A conditional generative adversarial network consists of a generator and a discriminator.

[0091] The generator uses the aging index and aging sensitivity index at T time points as conditional information to generate time-series feature data. In implementation, the generator takes a random noise vector and conditional information as input, passes it through a fully connected layer to map the noise and conditions into an intermediate vector, and then uses a transposed convolutional network or a temporal convolutional network (TCN) to generate the data. The four-dimensional features of each time step are used to generate time series feature data.

[0092] The discriminator takes temporal feature data and the corresponding aging index at T time points as input and determines whether the input data is real data. In implementation, the discriminator can use a temporal convolutional network structure to extract temporal features, and then output the probability that the data is real after passing through a fully connected layer and a sigmoid function.

[0093] During implementation, the training loss of the conditional generative adversarial network is calculated using the following formula:

[0094] in, This indicates the generation of adversarial loss. Indicates the time-series consistency loss. This represents the balance coefficient.

[0095] in, This is the existing standard formula for calculating adversarial loss. This invention adds temporal consistency loss to the adversarial loss, making the adjacent variation patterns (i.e., difference vectors) of the generated sequence as close as possible to the adjacent variation patterns of the real sequence. The generated sequence data thus more closely resembles the slow drift characteristics of the real aging process.

[0096] Specifically, the timing consistency loss is calculated using the following formula:

[0097] in, Expressing expectations, This represents the k-th difference vector calculated from the real sample; This represents the k-th difference vector calculated from the generated samples. Let T represent the square of the L2 norm of the vector, and let T represent the sequence length of the real sample.

[0098] The first difference of the real sequence is: ; The first difference of the generated sequence is divided into: .

[0099] The temporal consistency loss forces the two differencing groups to be statistically similar. Therefore, the temporal consistency loss is determined by the mean squared difference mentioned above.

[0100] In practice, a conditional generative adversarial network can be trained based on an initial sample set. When the conditional generative adversarial network converges, training is stopped, and a well-trained conditional generative adversarial network is obtained. A pre-trained conditional generative adversarial network can be used to generate samples and augment the initial sample set. For example, for the final aging index... A small number of real samples are used to generate multiple samples by inputting their aging index and aging sensitivity index into a pre-trained conditional generative adversarial network. These samples are then mixed with real samples to form an enhanced training set, resulting in the training sample set.

[0101] Furthermore, a multi-branch time-series fusion network model is constructed. This model is then trained using a training sample set to obtain a well-trained multi-branch time-series fusion network model. The time-series characteristic data of the surge arrester to be evaluated is then input into the trained multi-branch time-series fusion network model for aging index prediction, thus quickly and accurately obtaining the aging degree assessment result of the surge arrester.

[0102] The multi-branch time-series fusion network model outputs a continuous aging index in the range of 0 to 1. Compared with the traditional discrete level, it can reflect the gradual aging process more precisely, which is conducive to formulating differentiated maintenance strategies (such as early warning, trend analysis, etc.).

[0103] Specifically, multi-branch temporal fusion network models include: Gated loop unit branches are used to extract long-term temporal dependency features of the input data; Temporal convolutional network branches are used to extract multi-scale local features from the input data; Causal self-attention branch is used to extract global interaction features of the input data; The gated fusion layer is used to adaptively weight and fuse the features extracted from each branch, and to predict the aging index based on the fused features.

[0104] The multi-branch temporal fusion network model extracts time features at different levels through three parallel branches and integrates them adaptively using a gating fusion mechanism, thereby improving the model's ability to model complex aging processes and increasing prediction accuracy.

[0105] The input data for the multi-branch time-series fusion network model is the time-series characteristic data of the surge arrester.

