New energy power generation safety monitoring system based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a new energy power generation safety monitoring system based on artificial intelligence, and the system comprises a power spectrum estimation module which determines a near fundamental wave inter-harmonic band; the equivalent bandwidth calculation module is used for calculating to obtain an energy equivalent bandwidth change rate; the combined vector construction module is used for obtaining a combined vector; the risk index prediction module outputs a risk index through a risk prediction model; the self-adaptive threshold generation module is used for calculating to obtain a self-adaptive threshold; and the power generation risk judgment module judges whether a power generation risk exists or not according to the self-adaptive threshold value. According to the method, an electric ring coupling characteristic taking an ice-induced inter-harmonic spectrum dynamic broadening coefficient as a core is constructed, so that mechanical non-stationary fluctuation caused by fan blade icing and network side near fundamental wave inter-harmonic evolution are associated; meanwhile, the incidence relation of inter-harmonic broadening acceleration and risk increasing is embedded into a model structure, and new energy grid connection safety of a weak power grid and a capacitance compensation power distribution system is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an artificial intelligence-based safety monitoring system for new energy power generation. Background Technology

[0002] As the proportion of new energy grid connection continues to increase, a large number of wind farms with power electronic interfaces are connected to weak power grids or configured with capacitor compensation distribution systems. Coastal cold regions and high-altitude areas often experience humid and cold environments with supercooled water droplets, making wind turbine blades prone to icing. The units need to operate under frequent changing operating conditions. At the same time, grid connection standards require monitoring power quality indicators such as interharmonics and flicker. Furthermore, due to long cables and reactive power compensation devices, the distribution system forms a frequency band near the low frequency range that is more sensitive to external disturbances. Operators need to pay attention to the mechanical condition of the wind turbines, meteorological conditions, and grid-side power quality and stability boundaries.

[0003] Existing monitoring technologies largely follow two paths: on the electrical side, the total harmonic distortion rate (THD) or the amplitude of fixed-order harmonics (CHDs) are the core indicators; on the environmental side, temperature, humidity, wind speed, or icing are directly observed. Both methods are based on static thresholds or fixed-frequency statistics. While these methods can issue alerts when disturbances are significant, they are not sensitive to non-stationary changes during icing, and are particularly difficult to identify the rapid process of non-integer multiples of the fundamental frequency widening and increasing in intensity. Furthermore, relying solely on environmental alarms cannot determine whether the grid connection point is about to be excited, and simply observing the peak electrical energy cannot capture the key phenomenon of frequency band expansion over time, making it difficult to correlate environmental and electrical information for risk warning.

[0004] Blade icing alters aerodynamic shape and surface roughness, leading to abrupt changes in lift-drag characteristics and asymmetric forces between blades. This, in turn, causes strong non-stationary fluctuations in mechanical torque and speed disturbances. These fluctuations are transmitted to the grid side via the control loops of generators and converters, forming non-integer multiples of the fundamental frequency. Icing is intermittent in its melting, shedding, and re-condensation, and its intensity and dominant frequency drift over time, causing the grid-side spectrum to change from a stable single peak to a wideband energy distribution that expands over a short period. When this wideband distribution expands into and overlaps with the low-frequency sensitive bands of the power distribution system, it is easily excited into voltage flicker or continuous oscillation. This process of changing from a narrow spectrum to a wide spectrum, from frequency stability to drift, is the core reason why existing methods relying on fixed frequency points or static thresholds cannot capture it in time, making it difficult to provide early warning of grid-side risks caused by icing. Summary of the Invention

[0005] This invention provides a new energy power generation safety monitoring system based on artificial intelligence, which solves the technical problems mentioned in the background.

[0006] This invention provides an artificial intelligence-based safety monitoring system for new energy power generation, comprising: The power spectrum estimation module collects three-phase current at the grid connection point to construct a scalar current trajectory and simultaneously collects environmental parameters. It performs short-time power spectrum estimation on the scalar current trajectory to obtain the power spectral density. It determines the near-fundamental interharmonic band within a predetermined narrow frequency band on both sides of the fundamental frequency of the power grid and removes sub-frequency bands containing the fundamental frequency. Environmental parameters include: wet-bulb temperature, relative humidity, and wind speed; The equivalent bandwidth calculation module calculates the total energy and energy weighted center frequency within the near-fundamental interharmonic band, thereby determining the minimum symmetrical frequency band and energy equivalent bandwidth, and calculating the rate of change of energy equivalent bandwidth. The combined vector construction module takes the environmental parameters as input to a continuous monotonic learnable mapping, outputs the icing potential, multiplies it by the rate of change of energy equivalent bandwidth to obtain the dynamic broadening coefficient of the ice-induced interharmonic spectrum, and performs online exponential sliding mean and standard deviation estimation and time difference estimation on the coefficient to obtain the combined vector. The risk index prediction module takes the combined vector as input, constructs a risk prediction model containing non-negative causal convolution, a stable discrete state space kernel, and a monotonic output head, and outputs the risk index. The adaptive threshold generation module performs online quantile tracking of the risk index to obtain the upper quantile baseline, and calculates the adaptive threshold by combining the freezing potential and safety margin coefficient. The power generation risk assessment module determines whether the risk index is greater than or equal to an adaptive threshold. If so, it indicates that there is a power generation risk in the new energy power generation; otherwise, it indicates that there is no power generation risk.

[0007] Furthermore, the current samples of each phase corresponding to the same time stamp in the three-phase current discrete sequence are squared respectively. The results of the three square operations are summed, and the sum is divided by 3. Finally, the square root of the quotient is taken to obtain the discrete form of the scalar current trajectory. The frequency resolution is determined according to the sampling frequency and the length of the window function. Then, the frequency grid is determined according to the frequency resolution. Each frequency value in the frequency grid is an integer multiple of the frequency resolution. The discrete form of the scalar current trajectory is multiplied by the window function. Each multiplication result is multiplied by the corresponding complex exponential term. Finally, all multiplication results are summed to obtain the short-time Fourier transform result. The modulus of the short-time Fourier transform result is taken first. Then, the modulus is squared. The squared result is divided by the product of the window function length and the sampling frequency to obtain the power spectral density.

[0008] Furthermore, the fundamental frequency of the power grid is set to 50 Hz, and a symmetrical bandwidth parameter is set with the symmetrical bandwidth parameter being greater than 0. Taking the fundamental frequency of the power grid as the center, the frequency value obtained by subtracting the symmetrical bandwidth parameter from the fundamental frequency of the power grid is used as the lower limit, and the frequency value obtained by adding the symmetrical bandwidth parameter to the fundamental frequency of the power grid is used as the upper limit, thus forming a predetermined narrow frequency band on both sides of the fundamental frequency. The sub-frequency band range containing the fundamental frequency in the predetermined narrow frequency band is removed, and the remaining frequency range forms the near-fundamental interharmonic band, wherein the sub-frequency band range containing the fundamental frequency is a preset parameter.

[0009] Furthermore, the power spectral density corresponding to each frequency grid within the near-fundamental interharmonic band is multiplied by the frequency resolution to obtain a first product result. Then, all the first product results are added together to obtain the total energy within the band. The frequency value of each frequency grid within the near-fundamental interharmonic band is multiplied by the power spectral density corresponding to that frequency grid to obtain a second product result. All the second product results are added together and divided by the total energy within the band to obtain the energy-weighted center frequency.

