An impedance measurement method based on KAN neural network and continuous disturbance injection device

Abstract

The present invention discloses an impedance measurement method based on a KAN neural network and a continuous disturbance injection device, which relates to the technical field of power grid operation analysis, and includes the following steps: S1, setting disturbance parameters in a continuous disturbance injection device, and connecting the continuous disturbance injection device; S2, continuously injecting disturbance signals, and obtaining the response voltage and response current of the grid connection point of the system to be measured; S3, calculating the impedance amplitude and phase at the current disturbance frequency based on the spectrum recognition algorithm of the KAN neural network; S4, performing impedance calculation at the next frequency point until the total number of disturbance injections is reached, and completing the measurement of the impedance of the system to be measured; S5, displaying the impedance calculation results at different frequencies through visualization. The present invention significantly improves the accuracy and stability of impedance measurement in the power system by adopting the KAN neural network, and enhances the anti-noise performance and nonlinear signal processing capability.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid operation analysis, and in particular to an impedance measurement method based on a KAN neural network and a continuous disturbance injection device. Background Art

[0002] With the expansion of the scale of power systems and the widespread access to renewable energy, the nonlinear characteristics and complexity of power systems have gradually increased, which has put forward higher requirements for the impedance characteristics of power systems. Impedance characteristics are a measure of the ability of power system transmission lines and equipment to respond to electrical signals. It affects the power quality, short-circuit current, system resonance, and the working efficiency of protection equipment. Studying the impedance characteristics of power systems not only helps to evaluate the stability of system oscillations, prevent resonance and electromagnetic interference, but also provides a theoretical basis for designing appropriate protection strategies, optimizing grid operation, and maintaining power equipment. Especially in the application of new technologies such as new energy grid connection and flexible AC transmission technology (FACTS), it is particularly important to accurately grasp the impedance characteristics of power systems. Therefore, impedance measurement and analysis are of great significance to ensure the safe and efficient operation of power systems in complex operating environments.

[0003] In the prior art, impedance analysis generally uses the fast Fourier transform (FFT) algorithm to extract the voltage disturbance component and current disturbance component corresponding to the disturbance frequency in a section of the voltage and current waveform at the disturbance injection point. However, the spectrum leakage and fence effect of FFT may cause the extracted disturbance components to be inaccurate, affecting the accuracy of impedance analysis. At present, there are also technologies that use methods based on singular value decomposition (SVD) or ESPRIT to analyze the spectrum of complex signals, but when using these algorithms, there may be problems such as incorrect judgment of the number of harmonic components, too long time to calculate high-dimensional matrices resulting in the inability to monitor in real time, strong noise interference, and weak nonlinear signal processing capabilities. Summary of the invention

[0004] In order to solve the problems of poor accuracy of impedance analysis, strong noise interference and weak nonlinear signal processing capability in the existing power system, the present invention proposes an impedance measurement method based on a KAN neural network and a continuous disturbance injection device to solve the above problems.

[0005] The present application discloses an impedance measurement method based on a KAN neural network and a continuous disturbance injection device, comprising the following steps:

[0006] S1. Setting disturbance parameters in a continuous disturbance injection device, determining the total number of disturbance injections, and connecting the continuous disturbance injection device to the grid connection point of the system to be tested;

[0007] S2. Use a continuous disturbance injection device to continuously inject a disturbance signal with set disturbance parameters, and obtain a response voltage and a response current of a grid connection point of the system to be tested;

[0008] S3, the spectrum recognition algorithm based on KAN neural network calculates the impedance amplitude and phase at the current disturbance frequency;

[0009] S4. According to the frequency injection interval of the continuous disturbance injection device, the impedance calculation of the next frequency point is performed until the total number of disturbance injections is reached, and the impedance measurement of the system to be measured is completed;

[0010] S5. The impedance calculation results at different frequencies are displayed through visualization.

[0011] Preferably, the disturbance parameters include disturbance amplitude, starting frequency, ending frequency, frequency injection interval and duration of each frequency injection.

[0012] Preferably, the continuous disturbance injection device comprises:

[0013] Reading module: used to read the set disturbance parameters. The disturbance parameters are input from the external controller through the input interface of the reading module to ensure the flexibility and controllability of signal generation.