[0106] Specifically, the gated loop unit branches adopt a bidirectional single-layer gated loop unit; The first-level gated loop unit is used to calculate the first hidden state at each time step based on the input data at that time step and the hidden state output by the first-level gated loop unit at the previous time step. The second-level gated loop unit is used to calculate the second hidden state at each time step based on the first hidden feature output by the first-level gated loop unit at that time step and the second hidden state output by the second-level gated loop unit at the previous time step. The second hidden state at each time step is the temporal long-term dependency feature extracted from the gated recurrent unit branch.

[0107] The gated recurrent unit branch is used to extract long-term dependencies from the input data. This branch adopts a two-layer unidirectional GRU. The first layer outputs the complete hidden state sequence to the second layer. The second layer takes the hidden state output by the first layer as input and extracts the second hidden state at each time step as the global representation of the entire sequence.

[0108] Specifically, the first-level gated loop unit extracts the first hidden state using the following formula:

[0109]

[0110]

[0111]

[0112] in, This represents the input data at time t. This represents the hidden state of the output of the first-level gated loop unit at time t-1. , , , , and Both represent trainable weight matrices. , and Both represent bias vectors. This represents the Sigmoid function. This represents the tanh activation function. , and Indicates intermediate parameters. This represents the hidden state of the output of the first-level gated loop unit at time t. The table represents element-wise multiplication.

[0113] The first layer outputs the complete hidden state sequence. .

[0114] It should be noted that, for the first time step, the hidden state output by the first-level gated loop unit in the previous time step... The zero vector can be used.

[0115] The second-layer GRU takes the hidden state sequence of the first layer as input (replacing the original...). The operations are exactly the same, only the weight parameters are different. For the first time step, the hidden state output by the second-layer gated recurrent unit in the previous time step... The zero vector can be used.

[0116] The second layer outputs the complete hidden state sequence. .

[0117] The design of this branch allows GRU to gradually accumulate historical information from front to back, providing long-term dependency features for subsequent fusion layers.

[0118] The goal of temporal convolutional network branches is to extract local patterns at different time scales (e.g., short-term, medium-term, long-term) while ensuring causality.

[0119] Specifically, a temporal convolutional network branch includes at least three parallel sub-branches and a fusion layer; Each sub-branch independently performs dilated causal convolution to capture local features at different scales; The fusion layer is used to concatenate local features of different scales along the feature dimension and obtain multi-scale local features by pointwise convolution.

[0120] In implementation, for example, three parallel sub-branches can be used to capture short-term, medium-term, and long-term local features, respectively. Each sub-branch uses the same kernel size but a different expansion factor.

[0121] Dilated causal convolution introduces a dilation operation on top of causal convolution, which can increase or decrease the receptive field while using only the input at the current time step and before, thus ensuring causality.

[0122] For example, the first sub-branch has a kernel size of 3, an inflation factor of 1, and a receptive field of 3, capturing short-term patterns. The second sub-branch has a kernel size of 3, an inflation factor of 2, and a receptive field of 1 + 2 × 2 = 5, capturing medium-term patterns. The third sub-branch has a kernel size of 3, an inflation factor of 4, and a receptive field of 1 + 2 × 4 = 9, capturing long-term patterns.

[0123] The fusion layer concatenates local features at different scales along the feature dimension:

[0124] in, This represents the local features of the input at time t in the first sub-branch. This represents the local features of the input at time t in the second sub-branch. This represents the local features of the input at time t of the third sub-branch. This indicates the local features after splicing.

[0125] The fusion layer uses pointwise convolution (1×1 convolution) to... Dimensions reduced to (Dimension of input features of the temporal convolutional network branches) to obtain multi-scale local features at time t. .

[0126] Final Temporal Convolutional Network Branch Output .

[0127] Self-attention can directly model the dependency between any two time steps, without being limited by distance. This invention uses causal self-attention branches to extract global interaction features of the input data. Specifically, the causal self-attention branch is used to extract global interaction features of the input data, but it is necessary to ensure that the output at time t can only depend on the input at time t and before.

[0128] During implementation, the input data is first mapped linearly to obtain the query matrix, key matrix, and value matrix.