[0010] Furthermore, using the energy-weighted center frequency as the midpoint and half the radius of the symmetrical bandwidth variable as the interval radius, a symmetrical frequency interval is obtained. The power spectral density corresponding to each frequency grid within the intersection is multiplied by the frequency resolution to obtain the product result within the intersection. All product results within the intersection are added together to obtain the energy accumulation. If the energy accumulation is less than 95% of the total energy within the band, the symmetrical bandwidth variable is adjusted and increased until the energy accumulation is not less than 95% of the total energy within the band. The symmetrical bandwidth variable at this point is taken as the energy equivalent bandwidth. The intersection of the symmetrical frequency interval corresponding to this energy equivalent bandwidth and the near-fundamental interharmonic band is taken as the minimum symmetrical frequency band. Two adjacent time window centers are selected, and the energy equivalent bandwidth corresponding to each of the two time window centers and the time interval between the two time window centers are obtained respectively. The energy equivalent bandwidth corresponding to the energy equivalent bandwidth of the latter time window center is subtracted from the energy equivalent bandwidth corresponding to the former time window center, and the difference is divided by the time interval between the two time window centers to obtain the energy equivalent bandwidth change rate.

[0011] Further, the dynamic broadening coefficient of the ice-induced interharmonic spectrum is calculated, including the following steps: Step S201: Subtract the wet-bulb temperature from the preset temperature reference value to obtain the wet-bulb temperature deviation value, multiply the wet-bulb temperature deviation value with the first scale parameter, input the monotonic bounded function, and obtain the wet-bulb temperature mapping result. Step S202: Subtract the preset humidity reference value from the relative humidity to obtain the relative humidity deviation value, multiply the relative humidity deviation value with the second scale parameter, input the monotonic bounded function, and obtain the relative humidity mapping result. Step S203: Subtract the preset wind speed reference value from the wind speed to obtain the wind speed deviation value, multiply the wind speed deviation value with the third scale parameter, input the monotonic bounded function, and obtain the wind speed mapping result. Step S204: Multiply the wet-bulb temperature mapping result, the relative humidity mapping result, and the wind speed mapping result to obtain the icing potential; Step S205: Obtain the energy equivalent bandwidth change rate, which is the time change rate of the energy equivalent bandwidth within the near-fundamental interharmonic band. Step S206: Multiply the rate of change of energy equivalent bandwidth by the icing potential to obtain the dynamic broadening coefficient of ice-induced interharmonic spectrum.

[0012] Further, the combined vector is calculated, including the following steps: Step S301: Subtract the forgetting factor from 1, multiply by the exponential moving average of the previous time step, and add the product of the ice-induced interharmonic spectrum broadening coefficient and the forgetting factor at the current time step to obtain the exponential moving average at the current time step. Step S302: Subtract the forgetting factor from 1, multiply by the exponential moving variance of the previous time step, add the square of the difference between the current ice-induced interharmonic spectrum broadening coefficient and the exponential moving mean of the previous time step, and then multiply by the forgetting factor to obtain the exponential moving variance of the current time step. Step S303: Take the square root of the exponential moving variance at the current time to obtain the exponential moving standard deviation at the current time. Subtract the exponential moving mean at the current time from the ice-induced interharmonic spectrum broadening coefficient at the current time to obtain the difference in coefficient mean. Step S304: Add a preset constant to the exponential moving standard deviation at the current time to obtain the standard deviation correction value, and then divide the difference in the mean of the coefficients by the standard deviation correction value to obtain the standardized value. Step S305: Select the center time of two adjacent time windows, calculate the time interval between the center times of the two time windows, obtain the standardized values ​​corresponding to the center times of the two time windows respectively, and subtract the standardized value of the previous time from the standardized value of the current time to obtain the difference of standardized values. Step S306: Divide the difference of standardized values ​​by the time interval to obtain the first-order rate of change. Concatenate the standardized values ​​and the first-order rate of change to obtain the combined vector.

[0013] Furthermore, multiple non-negative causal convolutional branches are constructed, the kernel length of each branch is set, and multiple lag weights are configured for each branch. For each branch, the input combination vector and its historical combination vector are multiplied by the corresponding lag weights respectively, and all multiplication results are summed to obtain the convolutional output of each branch. Configure a trainable matrix for each non-negative causal convolution branch, multiply the convolution output of each branch with the corresponding trainable matrix, and sum all the multiplication results to obtain the driving force. Set the hidden state vector of the stable discrete state space kernel, and configure four trainable matrices, namely the first trainable matrix, the second trainable matrix, the third trainable matrix, and the fourth trainable matrix; apply hyperbolic tangent operation to each element of the first trainable matrix; Multiply the first trainable matrix (after hyperbolic tangent operation) with the hidden state vector at the current time step, multiply the second trainable matrix with the driving force, and add the two multiplication results to obtain the hidden state vector at the next time step; multiply the third trainable matrix with the hidden state vector at the current time step, and multiply the fourth trainable matrix with the driving force, and add the two multiplication results to obtain the intermediate risk quantity.

[0014] Furthermore, two trainable parameters are configured for the monotonic output head: a gain parameter and an offset parameter. A softplus function is applied to the gain parameter to obtain a non-negative gain. The intermediate risk is multiplied by the non-negative gain, and the result is added to the offset parameter to obtain the intermediate value before output. A sigmoid function is applied to the intermediate value before output to obtain the risk index. The loss function of the risk prediction model is set as the cross-entropy loss function. The labels of the historical training data are obtained offline. These labels are indicators of whether grid connection oscillation or flicker of new energy power generation has occurred within a preset time after the calculation of the historical risk index. The value of the label is 0 or 1.

[0015] Further, subtract the freezing potential from 1 to obtain the weight; determine if the risk index at the current moment is greater than the upper quantile baseline at the previous moment, then the indicator function result is 1, otherwise the indicator function result is 0; first subtract the quantile position from 1, then subtract the difference from the indicator function result, multiply by the step size, and then multiply by the weight to obtain the upper quantile baseline update; add the upper quantile baseline update to the upper quantile baseline at the previous moment to obtain the upper quantile baseline at the current moment; The deviation is obtained by subtracting the upper quantile baseline of the previous time from the current risk index and taking the absolute value of the difference. The mean absolute deviation index moving estimate of the current time is obtained by subtracting the forgetting factor from 1 and multiplying it by the mean absolute deviation index moving estimate of the previous time, adding the forgetting factor multiplied by the weight and then multiplying by the deviation. The adaptive threshold is obtained by adding the safety margin coefficient to the upper quantile baseline of the current time and multiplying it by the mean absolute deviation index moving estimate of the current time.

[0016] The beneficial effects of this invention are as follows: By constructing an electrical loop coupling feature with the ice-induced interharmonic spectrum broadening coefficient as the core, this invention correlates the mechanical non-stationary fluctuations caused by wind turbine blade icing with the evolution of near-fundamental interharmonics on the grid side; simultaneously, it embeds the correlation between the accelerated broadening of interharmonics and the increasing risk into the model structure, avoiding the counterintuitive output of unconstrained black-box models and improving the reliability of risk prediction; in addition, the adaptive threshold is dynamically adjusted through quantile tracking and inverse weighting of icing potential to adapt to seasonal operating condition drift and solve the problem of false alarms and missed alarms with fixed thresholds; ultimately, it can detect the risk of near-fundamental interharmonics expanding into the low-frequency sensitive frequency band of the distribution system in advance, accurately warn of voltage flicker or continuous oscillations caused by icing, and ensure the grid connection safety of new energy sources in weak power grids and capacitor-compensated distribution systems. Attached Figure Description

[0017] Fig. 1 This is a schematic diagram of a new energy power generation safety monitoring system based on artificial intelligence according to the present invention; Fig. 2 This is a flowchart of the calculation of the dynamic broadening coefficient of the ice-induced interharmonic spectrum according to the present invention; Fig. 3 This is a flowchart of the calculation of the combined vector according to the present invention.