[0014] Amplitude and frequency control module: used to determine the signal amplitude and control the signal frequency change. After the initial signal is generated, starting from the starting frequency, the signal frequency is gradually increased according to the set frequency injection interval and the injection duration of each frequency segment, generating multiple signals with different frequencies.

[0015] Signal generation module: According to the signal provided by the amplitude and frequency control module, a disturbance signal is generated in the time domain following the sinusoidal wave law. The frequency of the generated disturbance signal is the frequency of the signal generated by the amplitude and frequency control module, the amplitude is the set disturbance amplitude, and the duration of each signal segment is the set injection duration. The signal generation module maintains the output of the disturbance signal within the duration to ensure that the disturbance signal can be stably output at this frequency.

[0016] Output module: Injects the disturbance signal generated by the signal generation module into the grid connection point of the system to be tested, ensures the smooth output of the entire disturbance signal, filters out the mutations or noise that may be generated during frequency conversion, and ensures the continuity of the disturbance signal.

[0017] Preferably, the generation of the disturbance signal in S2 comprises the following steps:

[0018] S21, read the set disturbance parameters, and then generate the first signal according to the set starting frequency and disturbance amplitude. The signal is generated in the time domain according to the sine wave, and the duration of the generated signal is the set injection duration of each frequency segment;

[0019] S22, after the first segment signal is generated, gradually increasing the frequency of the signal according to the frequency injection interval to generate signals of multiple frequency segments;

[0020] S23. When the frequency increases to the set end frequency, the entire signal is composed of multiple sine waves of different frequencies. These signals are spliced ​​together in time sequence to finally obtain a disturbance signal whose frequency gradually increases with time.

[0021] According to the disturbance signal generation method, a disturbance voltage signal and a disturbance current signal are respectively injected into the grid connection point of the system to be tested, and the response voltage and response current of the grid connection point of the system to be tested are respectively obtained through a voltage measuring device and a current measuring device (such as a voltmeter and an ammeter).

[0022] Preferably, using a KAN neural network to fit the amplitude, frequency and phase of the response voltage and the response current comprises the following steps:

[0023] S31, input signal sampling modeling, assuming that the response voltage or response current obtained by S2 is , through the sampling period The continuous signal Discretize into ,in is a discrete time point;

[0024] S32, data input, the obtained noisy signal A window length is intercepted as the input of the KAN neural network. The choice of window length depends on the frequency component of the signal and the computing resources. After preprocessing including trend removal, denoising and normalization, the data is input to the input layer of the KAN neural network;

[0025] S33, signal feature decomposition, according to the Kolmogorov-Arnold representation theorem, the KAN neural network can decompose complex time series signals. The initial layer of the KAN neural network maps the signal obtained after S32 preprocessing into different feature functions through multiple intermediate neurons. The feature functions include sine wave components. As well as the linear noise part superimposed on it, the feature decomposition layer is based on the sine wave basis function expansion, which decomposes the noisy signal into the superposition of multiple sine waves and noise;

[0026] S34, nonlinear approximation, in the middle layer of the KAN neural network, nonlinear activation functions are used to perform nonlinear approximation on the feature function obtained in S33. Each activation function corresponds to a nonlinear transformation of an input feature. Through multi-layer mapping, the weight of each activation function is gradually adjusted to make the output signal close to the original signal. The KAN network can flexibly approximate any continuous function, so it can capture the components of the signal. The amplitude, frequency and phase characteristics of each sine wave;

[0027] S35, noise filtering and output fitting, after multiple iterations of training, the output of the KAN neural network gradually approaches the original signal , that is, extract the characteristics of each sinusoidal signal, and model the noise part as a random component. Since the KAN neural network has good nonlinear fitting ability, the final KAN neural network output can restore the signal well. The characteristics of each sine wave in the image, ignoring or minimizing noise interference;

[0028] S36, training optimization, in the fitting process, the loss function is used to measure the deviation between the prediction result and the original signal. If the deviation value is greater than the set tolerance threshold, the back propagation algorithm is used to optimize the learnable parameters in the KAN neural network to approximate the original sinusoidal signal. , reducing the impact of noise on the output.