[0129] To ensure causality, a mask matrix is ​​defined. :

[0130] Represents the mask matrix The tth row, the tth row The elements of the column.

[0131] The causal attention weights are then:

[0132] in, , This represents the vector at row t of the query matrix. The first key matrix represents the key matrix. row vectors This indicates the dimension of the attention head, and the superscript T indicates transpose.

[0133] The output of the single-causal attention head is .

[0134] in, Represents the value matrix of the first Row vectors.

[0135] The outputs of multiple causal attention heads are concatenated and mapped back to the original dimension using a projection matrix. Then, through residual connections, layer normalization, and a feedforward neural network, the causal self-attention features at each time step are obtained. This yields the global interaction features.

[0136] Causal self-attention branch output . This represents the causal self-attention characteristics at time 1. This represents the causal self-attention feature at time T.

[0137] The gated fusion layer projects the features of each time step in the three branches onto the same dimension through the corresponding fully connected layer, and then calculates the gate weights:

[0138] in, This represents vector concatenation. Represents the weight matrix. This represents the bias vector. This represents the feature at time t in the temporal long-term dependency features with dimension alignment. Let represent the feature at time t in the dimension-aligned multi-scale local features. This represents the feature at time t in the dimension-aligned global interaction features. The output is... ,and .

[0139] Fusion features at time t .

[0140] The aging index at time t is predicted based on the fusion features. In practice, the predicted aging index can be output through a fully connected layer + Sigmoid function.

[0141] The multi-branch temporal fusion network model captures short-term fluctuations (TCN small expansion rate), long-term trends (GRU), and global dependencies across time steps (attention) of the aging process simultaneously through three branches, significantly improving its ability to fit complex aging trajectories. The gated fusion layer enables the model to adaptively select the most reliable branch based on the input data, enhancing robustness.

[0142] In practice, the training loss of the multi-branch temporal fusion network model is calculated using the following formula:

[0143] in, Indicates the predicted loss. This represents the loss due to physical monotonicity constraints. This represents the trend consistency regularization loss. This represents the trainable parameters of a multi-branch temporal fusion network model. The L2 norm of a vector is the square of its value. , and This represents the weighting coefficient.

[0144] The training loss consists of three parts: data fitting loss, physical monotonicity constraint loss, and trend consistency regularization loss. To prevent overfitting, L2 regularization loss is also added.

[0145] In practice, the data fitting loss (prediction loss) is the loss between the predicted aging index and the actual aging index.

[0146] In implementation, a weighted mean square error can be used, assigning greater weight to samples with higher aging levels:

[0147] in, This indicates the number of samples in the current training batch. This represents the weight of the i-th sample. Let represent the aging index predicted by the model for the i-th sample at time T. Let represent the true aging index of the i-th sample at the T-th time.

[0148] The prediction at time T is the result that the model needs to output during inference; therefore, supervision is only applied to the prediction at the last time step.

[0149] During implementation, , A value of 0.1 can be used to double the weight of severely aged samples.

[0150] According to common knowledge about surge arrester aging, the following physical monotonicity constraints should be met: If the resistive current fundamental amplitude, odd-even harmonic ratio, approximate power consumption, and aging sensitivity index of input sample A are all greater than or equal to the corresponding values ​​of sample B (and at least one is strictly greater than), then the predicted aging index should be: .

[0151] During implementation, pairs of real samples are randomly sampled from each batch, and only those pairs with a clear monotonic relationship are retained to construct a set of sample pairs. That is, sample pairs that meet the following requirements: , , , .

[0152] The physical monotonicity constraint loss is:

[0153] in, Represents the set of sample pairs The number of sample pairs in the middle This represents the aging index predicted by the model for the j-th sample at time T. Let represent the aging index predicted by the model for the i-th sample at time T. Represents the marginal parameter. This represents the maximum difference in the aging sensitivity index within the sample set. This represents the aging sensitivity index of the i-th sample. This represents the aging sensitivity index of the j-th sample.