[0018] In the figure: Power spectrum estimation module 101, equivalent bandwidth calculation module 102, combined vector construction module 103, risk index prediction module 104, adaptive threshold generation module 105, power generation risk judgment module 106. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0020] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] like Figs. 1-3 As shown, a new energy power generation safety monitoring system based on artificial intelligence includes: The power spectrum estimation module 101 collects three-phase current at the grid connection point to construct a scalar current trajectory and simultaneously collects environmental parameters. It performs short-time power spectrum estimation on the scalar current trajectory to obtain the power spectral density. It determines the near-fundamental interharmonic band within a predetermined narrow frequency band on both sides of the fundamental frequency of the power grid and removes sub-frequency bands containing the fundamental frequency. Environmental parameters include: wet-bulb temperature, relative humidity, and wind speed; The equivalent bandwidth calculation module 102 calculates the total energy and energy weighted center frequency in the near-fundamental interharmonic band, thereby determining the minimum symmetrical frequency band and energy equivalent bandwidth, and calculating the rate of change of energy equivalent bandwidth. The combined vector construction module 103 inputs environmental parameters into a continuous monotonic learnable mapping, outputs the icing potential, multiplies it by the rate of change of energy equivalent bandwidth to obtain the dynamic broadening coefficient of the ice-induced interharmonic spectrum, and performs online exponential sliding mean and standard deviation estimation and time difference on the coefficient to obtain the combined vector. The risk index prediction module 104 takes the combined vector as input, constructs a risk prediction model containing non-negative causal convolution, a stable discrete state space kernel and a monotonic output head, and outputs the risk index. The adaptive threshold generation module 105 performs online quantile tracking of the risk index to obtain the upper quantile baseline, and calculates the adaptive threshold by combining the freezing potential and safety margin coefficient. The power generation risk assessment module 106 determines whether the risk index is greater than or equal to the adaptive threshold. If so, it indicates that there is a power generation risk in the new energy power generation; otherwise, it indicates that there is no power generation risk.

[0022] In one embodiment of the present invention, the phase current samples corresponding to the same time stamp in the three-phase current discrete sequence are squared respectively, the three squared results are summed, the sum is divided by 3, and finally the square root of the quotient is taken to obtain the discrete form of the scalar current trajectory; the frequency resolution is determined according to the sampling frequency and the length of the window function, and then the frequency grid is determined according to the frequency resolution, where each frequency value in the frequency grid is an integer multiple of the frequency resolution; the discrete form of the scalar current trajectory is multiplied by the window function, and each multiplication result is multiplied by the corresponding complex exponential term, and finally all multiplication results are summed to obtain the short-time Fourier transform result; the modulus of the short-time Fourier transform result is taken first, and then the modulus is squared, and the squared result is divided by the product of the window function length and the sampling frequency to obtain the power spectral density.

[0023] It should be noted that the three-phase current discrete sequence represents the set of current values ​​collected at the grid connection point, reflecting the changes of phase A, phase B, and phase C currents with discrete time points. Each discrete time point corresponds to a set of current values ​​for phase A, phase B, and phase C. The discrete form of the scalar current trajectory represents a single current value obtained by squaring, summing, dividing by three, and taking the square root of the current samples of each of the three phases at the same time stamp. This value does not contain phase information and only reflects the amplitude characteristics of the current. The sampling frequency represents the number of times the three-phase current is sampled per unit time, in Hertz, and determines the time interval for discretizing the current signal. Window functions are used to truncate and weight the discrete form of the scalar current trajectory, reducing spectral leakage caused by signal truncation in the time domain. Commonly used window functions include the Hanning window and the Hamming window. The window function exists in the form of a set of values, and the number of values ​​in the set is equal to the length of the window function. Frequency resolution represents the minimum frequency interval between two adjacent frequency components that can be distinguished in short-time power spectrum estimation, in Hertz. The frequency grid represents the frequency range used for short-time power spectrum estimation. The spectrum estimation is a set of discrete frequency points, where each frequency point's value is an integer multiple of the frequency resolution, covering the frequency range to be analyzed. The complex exponential term represents the exponential function in complex form used to convert the time-domain signal to the frequency domain. Its value is determined by a specific frequency value in the frequency grid, the discrete-time index, the natural constant, and pi. The value of the complex exponential term is equal to the natural constant raised to the power of (negative imaginary unit multiplied by two times pi, then multiplied by a specific frequency value in the frequency grid, then multiplied by the value of the discrete-time index). The discrete-time index refers to the position number of a sample point in the discrete form of the scalar current trajectory in the time series, incrementing sequentially from 0 at the initial sample point. The short-time Fourier transform result represents the complex value obtained after weighting the discrete form of the scalar current trajectory through a window function and transforming it with the complex exponential term. It is used to reflect the frequency domain characteristics of the signal under a specific time window and a specific frequency point. The power spectral density represents the power per unit frequency range, measured in watts per hertz, and is used to reflect the power distribution of the signal at different frequency points.

[0024] It should be noted that the time window uses a sliding method to traverse the entire discrete form of the scalar current trajectory. The sliding step size refers to the interval of the number of samples between the starting sample points of two adjacent time windows. The step size can be set according to the time resolution requirements, usually choosing 1 / 2 or 1 / 4 of the window function length to balance time resolution and computational efficiency. The frequency range of the frequency grid needs to cover the frequency range of the signal to be analyzed. For the grid-connected scenario of new energy power generation, the lowest frequency is usually set to 0 Hz, and the highest frequency is set to 1 / 2 of the sampling frequency (i.e., the Nyquist frequency) to avoid frequency aliasing. For example, if the sampling frequency is 1000 Hz and the Nyquist frequency is 500 Hz, then the frequency of the frequency grid is... The range is set to 0 Hz to 500 Hz, and frequency points are generated sequentially according to frequency resolution. The determination of the time window needs to be combined with the analysis time point and the window function length. First, the center sample point of each time window is determined. The center sample point starts from the initial sample point in the discrete form of the scalar current trajectory and moves sequentially according to the sliding step size. Based on the center sample point, two sample points are taken forward and backward (window function length minus one) to form the sample points within the time window. If the window function length is odd, the number of sample points taken forward and backward is equal. If the window function length is even, one more sample point can be taken forward or backward to ensure that the total number of sample points is equal to the window function length.

[0025] In one embodiment of the present invention, the fundamental frequency of the power grid is set to 50 Hz, and a symmetrical bandwidth parameter is set such that the symmetrical bandwidth parameter is greater than 0. Taking the fundamental frequency of the power grid as the center, the frequency value obtained by subtracting the symmetrical bandwidth parameter from the fundamental frequency of the power grid is used as the lower limit, and the frequency value obtained by adding the symmetrical bandwidth parameter to the fundamental frequency of the power grid is used as the upper limit, thereby forming a predetermined narrow frequency band on both sides of the fundamental frequency. The sub-frequency band range containing the fundamental frequency in the predetermined narrow frequency band is removed, and the remaining frequency range constitutes the near-fundamental interharmonic band, wherein the sub-frequency band range containing the fundamental frequency is a preset parameter.

[0026] It should be noted that the fundamental frequency of the power grid represents the main power supply frequency in the public power grid, which is the basic frequency of voltage and current during normal operation of the power system. The fundamental frequency of the power grid is set to 50 Hz. The symmetric bandwidth parameter represents the width of the predetermined narrow band centered on the fundamental frequency of the power grid. The value of this parameter directly determines the size of the frequency range covered by the predetermined narrow band. The predetermined narrow band is the initial frequency range for subsequent screening of near-fundamental interharmonics. The sub-band range containing the fundamental frequency represents a small frequency range within the predetermined narrow band that contains the fundamental frequency of the power grid and the fundamental leakage frequency. This is the part that needs to be removed from the predetermined narrow band. This range is a pre-set parameter, the purpose of which is to avoid interference of the fundamental signal with interharmonic analysis. The near-fundamental interharmonic band represents the continuous frequency range remaining after removing the sub-band range containing the fundamental frequency from the predetermined narrow band. This is the dedicated frequency range for subsequent calculations of interharmonic energy, equivalent bandwidth, etc.