[0029] Preferably, the activation function of the KAN neural network is a combination of multiple sinusoidal functions and linear functions, and the activation function is as follows:

[0030]

[0031] in, is the magnitude of each sine term, is the angular frequency of each sinusoidal term, is the phase of each sinusoidal term, is the number of sine terms, indicating the number of fitted sine components. , the angular frequency of the sine term , the phase of the sine term are all learnable parameters, Represents the linear terms in the signal.

[0032] Preferably, the signal obtained in S31 Depend on It is composed of a sinusoidal signal and noise. The expression is:

[0033]

[0034] in, is the sampling period, For the The amplitude of the sinusoidal component, For the The initial phase of the sinusoidal components, For the The attenuation coefficient of the sinusoidal component is For the The frequency of the sinusoidal component, is white noise with zero mean.

[0035] Preferably, the loss function in S36 is:

[0036]

[0037] in, is the output of the KAN neural network, is the number of sampling points.

[0038] Preferably, the impedance characteristic of the system to be measured is calculated by the following steps:

[0039] After the original signal is decomposed into multiple sinusoidal waves and noise components through the KAN neural network, the voltage signal and current signal can be represented as the superposition of multiple sinusoidal wave components;

[0040] The voltage signal is expressed as:

[0041]

[0042] in, For frequency The voltage amplitude at For frequency The phase angle of the voltage at is the noise component of the voltage signal.

[0043] For current signals, it can also be expressed as the superposition of multiple sinusoidal wave components:

[0044]

[0045] in, For frequency The current amplitude at For frequency The phase angle of the current at is the noise component of the current signal.

[0046] After being processed by the KAN neural network, each frequency can be accurately extracted The corresponding voltage amplitude and current amplitude , and minimize the noise impact, providing reliable input data for the subsequent impedance calculation.

[0047] Based on the extracted frequency components, the voltage at each frequency point can be and current To calculate the impedance . Using Ohm's law, impedance is calculated as:

[0048]

[0049] in, For frequency The complex impedance at contains amplitude and phase information. The amplitude reflects the total impedance at that frequency, the real part (resistance) represents the energy consumption, and the imaginary part (reactance) reflects the energy storage characteristics. This information is critical for analyzing the electrical characteristics of the system.

[0050] Preferably, the display method in S5 includes an impedance amplitude curve, an impedance phase curve and a Nyquist diagram.

[0051] Beneficial effects of the present invention:

[0052] (1) Improve measurement accuracy: The KAN neural network can decompose noisy voltage and current signals into multiple simple univariate functions by decomposing complex time series signals. Its powerful nonlinear approximation ability enables the system to accurately extract the characteristic components in the signal and reduce measurement errors, thereby significantly improving the accuracy of impedance measurement.

[0053] (2) The present invention effectively models and separates the noise components in the signal, and can stably extract useful signal components in an environment with strong noise interference, thereby improving the measurement stability of the power system under complex noise conditions.

[0054] (3) The present invention can handle various nonlinear factors existing in the power system, flexibly adapt to complex voltage and current waveforms, and accurately extract impedance characteristics.

[0055] (4) The KAN neural network of the present invention has good self-learning ability and can adaptively adjust the parameters of the model according to the operating status and signal characteristics of the power system. As the operating environment of the power system changes, the system can still maintain a high measurement accuracy, and its adaptability and robustness are greatly improved.

[0056] (5) The efficient computing performance of the present invention enables it to quickly process a large amount of signal data and output impedance measurement results in real time, meeting the high requirements for real-time monitoring during power system operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of an impedance measurement method based on a KAN neural network and a continuous disturbance injection device according to an embodiment of the present invention;

[0058] Figure 2 It is a structural schematic diagram of a continuous disturbance injection device according to an embodiment of the present invention;

[0059] Figure 3 It is a schematic diagram of a continuous disturbance injection device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and examples.