[0154] During implementation, the magnitude of the difference in aging sensitivity index is used as the weight. If the two samples... If the difference is large, the difference in the predicted aging value should also be relatively large. It is a training set The maximum possible difference (between real samples). Through... The magnitude of the difference directly adjusts the marginal requirements, making the constraints more reasonable.

[0155] Marginal parameters For example, the value is 0.05.

[0156] For time series samples, if the true aging index is known as the value at the end of the time window, it is generally expected that the aging index should monotonically increase with the feature within the time window.

[0157] Specifically, the trend consistency regularization loss is calculated using the following formula:

[0158] in, This represents the aging index predicted by the model for the i-th sample at time t. The aging index of the i-th sample at time t+1, as predicted by the model, is represented by T, which represents the sequence length. This indicates the number of samples in the current training batch.

[0159] Trend consistency regularization loss through penalty This phenomenon forces the aging index sequence output by the model to satisfy... The monotonically non-decreasing relationship of the model makes its predictions consistent with generally accepted physical facts, significantly enhancing the credibility and acceptability of the predictions, which is especially crucial for power operation and maintenance decisions that require highly reliable judgments.

[0160] A specific embodiment of the present invention discloses a surge arrester aging assessment system, characterized in that it includes: The sample set construction module is used to collect time-series characteristic data and corresponding aging index data of multiple surge arresters to construct an initial sample set. The data augmentation module is used to augment the initial sample set to obtain the training sample set; The model training module is used to construct a multi-branch temporal fusion network model for aging index prediction; the constructed multi-branch temporal fusion network model is trained based on the training sample set to obtain a trained multi-branch temporal fusion network model. The evaluation module is used to input the time-series characteristic data of the surge arrester to be evaluated into the trained multi-branch time-series fusion network model to predict the aging index and obtain the aging degree evaluation result of the surge arrester to be evaluated.

[0161] The above-described method and system embodiments are based on the same principles, and their related aspects can be referenced from each other to achieve the same technical effects. For specific implementation processes, please refer to the foregoing embodiments, which will not be repeated here.

[0162] Electronic device example: One specific embodiment of this application discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the surge arrester aging assessment method in the method embodiment.

[0163] Examples of readable storage media: One specific embodiment of this application discloses a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the surge arrester aging assessment method in the method embodiment.

[0164] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0165] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the aging of surge arresters, characterized in that, Includes the following steps: An initial sample set was constructed by collecting time-series characteristic data and corresponding aging index data of multiple surge arresters. Data augmentation is performed on the initial sample set to obtain the training sample set; A multi-branch temporal fusion network model was constructed for aging index prediction. The multi-branch temporal fusion network model is trained based on the training sample set to obtain the trained multi-branch temporal fusion network model. The time-series characteristic data of the surge arrester to be evaluated is input into the trained multi-branch time-series fusion network model to predict the aging index and obtain the aging degree evaluation result of the surge arrester to be evaluated. The aging index is continuous data. The multi-branch temporal fusion network model includes: Gated loop unit branches are used to extract long-term temporal dependency features of the input data; Temporal convolutional network branches are used to extract multi-scale local features from the input data; Causal self-attention branch is used to extract global interaction features of the input data; A gated fusion layer is used to adaptively weight and fuse the features extracted from each branch, and to predict the aging index based on the fused features. The gated loop unit branch adopts a bidirectional single-layer gated loop unit; The first-level gated loop unit is used to calculate the first hidden state at each time step based on the input data at that time step and the hidden state output by the first-level gated loop unit at the previous time step. The second-level gated loop unit is used to calculate the second hidden state at each time step based on the first hidden feature output by the first-level gated loop unit at that time step and the second hidden state output by the second-level gated loop unit at the previous time step. The second hidden state at each time step is the temporal long-term dependency feature extracted from the gated recurrent unit branch; The first-level gated loop unit extracts the first hidden state using the following formula: in, This represents the input data at time t. This represents the hidden state of the output of the first-level gated loop unit at time t-1. , , , , and Both represent trainable weight matrices. , and Both represent bias vectors. This represents the Sigmoid function. This represents the tanh activation function. , and Indicates intermediate parameters. This represents the hidden state of the output of the first-level gated loop unit at time t. Element-by-element multiplication; The training loss of a multi-branch temporal fusion network model is calculated using the following formula: in, Indicates the predicted loss. This represents the loss due to physical monotonicity constraints. This represents the trend consistency regularization loss. This represents the trainable parameters of a multi-branch temporal fusion network model. The L2 norm of a vector is the square of its value. , and This represents the weighting coefficient.