[0027] It should be noted that the symmetrical bandwidth parameter needs to be set in conjunction with two core factors: first, the frequency range requirements for near-fundamental interharmonic monitoring in the grid connection specifications for new energy power generation systems (for example, some specifications require monitoring interharmonics within 5 Hz on both sides of the fundamental frequency); and second, the frequency distribution pattern of interharmonics caused by actual icing faults on site (icing usually causes interharmonics to concentrate in the range of 3 to 8 Hz on both sides of the fundamental frequency). For example, if on-site data shows that interharmonics caused by icing are mostly distributed within 5 Hz on both sides of 50 Hz, the symmetrical bandwidth parameter can be set to 5 Hz to ensure that the predetermined narrow frequency band can completely cover potential interharmonics, while avoiding the inclusion of too many irrelevant frequency components that would increase the computational load; including the fundamental frequency... The sub-band range needs to be set around the fluctuation range of the fundamental frequency and the fundamental leakage range in the spectrum estimation. Usually, the center of this sub-band is the fundamental frequency of the power grid (50 Hz), and the bandwidth is controlled between 0.5 Hz and 1 Hz. In specific settings, first determine the allowable fluctuation range of the fundamental frequency (for example, the national standard stipulates that the fundamental frequency of the power grid can fluctuate by ±0.5 Hz), and then consider the fundamental leakage that may be generated by the short-time Fourier transform (about ±0.1 Hz). Therefore, the sub-band range containing the fundamental frequency can be set to 49.5 Hz to 50.5 Hz. This range can completely cover the fundamental frequency itself and the fundamental leakage frequency, effectively avoiding interference of the fundamental signal on the subsequent interharmonic characteristic calculation.

[0028] In one embodiment of the present invention, the power spectral density corresponding to each frequency grid in the near-fundamental interharmonic band is multiplied by the frequency resolution to obtain a first product result, and then all the first product results are added together to obtain the total energy in the band; the frequency value of each frequency grid in the near-fundamental interharmonic band is multiplied by the power spectral density corresponding to that frequency grid to obtain a second product result, and all the second product results are added together and divided by the total energy in the band to obtain the energy-weighted center frequency.

[0029] It should be noted that the first product result is used for subsequent cumulative calculation of the total energy within the band, reflecting the energy contribution of a single frequency grid; the second product result is used for subsequent cumulative calculation of the energy-weighted center frequency, reflecting the correlation contribution of frequency and energy of a single frequency grid; the total energy within the band is used to reflect the total energy contained within the near-fundamental interharmonic band; when the frequency value of the frequency grid is completely within the near-fundamental interharmonic band, the original value of the power spectral density corresponding to that frequency grid is directly taken; when the frequency value of the frequency grid partially overlaps with the near-fundamental interharmonic band (i.e., the frequency grid crosses the harmonic band boundary), linear interpolation is used to calculate the power spectral density corresponding to the part of the frequency grid within the harmonic band, that is, the overlap ratio between the frequency grid and the harmonic band boundary is first determined. Then, the overlap ratio is multiplied by the original power spectral density of the frequency grid to obtain the power spectral density value used for calculation, ensuring accurate energy calculation. When screening frequency grids within the near-fundamental interharmonic band, the center frequency value of the frequency grid is used as the criterion. That is, the center frequency value of each frequency grid (the average of the starting and ending frequency values ​​of the frequency grid) is first obtained, and then it is determined whether the center frequency value is within the frequency range of the near-fundamental interharmonic band (i.e., greater than or equal to the lower limit frequency value of the near-fundamental interharmonic band and less than or equal to the upper limit frequency value of the near-fundamental interharmonic band). If the judgment result is yes, the frequency grid belongs to the calculation object within the near-fundamental interharmonic band; if the judgment result is no, the frequency grid is excluded and does not participate in subsequent multiplication and accumulation operations.

[0030] In one embodiment of the present invention, a symmetrical frequency interval is obtained by taking the energy-weighted center frequency as the midpoint and half of the symmetrical bandwidth variable as the interval radius; the power spectral density corresponding to each frequency grid in the intersection is multiplied by the frequency resolution to obtain the product result in the intersection, and all product results in the intersection are added to obtain the energy accumulation; if the energy accumulation is less than 95% of the total energy in the band, the symmetrical bandwidth variable is adjusted and increased until the energy accumulation is not less than 95% of the total energy in the band, and the symmetrical bandwidth variable at this time is taken as the energy equivalent bandwidth; the intersection of the symmetrical frequency interval corresponding to the energy equivalent bandwidth and the near-fundamental interharmonic band is taken as the minimum symmetrical frequency band; two adjacent time window centers are selected, and the energy equivalent bandwidth corresponding to each of the two time window centers and the time interval between the two time window centers are obtained respectively; the energy equivalent bandwidth corresponding to the energy equivalent bandwidth corresponding to the latter time window center is subtracted from the energy equivalent bandwidth corresponding to the former time window center, and the difference is divided by the time interval between the two time window centers to obtain the energy equivalent bandwidth change rate.

[0031] It should be noted that the symmetric bandwidth variable represents an initial value greater than 0, which can be increased through subsequent adjustments, and is measured in Hertz. It is used to determine the radius of the symmetric frequency interval. The symmetric frequency interval represents a continuous frequency range with the energy weighting center frequency as the midpoint and half the radius of the symmetric bandwidth variable as the radius. This is the basic interval for calculating the energy accumulation by taking the intersection with the near-fundamental interharmonic band. The intersection of the symmetric frequency interval and the near-fundamental interharmonic band represents the overlapping frequency range that simultaneously belongs to both the symmetric frequency interval and the near-fundamental interharmonic band. Only the frequency grids within this range participate in the energy accumulation calculation, ensuring that the calculation range does not exceed the near-fundamental interharmonic band. The lower and upper limit frequency values ​​of the symmetric frequency interval and the near-fundamental interharmonic band are obtained. The larger of the two lower limit frequency values ​​is taken as the lower limit frequency value of the intersection, and the smaller of the two upper limit frequency values ​​is taken as the upper limit frequency value of the intersection. The continuous frequency range formed by the lower and upper limits is the intersection. The product result within the intersection represents a single frequency in the intersection. The power spectral density corresponding to the grid is multiplied by the frequency resolution, reflecting the energy contribution of a single frequency grid in the intersection. The energy accumulation is used to determine whether the current symmetric bandwidth variable meets the requirement of containing 95% of the total energy in the band, and is the basis for adjusting the symmetric bandwidth variable. The energy equivalent bandwidth represents the value of the symmetric bandwidth variable when the energy accumulation is not less than 95% of the total energy in the band, in Hertz, reflecting the equivalent width of the energy distribution in the near-fundamental interharmonic band. The minimum symmetric frequency band represents the intersection of the symmetric frequency interval corresponding to the energy equivalent bandwidth and the near-fundamental interharmonic band, which is the smallest continuous frequency range containing 95% of the energy in the near-fundamental interharmonic band, reflecting the core frequency domain region of concentrated energy distribution. The centers of two adjacent time windows represent the center moments of two adjacent time windows used for power spectrum estimation in the time series (the former is the center of the previous time window, and the latter is the center of the subsequent time window). The rate of change of the energy equivalent bandwidth is used to reflect the speed at which the energy equivalent bandwidth changes over time.