[0061] The embodiment of the present application discloses an impedance measurement method based on a KAN neural network and a continuous disturbance injection device, the process is as follows: Figure 1 As shown, the following steps are included:

[0062] S1. Set disturbance parameters in the continuous disturbance injection device. The disturbance parameters include disturbance amplitude, starting frequency, ending frequency, frequency injection interval and injection duration of each frequency segment. Determine the total number of disturbance injections based on the above disturbance parameters. , and then connect the continuous disturbance injection device to the grid connection point of the system under test. Among them, the total number of disturbance injections .

[0063] like Figure 2 As shown, the structure of the continuous disturbance injection device includes:

[0064] Reading module: used to read the set disturbance parameters. The disturbance parameters are input from the external controller through the input interface of the reading module to ensure the flexibility and controllability of signal generation.

[0065] Amplitude and frequency control module: used to determine the signal amplitude and control the signal frequency change. After the initial signal is generated, starting from the starting frequency, the signal frequency is gradually increased according to the set frequency injection interval and the injection duration of each frequency segment, generating multiple signals with different frequencies.

[0066] Signal generation module: According to the signal provided by the amplitude and frequency control module, a disturbance signal is generated in the time domain following the sinusoidal wave law. The frequency of the generated disturbance signal is the frequency of the signal generated by the amplitude and frequency control module, the amplitude is the set disturbance amplitude, and the duration of each signal segment is the set injection duration. The signal generation module maintains the output of the disturbance signal within the duration to ensure that the disturbance signal can be stably output at this frequency.

[0067] Output module: Injects the disturbance signal generated by the signal generation module into the grid connection point of the system to be tested, ensures the smooth output of the entire disturbance signal, filters out the mutations or noise that may be generated during frequency conversion, and ensures the continuity of the disturbance signal.

[0068] The principle of the continuous disturbance injection device is as follows Figure 3 As shown, it includes a ramp function generator CH1, a climbing function generator CH2, a constant input CH3 and an adjustable sine wave generator SSSG1. Figure 3MAG indicates the amplitude setting, FS / TS is the frequency injection interval and its duration, FRES indicates the starting injection frequency, and FREE indicates the ending injection frequency. The ramp function generator CH1 is used to generate a fixed amplitude signal, the ramp function generator CH2 is used to generate a frequency signal that changes with time, and the constant input CH3 is used to generate an initial phase signal. The initial phase defaults to 0, and the initial phases of the other two phases can be generated by ±120. The three signals are input into the adjustable sine wave generator SSG1 to generate the required disturbance sine signal output.

[0069] The continuous disturbance injection device in this embodiment allows the user to flexibly set disturbance parameters, which is suitable for different application scenarios. It can generate a continuous and stable sine wave signal and ensure a smooth transition of the frequency change process to avoid errors caused by frequency mutations. Each functional module is designed independently, making the device easy to upgrade, maintain and troubleshoot.

[0070] S2. Use a continuous disturbance injection device to continuously inject a disturbance signal. The method for generating the disturbance signal is as follows:

[0071] S21. Read the set disturbance parameters, and then generate the first signal according to the set starting frequency and disturbance amplitude. The signal is generated in the time domain as a sine wave. The duration of the generated signal is the set injection duration of each frequency segment, ensuring that the signal is output at this frequency for a sufficiently long time.

[0072] S22, after the first segment signal is generated, gradually increasing the frequency of the signal according to the frequency injection interval to generate signals of multiple frequency segments;

[0073] S23. When the frequency increases to the set end frequency, the entire signal is composed of multiple sine waves of different frequencies. These signals are spliced ​​together in time sequence to finally obtain a disturbance signal whose frequency gradually increases with time.

[0074] According to the disturbance signal generation method, a disturbance voltage signal and a disturbance current signal are respectively injected into the grid connection point of the system to be tested, and the response voltage and response current of the grid connection point of the system to be tested are respectively obtained through a voltage measuring device and a current measuring device (such as a voltmeter and an ammeter).

[0075] S3, based on the spectrum recognition algorithm of KAN neural network (Kolmogorov-Arnold Networks), calculates the impedance amplitude and phase at the current disturbance frequency.