2. The surge arrester aging assessment method according to claim 1, characterized in that, The initial sample set is augmented using the following method to obtain the training sample set: Construct a conditional generative adversarial network, including a generator and a discriminator; The generator is used to generate time-series feature data based on the aging index and the aging sensitivity index; The discriminator is used to determine whether the input time-series feature data and the corresponding aging index data are real data; The conditional generative adversarial network is trained based on the initial sample set to obtain a trained conditional generative adversarial network. Samples are generated based on a pre-trained conditional generative adversarial network, and data augmentation is performed on the initial sample set.

3. The surge arrester aging assessment method according to claim 2, characterized in that, The training loss of the conditional generative adversarial network is calculated using the following formula: in, This indicates the generation of adversarial loss. Indicates the time-series consistency loss. This represents the balance coefficient.

4. The surge arrester aging assessment method according to claim 3, characterized in that, The timing consistency loss is calculated using the following formula: in, Expressing expectations, This represents the k-th difference vector calculated from the real sample; This represents the k-th difference vector calculated from the generated samples. Let T represent the square of the L2 norm of the vector, and let T represent the sequence length of the real sample.

5. The surge arrester aging assessment method according to claim 1, characterized in that, The temporal convolutional network branch includes at least three parallel sub-branches and a fusion layer; Each sub-branch independently performs dilated causal convolution to capture local features at different scales; The fusion layer is used to concatenate local features of different scales along the feature dimension and obtain multi-scale local features by pointwise convolution.

6. The surge arrester aging assessment method according to claim 5, characterized in that, The physical monotonicity constraint loss is calculated using the following formula: in, Represents the set of sample pairs The number of sample pairs in the middle This represents the aging index predicted by the model for the j-th sample at time T. Let represent the aging index predicted by the model for the i-th sample at time T. Represents the marginal parameter. This represents the maximum difference in the aging sensitivity index within the sample set. This represents the aging sensitivity index of the i-th sample. This represents the aging sensitivity index of the j-th sample.

7. The surge arrester aging assessment method according to claim 6, characterized in that, The trend consistency regularization loss is calculated using the following formula: in, This represents the aging index predicted by the model for the i-th sample at time t. The aging index of the i-th sample at time t+1, as predicted by the model, is represented by T, which represents the sequence length. This indicates the number of samples in the current training batch.

8. The surge arrester aging assessment method according to claim 1, characterized in that, The timing characteristic data of the surge arrester are obtained using the following method: At each moment, the voltage signal and total leakage current signal across the surge arrester are collected synchronously; Based on the Discrete Fourier Transform, the voltage amplitude, voltage phase angle, current amplitude, and current phase angle corresponding to each harmonic are extracted from the voltage signal and the total leakage current signal, respectively. Based on the voltage amplitude, voltage phase angle, current amplitude, and current phase angle corresponding to each harmonic, the capacitive current components corresponding to each harmonic are obtained through orthogonal projection calculation and temperature-frequency compensation model. The capacitive current components corresponding to each harmonic are superimposed to obtain the total capacitive current, and the resistive current component of the surge arrester is decoupled from the total leakage current signal based on the total capacitive current. Based on the resistive current component at each moment, the characteristic data of that moment are obtained, and the timing characteristic data of the surge arrester are obtained.