[0032] It should be noted that the initial value of the symmetrical bandwidth variable needs to be set in conjunction with the near-fundamental interharmonic band width, typically taking 1 / 10 to 1 / 5 of the harmonic band width. For example, when the harmonic band width is 10 Hz, the initial value is set to 1 to 2 Hz. This setting avoids multiple adjustments due to an initial value that is too small, or energy accumulation far exceeding the total energy within the 95% band due to an initial value that is too large, thus balancing adjustment efficiency and accuracy. The adjustment step size of the symmetrical bandwidth variable adopts an adaptive step size, with the initial step size taking 1 / 5 to 1 / 10 of the initial value (for example, when the initial value is 1 Hz, the step size is 0.1 to 0.2 Hz). If the energy accumulation is related to the 95% band width... If the difference in total energy within the 95% band exceeds 5%, maintain the current step size; if the difference is less than 1%, reduce the step size to half of the current size to avoid skipping the minimum equivalent bandwidth and ensure accuracy; the time interval must be consistent with and fixed to the sliding step size of the time window in short-time power spectrum estimation. For example, if the sliding step size is 0.1 seconds, the interval should also be set to 0.1 seconds; the accuracy of energy accumulation judgment should be controlled within 0.1%, that is, when the absolute value of the difference between the energy accumulation and the total energy within the 95% band is less than 0.1%, it is considered to meet the requirements; avoid over-adjustment due to small errors in power spectral density calculation to ensure stable equivalent bandwidth.

[0033] In one embodiment of the present invention, such as Fig. 2 As shown, the calculation of the ice-induced interharmonic spectrum dynamic broadening coefficient includes the following steps: Step S201: Subtract the wet-bulb temperature from the preset temperature reference value to obtain the wet-bulb temperature deviation value, multiply the wet-bulb temperature deviation value with the first scale parameter, input the monotonic bounded function, and obtain the wet-bulb temperature mapping result. Step S202: Subtract the preset humidity reference value from the relative humidity to obtain the relative humidity deviation value, multiply the relative humidity deviation value with the second scale parameter, input the monotonic bounded function, and obtain the relative humidity mapping result. Step S203: Subtract the preset wind speed reference value from the wind speed to obtain the wind speed deviation value, multiply the wind speed deviation value with the third scale parameter, input the monotonic bounded function, and obtain the wind speed mapping result. Step S204: Multiply the wet-bulb temperature mapping result, the relative humidity mapping result, and the wind speed mapping result to obtain the icing potential; Step S205: Obtain the energy equivalent bandwidth change rate, which is the time change rate of the energy equivalent bandwidth within the near-fundamental interharmonic band. Step S206: Multiply the rate of change of energy equivalent bandwidth by the icing potential to obtain the dynamic broadening coefficient of ice-induced interharmonic spectrum.

[0034] It should be noted that the wet-bulb temperature deviation value represents the difference between the preset temperature benchmark value and the wet-bulb temperature, used to quantify the contribution of wet-bulb temperature to icing conditions; the larger the difference, the stronger the icing potential. The wet-bulb temperature mapping result represents the value obtained by inputting the product of the wet-bulb temperature deviation value and the first scale parameter into a monotonically bounded function. This value is between 0 and 1, reflecting the contribution of the wet-bulb temperature dimension to the icing potential. The first scale parameter represents a learnable parameter greater than 0, used to adjust the influence weight of the wet-bulb temperature deviation value on subsequent mapping results. The monotonically bounded function represents a function whose output value monotonically increases as the input value increases, and whose output value always remains within a fixed range. The relative humidity deviation value represents the contribution of relative humidity to icing conditions; the larger the difference, the more favorable the high humidity environment is for icing. The second scale parameter represents a learnable parameter greater than 0, used to adjust the influence weight of the relative humidity deviation value on subsequent mapping results. The relative humidity mapping result represents the value obtained by inputting the product of the relative humidity deviation value and the second-scale parameter into a monotonically bounded function. This value is between 0 and 1, reflecting the contribution of relative humidity to icing potential. The wind speed deviation value is used to quantify the contribution of wind speed to icing conditions; the larger the difference, the more favorable the convective heat transfer is for icing. The third-scale parameter represents a learnable parameter greater than 0, used to adjust the influence weight of the wind speed deviation value on subsequent mapping results. The wind speed mapping result represents the value obtained by inputting the product of the wind speed deviation value and the third-scale parameter into a monotonically bounded function. This value is between 0 and 1, reflecting the contribution of wind speed to icing potential. Icing potential represents a comprehensive index integrating environmental parameters; the larger the value, the more favorable the environment is for icing. The ice-induced interharmonic spectrum dynamic broadening coefficient represents the core fusion characteristic of electrical parameters (bandwidth change) and environmental quantities (icing potential); the larger the value, the higher the possibility of grid connection risk caused by icing.

[0035] It should be noted that the preset temperature, preset humidity, and preset wind speed reference values ​​are all preset parameters. Preferably, the preset temperature reference value is set to 0 degrees Celsius, because 0 degrees Celsius is the critical temperature for water to freeze. Below this temperature, the wet-bulb temperature deviation is positive, reflecting the cooling potential. The preset humidity reference value is set to 0.6 (i.e., 60%), because when the relative humidity is higher than 60%, there is sufficient water vapor in the air, which is conducive to freezing. The preset wind speed reference value is set to 3 meters per second, because when the wind speed is higher than 3 meters per second, convective heat transfer is enhanced, which can accelerate the cooling and freezing process of the equipment surface. The monotonic bounded function adopts the Sigmoid function, whose expression is that the function output value equals 1 divided by (1 / (1+1)) / (1+1)). (Adding the negative power of the natural constant); This function can compress any input value to the range of 0 to 1, and the larger the input value, the closer the output value is to 1, that is, the more favorable the environment for icing, the larger the mapping result, while avoiding the excessive influence of extreme values ​​on the icing potential; The first-scale parameter, the second-scale parameter, and the third-scale parameter are determined through offline training. The training data are the environmental parameters (wet-bulb temperature, relative humidity, wind speed) and electrical parameters (energy equivalent bandwidth change rate) corresponding to historical icing events; With the goal of maximizing the correlation between the icing potential and the actual probability of icing, the values ​​of the three-scale parameters are adjusted through the gradient descent optimization algorithm, and finally stable parameter values ​​greater than zero are obtained, which will not be elaborated here.

[0036] In one embodiment of the present invention, such as Fig. 3 As shown, the calculation of the combined vector includes the following steps: Step S301: Subtract the forgetting factor from 1, multiply by the exponential moving average of the previous time step, and add the product of the ice-induced interharmonic spectrum broadening coefficient and the forgetting factor at the current time step to obtain the exponential moving average at the current time step. Step S302: Subtract the forgetting factor from 1, multiply by the exponential moving variance of the previous time step, add the square of the difference between the current ice-induced interharmonic spectrum broadening coefficient and the exponential moving mean of the previous time step, and then multiply by the forgetting factor to obtain the exponential moving variance of the current time step. Step S303: Take the square root of the exponential moving variance at the current time to obtain the exponential moving standard deviation at the current time. Subtract the exponential moving mean at the current time from the ice-induced interharmonic spectrum broadening coefficient at the current time to obtain the difference in coefficient mean. Step S304: Add a preset constant to the exponential moving standard deviation at the current time to obtain the standard deviation correction value, and then divide the difference in the mean of the coefficients by the standard deviation correction value to obtain the standardized value. Step S305: Select the center time of two adjacent time windows, calculate the time interval between the center times of the two time windows, obtain the standardized values ​​corresponding to the center times of the two time windows respectively, and subtract the standardized value of the previous time from the standardized value of the current time to obtain the difference of standardized values. Step S306: Divide the difference of standardized values ​​by the time interval to obtain the first-order rate of change. Concatenate the standardized values ​​and the first-order rate of change to obtain the combined vector.