[0076] The KAN neural network is inspired by the Kolmogorov-Arnold representation theorem, which states that any continuous function can be represented by a set of single-variable functions, which provides theoretical support for the neural network to approximate arbitrarily complex functions. In a neural network, the activation function is a function used to perform a nonlinear transformation on the output of each neuron, which determines the output result after the input signal passes through the neuron. The activation function maps the input to an output range and introduces nonlinear capabilities, allowing the neural network to approximate complex functions.

[0077] In traditional neural networks, activation functions are usually fixed. Each neuron uses a predefined activation function that has no adjustable parameters and performs nonlinear transformations based only on the input value. The activation function of the KAN neural network is not a predefined fixed function, but a learnable function. This means that the activation function of the KAN neural network is not just a single mathematical formula, but a function that is dynamically adjusted based on the characteristics of the data.

[0078] In order to fit the voltage and current signals containing multiple frequency components, the sine activation function can more effectively process signals with strong periodicity and oscillation. It is more suitable for processing periodic features than traditional activation functions (such as ReLU, Sigmoid, Tanh), especially when processing multi-frequency sinusoidal signals, harmonic signals or noisy sine waves, the sinusoidal activation function behaves more naturally and effectively. It is defined as a combination of multiple sinusoidal functions, each with different frequency, phase and amplitude. Its expression is:

[0079]

[0080] In order to further meet the needs of fitting various complex signals in the power system, a combination of multiple sinusoidal functions and linear functions is used as the activation function in this embodiment:

[0081]

[0082] in, is the magnitude of each sine term, is the angular frequency of each sinusoidal term, is the phase of each sinusoidal term, is the number of sine terms, indicating the number of fitted sine components. , the angular frequency of the sine term , the phase of the sine term are all learnable parameters, Represents the linear term in the signal and can be used to fit the low-frequency noise or trend term in the signal.

[0083] This combined activation function can process complex signals, especially those containing periodicity, nonlinearity, sparsity and attenuation characteristics, and improves the performance of the KAN neural network in fitting current signals and voltage signals.

[0084] In this application, the KAN neural network is used to fit the amplitude, frequency and phase of the response voltage and response current, including the following steps:

[0085] S31. Input signal sampling modeling. The voltage signal and current signal of the system after the disturbance are collected through the measuring device. The sampling frequency of the signal should be high enough, three to five times the highest frequency component of the signal collected by the measuring device, to ensure effective sampling of the high-frequency components and noise in the signal.

[0086] Assume that the response voltage or response current obtained by S2 is , through the sampling period The continuous signal Discretize into ,in is a discrete time point.

[0087] Signal Depend on It is composed of a sinusoidal signal and noise. The expression is:

[0088]

[0089] in, is the sampling period, For the The amplitude of the sinusoidal component, For the The initial phase of the sinusoidal components, For the The attenuation coefficient of the sinusoidal component is For the The frequency of the sinusoidal component, is white noise with zero mean.

[0090] S32, data input, the obtained noisy signal A window length is intercepted as the input of the KAN neural network, according to the highest frequency of the signal Select the window length. The calculation formula is: , Take a value between 3 and 5. After preprocessing operations including trend removal, denoising and normalization, the signal is input to the input layer of the KAN neural network. Denoising uses filters or wavelet denoising methods, and normalization ensures that the signal amplitude is within a reasonable range, such as [0, 1] or [-1, 1].

[0091] S33, signal feature decomposition, according to the Kolmogorov-Arnold representation theorem, the KAN neural network can decompose complex time series signals. The initial layer of the KAN neural network maps the signal obtained after S32 preprocessing into different feature functions through multiple intermediate neurons. The feature functions include sine wave components. As well as the linear noise part superimposed on it, the feature decomposition layer is expanded based on the sine wave basis function to decompose the noisy signal into the superposition of multiple sine waves and noise.

[0092] S34, nonlinear approximation, in the middle layer of the KAN neural network, nonlinear activation functions are used to perform nonlinear approximation on the feature function obtained in S33. Each activation function corresponds to a nonlinear transformation of an input feature. Through multi-layer mapping, the weight of each activation function is gradually adjusted to make the output signal close to the original signal. The KAN network can flexibly approximate any continuous function, so it can capture the components of the signal. The amplitude, frequency and phase characteristics of each sine wave.