9. The surge arrester aging assessment method according to claim 8, characterized in that, The characteristic data at this moment include the fundamental amplitude of the resistive current, the odd-even harmonic ratio of the resistive current, the approximate power consumption, and the aging sensitivity index.

10. The surge arrester aging assessment method according to claim 9, characterized in that, The characteristic data for each moment are obtained based on the resistive current component, including: Based on the discrete Fourier transform, the resistive current amplitude and resistive current phase angle corresponding to each harmonic are extracted from the decoupled resistive current components. The fundamental amplitude of the resistive current and the odd-even harmonic ratio of the resistive current are extracted based on the resistive current amplitude corresponding to each harmonic. The approximate power consumption of the surge arrester is calculated based on the resistive current amplitude, resistive current phase angle, voltage amplitude, and voltage phase angle corresponding to each harmonic.

11. The surge arrester aging assessment method according to claim 10, characterized in that, The aging sensitivity index is obtained using the following method: in, This represents the aging sensitivity index at time t. This represents the fundamental amplitude of the resistive current at time t. This represents the odd-even harmonic ratio of the resistive current at time t. This represents the approximate power consumption at time t. This indicates the fundamental resistive current of the surge arrester at the time of manufacture. This indicates the approximate power consumption of the surge arrester at the time of manufacture.

12. A surge arrester aging assessment system, characterized in that, include: The sample set construction module is used to collect time-series characteristic data and corresponding aging index data of multiple surge arresters to construct an initial sample set. The data augmentation module is used to augment the initial sample set to obtain the training sample set; The model training module is used to build a multi-branch temporal fusion network model for aging index prediction. The multi-branch temporal fusion network model is trained based on the training sample set to obtain the trained multi-branch temporal fusion network model. The evaluation module is used to input the time-series characteristic data of the surge arrester to be evaluated into the trained multi-branch time-series fusion network model to predict the aging index and obtain the aging degree evaluation result of the surge arrester to be evaluated. The aging index is continuous data. The multi-branch temporal fusion network model includes: Gated loop unit branches are used to extract long-term temporal dependency features of the input data; Temporal convolutional network branches are used to extract multi-scale local features from the input data; Causal self-attention branch is used to extract global interaction features of the input data; A gated fusion layer is used to adaptively weight and fuse the features extracted from each branch, and to predict the aging index based on the fused features. The gated loop unit branch adopts a bidirectional single-layer gated loop unit; The first-level gated loop unit is used to calculate the first hidden state at each time step based on the input data at that time step and the hidden state output by the first-level gated loop unit at the previous time step. The second-level gated loop unit is used to calculate the second hidden state at each time step based on the first hidden feature output by the first-level gated loop unit at that time step and the second hidden state output by the second-level gated loop unit at the previous time step. The second hidden state at each time step is the temporal long-term dependency feature extracted from the gated recurrent unit branch; The first-level gated loop unit extracts the first hidden state using the following formula: in, This represents the input data at time t. This represents the hidden state of the output of the first-level gated loop unit at time t-1. , , , , and Both represent trainable weight matrices. , and Both represent bias vectors. This represents the Sigmoid function. This represents the tanh activation function. , and Indicates intermediate parameters. This represents the hidden state of the output of the first-level gated loop unit at time t. Element-by-element multiplication; The training loss of a multi-branch temporal fusion network model is calculated using the following formula: in, Indicates the predicted loss. This represents the loss due to physical monotonicity constraints. This represents the trend consistency regularization loss. This represents the trainable parameters of a multi-branch temporal fusion network model. The L2 norm of a vector is the square of its value. , and This represents the weighting coefficient.

13. An electronic device, characterized in that, The electronic device includes a processor and a memory; wherein the memory stores a computer program, which, when executed by the processor, implements the steps of the surge arrester aging assessment method as described in any one of claims 1-11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the surge arrester aging assessment method as described in any one of claims 1-11.

Images