[0037] It should be noted that the forgetting factor is controlled between 0 and 1, and is used to control the ratio of historical data weight to current data weight in exponential moving average statistics. The larger the value, the more significant the impact of current data on the statistical results; the smaller the value, the higher the weight of historical data. Specifically, the forgetting factor can be set according to specific needs and scenarios. For example, if the model needs to be more sensitive to recent changes in the dynamic broadening coefficient of glacial interharmonic spectrum (e.g., during seasonal transitions), it can be set to 0.08 to 0.1; if more stable statistical results are needed (e.g., during periods of stable operating conditions), it can be set to 0.01 to 0.03; for example, during the high-risk period of icing in wind farms in winter, the forgetting factor is set to 0.1 to ensure timely capture of abnormal fluctuations in the coefficient; the exponential moving average of the previous moment is used to reflect the average level of the historical dynamic broadening coefficient of glacial interharmonic spectrum. The exponential moving average at the current moment is used to reflect the real-time average level of the ice-induced interharmonic spectrum dynamic broadening coefficient; the coefficient mean difference is used to quantify the degree to which the current coefficient deviates from the real-time average level; the preset constant is a preset parameter to prevent the denominator from being 0, preferably set to 10 to the power of -6; the standardized value is used to eliminate the difference in the ice-induced interharmonic spectrum dynamic broadening coefficient caused by scale drift; the standardized value difference is used to quantify the temporal change amplitude of the standardized coefficient; the first-order rate of change is used to reflect the rate of change of the standardized value over time; the combined vector represents a two-dimensional vector formed by concatenating the standardized value and the first-order rate of change at the current moment in a predetermined order (standardized value first, first-order rate of change second), integrating the static standardization characteristics and dynamic rate of change characteristics of the ice-induced interharmonic spectrum dynamic broadening coefficient.

[0038] It should be noted that the initial values ​​for the exponential sliding statistics are the statistical values ​​of the ice-induced interharmonic spectrum broadening coefficients corresponding to the center times of the first 10 time windows after system startup, i.e., the initial value of the exponential sliding mean of the previous time moment, which is set as the arithmetic mean of the first 10 coefficients; the initial value of the exponential sliding variance of the previous time moment is set as the arithmetic mean of the squares of the deviations of the first 10 coefficients from this arithmetic mean; the center times of adjacent time windows must be strictly selected according to the time window sliding step size of the short-time power spectrum estimation, that is, the interval between two times is equal to the sliding step size (for example, if the sliding step size is 0.1 seconds, then the interval between adjacent times is 0.1 seconds), and must be selected in the order of the time series (a certain time cannot be skipped) to ensure uniform time intervals and avoid time scale confusion in the calculation of the first-order rate of change.

[0039] In one embodiment of the present invention, multiple non-negative causal convolutional branches are constructed, the kernel length of each branch is set, and multiple lag weights are configured for each branch; for each branch, the input combination vector and its historical time combination vector are multiplied by the corresponding lag weights respectively, and all multiplication results are added together to obtain the convolutional output of each branch. Configure a trainable matrix for each non-negative causal convolution branch, multiply the convolution output of each branch with the corresponding trainable matrix, and sum all the multiplication results to obtain the driving force. Set the hidden state vector of the stable discrete state space kernel, and configure four trainable matrices, namely the first trainable matrix, the second trainable matrix, the third trainable matrix, and the fourth trainable matrix; apply hyperbolic tangent operation to each element of the first trainable matrix; Multiply the first trainable matrix (after hyperbolic tangent operation) with the hidden state vector at the current time step, multiply the second trainable matrix with the driving force, and add the two multiplication results to obtain the hidden state vector at the next time step; multiply the third trainable matrix with the hidden state vector at the current time step, and multiply the fourth trainable matrix with the driving force, and add the two multiplication results to obtain the intermediate risk quantity.

[0040] It should be noted that the non-negative causal convolution branch represents a parallel computation branch used for extracting temporal features of the combined vector at multiple scales; the kernel length represents the number of lag weights in each non-negative causal convolution branch; the lag weights represent non-negative parameters configured for each non-negative causal convolution branch, used to allocate the contribution weights of the combined vector at the current time and different historical time points to the branch output; the larger the weight, the stronger the influence of the combined vector at the corresponding time point; the convolution output represents the output value of a single non-negative causal convolution branch after extracting the temporal features of the combined vector, reflecting the feature strength at that branch's time scale; the trainable matrix represents the value configured for each non-negative causal convolution branch. The adjustable matrix parameters during training are used to map the branch convolution outputs to a unified dimension; the driving quantity represents the value after integrating the output features of all non-negative causal convolution branches, used to drive the state update of the stable discrete state space kernel and the calculation of intermediate risk quantity, and is the key variable connecting the convolutional layer and the state space layer; the hidden state vector refers to the vector in the stable discrete state space kernel used to store historical time-series information, which can accumulate the long-term features of the combined vector, and its value is dynamically updated with the driving quantity; the intermediate risk quantity represents the risk feature value output by the stable discrete state space kernel without undergoing the final nonlinear mapping, integrating historical hidden state information and current driving information.

[0041] It should be noted that the number of non-negative causal convolution branches is determined based on the temporal scale distribution of icing risk features in historical data, typically set to 3 to 5. For example, if historical data shows that risk features exist at three main time scales: 0.5 seconds, 1 second, and 2 seconds, then 3 branches are set to correspond to these three scales respectively, ensuring that no multi-scale features are missed. The convolution kernel length is related to the target time scale of the branch and the sampling interval, and the calculation formula is: convolution kernel length = target time scale ÷ sampling interval. For example, if the target time scale is 1 second and the sampling interval is 0.1 seconds, then the convolution kernel length is set to 10, ensuring that the branch can cover the temporal features within 1 second. The initial value of the hidden state vector is set to an all-zero vector, that is, the value of each element in the vector is 0. The branches are trainable. The matrix has the following dimensions: the driving vector dimension × the branch convolution output dimension. The driving vector dimension is usually set to 8 to 16, and the branch convolution output dimension is the same as the combined vector dimension (2-dimensional). The first trainable matrix has the following dimensions: the hidden state vector dimension × the hidden state vector dimension, which is the same as the driving vector dimension. The second trainable matrix has the following dimensions: the hidden state vector dimension × the driving vector dimension. The third trainable matrix has the following dimensions: 1 × the hidden state vector dimension (outputting a single-value intermediate risk). The fourth trainable matrix has the following dimensions: 1 × the driving vector dimension. For example, if the driving vector dimension is 10, then the first trainable matrix is ​​10 × 10, the second trainable matrix is ​​10 × 10, the third trainable matrix is ​​1 × 10, and the fourth trainable matrix is ​​1 × 10.

[0042] In one embodiment of the present invention, two trainable parameters of the monotonic output head are configured, namely a gain parameter and an offset parameter; a softplus function is applied to the gain parameter to obtain a non-negative gain; the intermediate risk quantity is multiplied by the non-negative gain, and the result of the multiplication is added to the offset parameter to obtain the intermediate value before output; a sigmoid function is applied to the intermediate value before output to obtain the risk index. The loss function of the risk prediction model is set as the cross-entropy loss function. The labels of the historical training data are obtained offline. These labels are indicators of whether grid connection oscillation or flicker of new energy power generation has occurred within a preset time after the calculation of the historical risk index. The value of the label is 0 or 1.