[0093] S35, noise filtering and output fitting, after multiple iterations of training, the output of the KAN neural network gradually approaches the original signal , that is, extract the characteristics of each sinusoidal signal and model the noise part as a random component. Since the KAN neural network has good nonlinear fitting ability, the final KAN neural network output can restore the signal well. The characteristics of each sine wave in , ignoring or minimizing noise interference.

[0094] S36, training optimization, during the fitting process, use the loss function to measure the difference between the predicted result and the original signal If the deviation is greater than the set tolerance threshold, the back propagation algorithm is used to optimize the learnable parameters of the KAN neural network: the amplitude of the sine term , the angular frequency of the sine term , the phase of the sine term , making the fitting effect more accurate. Through the back propagation algorithm, the network parameters are continuously optimized, the feature extraction process of the signal is continuously optimized, the cumulative error is reduced, and the measurement results are closer to the actual impedance value through multiple iterations. Through continuous optimization, the KAN neural network will effectively approximate the original sinusoidal signal , and reduce the impact of noise on the output.

[0095]

[0096] in, is the output of the KAN neural network, is the number of sampling points.

[0097] S4. Match the frequency injection interval of the continuous disturbance injection device and measure the voltage data and current data after the disturbance is injected, and calculate the impedance of the next frequency point until a complete scan is completed. Record the number of scans as ,like , then obtain the next segment of data for calculation until the total number of disturbance injections is reached , complete the measurement of the impedance of the system to be tested.

[0098] After the KAN neural network performs nonlinear approximation and denoising on the input voltage and current signals, the amplitude and phase characteristics of each frequency component are obtained. Specifically, the KAN neural network decomposes the original signal into multiple sine waves and noise components.

[0099] For voltage signals, it can be expressed as the superposition of multiple sinusoidal wave components:

[0100]

[0101] in, For frequency The voltage amplitude at For frequency The phase angle of the voltage at is the noise component of the voltage signal.

[0102] For current signals, it can also be expressed as the superposition of multiple sinusoidal wave components:

[0103]

[0104] in, For frequency The current amplitude at For frequency The phase angle of the current at is the noise component of the current signal.

[0105] After being processed by the KAN neural network, each frequency can be accurately extracted The corresponding voltage amplitude and current amplitude , and minimize the noise impact, providing reliable input data for the subsequent impedance calculation.

[0106] Based on the extracted frequency components, the voltage at each frequency point can be and current To calculate the impedance . Using Ohm's law, impedance is calculated as:

[0107]

[0108] in, For frequency The complex impedance at contains amplitude and phase information. The amplitude reflects the total impedance at that frequency, the real part (resistance) represents the energy consumption, and the imaginary part (reactance) reflects the energy storage characteristics. This information is critical for analyzing the electrical characteristics of the system.

[0109] S5. To facilitate subsequent analysis, the impedance calculation results at different frequencies are displayed visually. The display methods used include:

[0110] Impedance amplitude curve: shows the change trend of impedance amplitude with frequency, which is used to visually observe the impedance change of the system at various frequencies. If the impedance amplitude is found to change sharply at certain frequency points, it can be inferred that the system may resonate at this frequency and further suppression measures are needed.

[0111] Impedance phase curve: shows the change of impedance phase with frequency, and observes the energy transmission and storage characteristics of the system. A phase close to 0° usually reflects the resistance characteristics of the system, while a phase close to ±90° in the high frequency band reflects the enhancement of inductance or capacitance effects.

[0112] Nyquist diagram: The real and imaginary parts of the complex impedance are plotted on the complex plane to form a Nyquist diagram, which is used to analyze the stability and resonance risk of the system. The position close to the origin in the diagram represents the frequency point with smaller impedance, and the farther away from the origin, the greater the impedance. For the resonant frequency point, sharp extreme points may appear on the Nyquist diagram, reflecting that the system may have abnormal operation at this frequency.

[0113] In summary, this application can support impedance measurement under complex power system conditions, including large-scale distributed generation systems, flexible AC transmission systems (FACTS), high-frequency harmonics and other complex signal environments. Compared with existing impedance measurement technologies, the method proposed in this application solves the limitations of FFT and other traditional spectrum analysis tools in power systems through more flexible modeling capabilities, providing new solutions for more accurate power quality assessment, harmonic analysis and dynamic system optimization.

[0114] The application can also be applied in the following scenarios:

[0115] Wind farm grid-connected impedance measurement:

[0116] Wind power generation has frequent output power fluctuations due to wind speed changes, which can easily cause grid harmonics and voltage instability. The KAN neural network method of the present invention can capture the nonlinear signals and high-frequency disturbances generated by wind farms during the grid connection process in real time, thereby accurately calculating the impedance of the grid connection point and preventing system resonance or instability.

[0117] Harmonic analysis of photovoltaic power generation system: Photovoltaic power generation is also affected by lighting conditions, and its output current shows strong nonlinear and unstable characteristics. Traditional impedance measurement methods often make errors when processing such signals. The KAN neural network in this application can adaptively process complex signals in photovoltaic power generation systems, quickly identify different frequency components, improve the accuracy of impedance measurement, and help optimize harmonic management and power quality control of power grids.

[0118] Grid fault monitoring under the connection of new energy to the grid: When the new energy power generation system is connected to the grid, instantaneous faults or abnormal fluctuations may occur in the grid. Through the time-frequency analysis capability of the KAN neural network, these fault signals can be monitored and identified in real time, the characteristic components of current and voltage can be quickly extracted, and the impedance can be accurately calculated, thereby providing an important basis for fault location and fault protection.

[0119] Dynamic impedance monitoring in distributed power generation systems: With the widespread application of distributed power generation systems, the distributed characteristics of the power grid make the impedance characteristics of each node more complex and changeable. The KAN neural network shows excellent adaptability when processing multi-node, nonlinear dynamic signals, and can effectively monitor the impedance changes of distributed power grids to ensure the stable operation of the power system.

[0120] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An impedance measurement method based on a KAN neural network and a continuous disturbance injection device, characterized in that: The following steps are involved: S1. Setting disturbance parameters in a continuous disturbance injection device, determining the total number of disturbance injections, and connecting the continuous disturbance injection device to the grid connection point of the system to be tested; S2. Use a continuous disturbance injection device to continuously inject a disturbance signal with set disturbance parameters, and obtain a response voltage and a response current of a grid connection point of the system to be tested; S3, the spectrum recognition algorithm based on KAN neural network calculates the impedance amplitude and phase at the current disturbance frequency; S4. According to the frequency injection interval of the continuous disturbance injection device, the impedance calculation of the next frequency point is performed until the total number of disturbance injections is reached, and the impedance measurement of the system to be measured is completed; S5. The impedance calculation results at different frequencies are displayed through visualization.

2. The impedance measurement method based on the KAN neural network and the continuous disturbance injection device according to claim 1 is characterized in that: The disturbance parameters include disturbance amplitude, starting frequency, ending frequency, frequency injection interval and each frequency injection duration.

3. The impedance measurement method based on the KAN neural network and the continuous disturbance injection device according to claim 2 is characterized in that: The continuous disturbance injection device comprises: Reading module: used to read the set disturbance parameters, which are input from the external controller through the input interface of the reading module; Amplitude and frequency control module: used to determine the signal amplitude and control the signal frequency change. After the initial signal is generated, starting from the starting frequency, multiple signals with different frequencies are generated according to the set frequency injection interval and the injection duration of each frequency segment. Signal generation module: generates a disturbance signal in the time domain according to the signal provided by the amplitude and frequency control module in accordance with the sine wave law. The frequency of the generated disturbance signal is the frequency of the signal generated by the amplitude and frequency control module, the amplitude is the set disturbance amplitude, and the duration of each signal segment is the set injection duration; Output module: Injects the disturbance signal generated by the signal generation module into the grid connection point of the system to be tested, ensures the smooth output of the entire disturbance signal, filters out the mutations or noise that may be generated during frequency conversion, and ensures the continuity of the disturbance signal.