[0043] It should be noted that the monotonic output head represents the output layer structure in the neural network used to map the intermediate risk quantity to a risk index ranging from 0 to 1; the gain parameter represents a trainable parameter in the monotonic output head used to adjust the amplification factor of the intermediate risk quantity, with an initial value of random non-negative. During training, it is optimized according to the loss function and finally constrained to non-negativity by the softplus function to ensure that the gain effect on the intermediate risk quantity does not reverse; the offset parameter represents a trainable parameter in the monotonic output head used to adjust the overall offset of the intermediate value before output, and its value can be positive or negative. It is optimized during training to compensate for the basic offset between the intermediate risk quantity and the risk index; the non-negative gain controls the contribution strength of the intermediate risk quantity to the risk index; the larger the value, the more significant the impact of changes in the intermediate risk quantity on the risk index; the risk index represents the final output of the monotonic output head, with a value between 0 and 1. The closer the value is to 1, the higher the probability of future grid connection risk; the closer it is to 0, the lower the risk. The loss function of the risk prediction model is the cross-entropy loss function, which measures the deviation between the predicted risk index value and the true historical label value. During training, this function is minimized to optimize all trainable parameters of the neural network. The preset duration is a preset parameter, which counts whether a risk event has occurred within the time limit starting from the time of historical risk index calculation. Preferably, the preset duration is set to 30 seconds. Grid connection oscillation of new energy power generation refers to the periodic fluctuation of grid current or voltage caused by factors such as icing when the new energy power generation system is connected to the grid. The frequency is usually close to the fundamental frequency, which will affect the stability of grid power supply. Flicker refers to the rapid fluctuation of grid voltage amplitude caused by factors such as interharmonics when the new energy power generation system is connected to the grid, which will cause light flicker and affect the normal operation of electrical equipment.

[0044] It should be noted that the labeling of historical training data adopts the power grid operation standard for judgment, namely: 1. Grid-connected oscillation: When the current fluctuation amplitude at the grid connection point exceeds 5% of the rated current and the fluctuation frequency is in the range of 45 to 55 Hz for more than 1 second, it is judged as oscillation; 2. Flicker: When the voltage amplitude fluctuation frequency at the grid connection point is in the range of 0.5 to 35 Hz and the fluctuation amplitude exceeds 0.9% of the rated voltage, it is judged as flicker. If any of the above events occurs within the preset time, the label is marked as 1, otherwise it is marked as 0; the initial value of the gain parameter is set to 0.1; the initial value of the offset parameter is set to 0 (no initial offset, allowing the model to learn autonomously) to ensure stable output in the early stage of model training, which will not be elaborated here.

[0045] In one embodiment of the present invention, the weight is obtained by subtracting the freezing potential from 1; if the risk index at the current moment is greater than the upper quantile baseline at the previous moment, the result of the indicator function is 1, otherwise the result of the indicator function is 0; the quantile position is subtracted from 1, and then the difference is subtracted from the result of the indicator function, multiplied by the step size, and then multiplied by the weight to obtain the upper quantile baseline update amount; the upper quantile baseline at the previous moment is added to the upper quantile baseline update amount to obtain the upper quantile baseline at the current moment. The deviation is obtained by subtracting the upper quantile baseline of the previous time from the current risk index and taking the absolute value of the difference. The mean absolute deviation index moving estimate of the current time is obtained by subtracting the forgetting factor from 1 and multiplying it by the mean absolute deviation index moving estimate of the previous time, adding the forgetting factor multiplied by the weight and then multiplying by the deviation. The adaptive threshold is obtained by adding the safety margin coefficient to the upper quantile baseline of the current time and multiplying it by the mean absolute deviation index moving estimate of the current time.

[0046] It should be noted that the higher the icing intensity, the lower the weight; during normal periods, the weight is close to 1, ensuring that the baseline reflects normal operating conditions. The indicator function is used to determine the relative magnitude of the current risk index and the upper quantile baseline of the previous moment, and the output result is only 1 or 0. The quantile position ranges from 0 to 1; the smaller the quantile position, the higher the upper quantile baseline (for example, a quantile position of 0.05 corresponds to the 95th upper quantile, meaning that during normal periods, the 95th percentile risk index is lower than this baseline). The step size is a small positive number used to control the update amplitude of the upper quantile baseline; the smaller the step size, the smoother the baseline update, avoiding sharp changes in the baseline due to a single abnormal risk index. The upper quantile baseline update represents the incremental value used to adjust the upper quantile baseline. It can be positive or negative. A positive increment raises the baseline, while a negative increment lowers the baseline, ensuring that the baseline gradually approaches the true upper quantile. The deviation reflects the extent to which the current risk index deviates from the historical baseline. The safety margin coefficient represents the positive value used to amplify the average absolute deviation estimate and ensure that the adaptive threshold covers normal fluctuations. The larger the coefficient, the higher the threshold and the lower the false alarm rate, but the risk of missed alarms may increase. The adaptive threshold represents the final judgment threshold obtained by combining the current upper quantile baseline and the weighted average absolute deviation, which can dynamically adapt to fluctuations under normal operating conditions.

[0047] It should be noted that the quantile value ranges from 0.01 to 0.1, and is set based on the concentration of the risk index distribution during normal periods. Specifically, if the risk index distribution is concentrated (small fluctuations) during normal periods, a quantile of 0.01 to 0.03 is used (corresponding to the 99th to 97th upper quantile); if the distribution is dispersed (large fluctuations), a quantile of 0.05 to 0.1 is used (corresponding to the 95th to 90th upper quantile). For example, the risk index of a wind farm during normal periods is mostly below 0.3, so a quantile of 0.05 is used to ensure that the risk index of 95% of normal samples is below the baseline. The step size ranges from 0.01 to 0.05, and is set according to the principle of prioritizing baseline update stability. That is, the step size needs to be small enough that the amount of a single update does not exceed 5% of the baseline at the previous moment, avoiding large jumps in the baseline risk index due to occasional anomalies. For example, at the previous moment... The baseline is 0.4, the step size is 0.02, and the maximum update amount per time (weight 1, indicator function 1, quantile position 0.05) is 0.02×1×(1-0.95)=0.001, which is only 0.25% of the baseline, ensuring stability. The safety margin coefficient ranges from 1.5 to 2.5. The adjustment principle is as follows: if the field has a low tolerance for false alarms (e.g., important load grid connection), take 2.0 to 2.5 to increase the threshold and reduce false alarms; if sensitive to missed alarms (e.g., remote wind farms), take 1.5 to 1.8 to decrease the threshold and reduce missed alarms; for example, if the grid connection end is connected to important loads such as hospitals, the safety margin coefficient is 2.2 to ensure that normal fluctuations will not trigger false alarms. In addition, when the risk index value at the current moment is equal to the upper quantile baseline value at the previous moment, the indicator function result is uniformly taken as 0.

[0048] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0049] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A new energy power generation safety monitoring system based on artificial intelligence, characterized in that, include: The power spectrum estimation module collects three-phase current at the grid connection point to construct a scalar current trajectory and simultaneously collects environmental parameters. It performs short-time power spectrum estimation on the scalar current trajectory to obtain the power spectral density. It determines the near-fundamental interharmonic band within a predetermined narrow frequency band on both sides of the fundamental frequency of the power grid and removes sub-frequency bands containing the fundamental frequency. Environmental parameters include: wet-bulb temperature, relative humidity, and wind speed; The equivalent bandwidth calculation module calculates the total energy and energy weighted center frequency within the near-fundamental interharmonic band, thereby determining the minimum symmetrical frequency band and energy equivalent bandwidth, and calculating the rate of change of energy equivalent bandwidth. The combined vector construction module takes the environmental parameters as input to a continuous monotonic learnable mapping, outputs the icing potential, multiplies it by the rate of change of energy equivalent bandwidth to obtain the dynamic broadening coefficient of the ice-induced interharmonic spectrum, and performs online exponential sliding mean and standard deviation estimation and time difference estimation on the coefficient to obtain the combined vector. The risk index prediction module takes the combined vector as input, constructs a risk prediction model containing non-negative causal convolution, a stable discrete state space kernel, and a monotonic output head, and outputs the risk index. The adaptive threshold generation module performs online quantile tracking of the risk index to obtain the upper quantile baseline, and calculates the adaptive threshold by combining the freezing potential and safety margin coefficient. The power generation risk assessment module determines whether the risk index is greater than or equal to an adaptive threshold. If so, it indicates that there is a power generation risk in the new energy power generation; otherwise, it indicates that there is no power generation risk.