4. The impedance measurement method based on the KAN neural network and the continuous disturbance injection device according to claim 3 is characterized in that: The generation of the disturbance signal in S2 comprises the following steps: S21, read the set disturbance parameters, and then generate the first signal according to the set starting frequency and disturbance amplitude. The signal is generated in the time domain according to the sine wave, and the duration of the generated signal is the set injection duration of each frequency segment; S22, after the first segment signal is generated, gradually increasing the frequency of the signal according to the frequency injection interval to generate signals of multiple frequency segments; S23. When the frequency increases to the set end frequency, the entire signal is composed of multiple sine waves of different frequencies. These signals are spliced ​​together in time sequence to finally obtain a disturbance signal whose frequency gradually increases with time.

5. The impedance measurement method based on the KAN neural network and the continuous disturbance injection device according to claim 4 is characterized in that: The KAN neural network is used to fit the amplitude, frequency and phase of the response voltage and response current, including the following steps: S31, input signal sampling modeling, assuming that the response voltage or response current obtained by S2 is , through the sampling period The continuous signal Discretize into ,in is a discrete time point; S32, data input, the signal A window length is intercepted as the input of the KAN neural network, and the signal After preprocessing including trend removal, denoising and normalization, the data is input into the input layer of the KAN neural network; S33, signal feature decomposition, using the initial layer of the KAN neural network to map the signal obtained after S32 preprocessing into different feature functions through multiple intermediate neurons. The feature function includes a sine wave component and a linear noise part superimposed thereon. The noisy signal is decomposed into a superposition of multiple sine waves and noise through the feature decomposition layer; S34, nonlinear approximation, in the middle layer of the KAN neural network, the signal feature function obtained in S33 is nonlinearly approximated using a nonlinear activation function. Each activation function corresponds to a nonlinear transformation of an input feature. Through multi-layer mapping, the weight of each activation function is gradually adjusted to make the output signal close to the original signal. , capturing the constituent signal The amplitude, frequency and phase characteristics of each sine wave; S35, noise filtering and output fitting, after multiple iterations, the output of the KAN neural network gradually approaches the original signal , that is, extracting the features of each sinusoidal signal and modeling the noise part as a random component. The final KAN neural network output can restore the signal The characteristics of each sine wave in; S36, training optimization, in the fitting process, the loss function is used to measure the deviation between the prediction result and the original signal. If the deviation value is greater than the set tolerance threshold, the back propagation algorithm is used to optimize the learnable parameters in the KAN neural network to approximate the original sinusoidal signal. , reducing the impact of noise on the output.

6. The impedance measurement method based on the KAN neural network and the continuous disturbance injection device according to claim 5 is characterized in that: The activation function of the KAN neural network is a combination of multiple sinusoidal functions and linear functions. The activation function is as follows: in, is the magnitude of each sine term, is the angular frequency of each sinusoidal term, is the phase of each sinusoidal term, is the number of sine terms, indicating the number of fitted sine components and the amplitude of the sine terms , the angular frequency of the sine term , the phase of the sine term are all learnable parameters, Represents the linear terms in the signal.

7. The impedance measurement method based on the KAN neural network and the continuous disturbance injection device according to claim 6 is characterized in that: The signal obtained in S31 Depend on It is composed of a sinusoidal signal and noise. The expression is: in, is the sampling period, For the The amplitude of the sinusoidal component, For the The initial phase of the sinusoidal components, For the The attenuation coefficient of the sinusoidal component is For the The frequency of the sinusoidal component, is white noise with zero mean.

8. The impedance measurement method based on the KAN neural network and the continuous disturbance injection device according to claim 7 is characterized in that: The loss function in S36 is: in, is the output of the KAN neural network, is the number of sampling points.

9. The impedance measurement method based on the KAN neural network and the continuous disturbance injection device according to claim 8 is characterized in that: The impedance calculation formula of the system to be measured is: in, For frequency The complex impedance at , containing both magnitude and phase information, For frequency The voltage at For frequency The current at For frequency The voltage amplitude at For frequency The phase angle of the voltage at For frequency The current amplitude at For frequency The phase angle of the current.

10. The impedance measurement method based on the KAN neural network and the continuous disturbance injection device according to claim 9 is characterized in that: The display method in S5 includes an impedance amplitude curve, an impedance phase curve and a Nyquist diagram.

Images