2. The new energy power generation safety monitoring system based on artificial intelligence according to claim 1, characterized in that, The discrete current samples of each phase corresponding to the same time stamp in the three-phase current discrete sequence are squared. The results of the three square operations are summed, and the sum is divided by 3. Finally, the square root of the quotient is taken to obtain the discrete form of the scalar current trajectory. The frequency resolution is determined according to the sampling frequency and the length of the window function. Then, the frequency grid is determined according to the frequency resolution. Each frequency value in the frequency grid is an integer multiple of the frequency resolution. The discrete form of the scalar current trajectory is multiplied by the window function. Each multiplication result is multiplied by the corresponding complex exponential term. Finally, all multiplication results are summed to obtain the short-time Fourier transform result. The modulus of the short-time Fourier transform result is taken first. Then, the modulus is squared. The squared result is divided by the product of the window function length and the sampling frequency to obtain the power spectral density.

3. The new energy power generation safety monitoring system based on artificial intelligence according to claim 2, characterized in that, The fundamental frequency of the power grid is set to 50 Hz. A symmetrical bandwidth parameter is set and is greater than 0. Taking the fundamental frequency of the power grid as the center, the frequency value obtained by subtracting the symmetrical bandwidth parameter from the fundamental frequency of the power grid is used as the lower limit, and the frequency value obtained by adding the symmetrical bandwidth parameter to the fundamental frequency of the power grid is used as the upper limit, thus forming a predetermined narrow frequency band on both sides of the fundamental frequency. The sub-frequency band containing the fundamental frequency in the predetermined narrow frequency band is removed, and the remaining frequency range constitutes the near-fundamental interharmonic band, wherein the sub-frequency band containing the fundamental frequency is a preset parameter.

4. The artificial intelligence-based new energy power generation safety monitoring system according to claim 3, characterized in that, Multiply the power spectral density corresponding to each frequency grid within the near-fundamental interharmonic band by the frequency resolution to obtain the first product result. Then, sum all the first product results to obtain the total energy within the band. Multiply the frequency value of each frequency grid within the near-fundamental interharmonic band by the power spectral density corresponding to that frequency grid to obtain the second product result. Sum all the second product results and divide by the total energy within the band to obtain the energy-weighted center frequency.

5. The new energy power generation safety monitoring system based on artificial intelligence according to claim 4, characterized in that, Using the energy-weighted center frequency as the midpoint and half the radius of the symmetrical bandwidth variable as the interval radius, a symmetrical frequency interval is obtained. The power spectral density corresponding to each frequency grid within the intersection is multiplied by the frequency resolution to obtain the product result within the intersection. All product results within the intersection are added together to obtain the energy accumulation. If the energy accumulation is less than 95% of the total energy within the band, the symmetrical bandwidth variable is adjusted and increased until the energy accumulation is not less than 95% of the total energy within the band. The symmetrical bandwidth variable at this point is taken as the energy equivalent bandwidth. The intersection of the symmetrical frequency interval corresponding to this energy equivalent bandwidth and the near-fundamental interharmonic band is taken as the minimum symmetrical frequency band. Two adjacent time window centers are selected, and the energy equivalent bandwidth corresponding to each of the two time window centers and the time interval between the two time window centers are obtained respectively. The energy equivalent bandwidth corresponding to the energy equivalent bandwidth of the previous time window center is subtracted from the energy equivalent bandwidth corresponding to the energy equivalent bandwidth of the latter time window center, and the difference is divided by the time interval between the two time window centers to obtain the rate of change of the energy equivalent bandwidth.

6. The new energy power generation safety monitoring system based on artificial intelligence according to claim 1, characterized in that, The calculation of the dynamic broadening factor of the ice-induced interharmonic spectrum includes the following steps: Step S201: Subtract the wet-bulb temperature from the preset temperature reference value to obtain the wet-bulb temperature deviation value, multiply the wet-bulb temperature deviation value with the first scale parameter, input the monotonic bounded function, and obtain the wet-bulb temperature mapping result. Step S202: Subtract the preset humidity reference value from the relative humidity to obtain the relative humidity deviation value, multiply the relative humidity deviation value with the second scale parameter, input the monotonic bounded function, and obtain the relative humidity mapping result. Step S203: Subtract the preset wind speed reference value from the wind speed to obtain the wind speed deviation value, multiply the wind speed deviation value with the third scale parameter, input the monotonic bounded function, and obtain the wind speed mapping result. Step S204: Multiply the wet-bulb temperature mapping result, the relative humidity mapping result, and the wind speed mapping result to obtain the icing potential; Step S205: Obtain the energy equivalent bandwidth change rate, which is the time change rate of the energy equivalent bandwidth within the near-fundamental interharmonic band. Step S206: Multiply the rate of change of energy equivalent bandwidth by the icing potential to obtain the dynamic broadening coefficient of ice-induced interharmonic spectrum.

7. The new energy power generation safety monitoring system based on artificial intelligence according to claim 6, characterized in that, The calculation of the combined vector includes the following steps: Step S301: Subtract the forgetting factor from 1, multiply by the exponential moving average of the previous time step, and add the product of the ice-induced interharmonic spectrum broadening coefficient and the forgetting factor at the current time step to obtain the exponential moving average at the current time step. Step S302: Subtract the forgetting factor from 1, multiply by the exponential moving variance of the previous time step, add the square of the difference between the current ice-induced interharmonic spectrum broadening coefficient and the exponential moving mean of the previous time step, and then multiply by the forgetting factor to obtain the exponential moving variance of the current time step. Step S303: Take the square root of the exponential moving variance at the current time to obtain the exponential moving standard deviation at the current time. Subtract the exponential moving mean at the current time from the ice-induced interharmonic spectrum broadening coefficient at the current time to obtain the difference in coefficient mean. Step S304: Add a preset constant to the exponential moving standard deviation at the current time to obtain the standard deviation correction value, and then divide the difference in the mean of the coefficients by the standard deviation correction value to obtain the standardized value. Step S305: Select the center time of two adjacent time windows, calculate the time interval between the center times of the two time windows, obtain the standardized values ​​corresponding to the center times of the two time windows respectively, and subtract the standardized value of the previous time from the standardized value of the current time to obtain the difference of standardized values. Step S306: Divide the difference of standardized values ​​by the time interval to obtain the first-order rate of change. Concatenate the standardized values ​​and the first-order rate of change to obtain the combined vector.

8. The new energy power generation safety monitoring system based on artificial intelligence according to claim 1, characterized in that, Construct multiple non-negative causal convolutional branches, set the kernel length of each branch, and configure multiple lag weights for each branch; for each branch, multiply the input combination vector and its historical combination vector by the corresponding lag weights respectively, and sum all the multiplication results to obtain the convolution output of each branch; Configure a trainable matrix for each non-negative causal convolution branch, multiply the convolution output of each branch with the corresponding trainable matrix, and sum all the multiplication results to obtain the driving force. Set the hidden state vector of the stable discrete state space kernel, and configure four trainable matrices, namely the first trainable matrix, the second trainable matrix, the third trainable matrix, and the fourth trainable matrix; apply hyperbolic tangent operation to each element of the first trainable matrix; Multiply the first trainable matrix (after hyperbolic tangent operation) with the hidden state vector at the current time step, multiply the second trainable matrix with the driving force, and add the two multiplication results to obtain the hidden state vector at the next time step; multiply the third trainable matrix with the hidden state vector at the current time step, and multiply the fourth trainable matrix with the driving force, and add the two multiplication results to obtain the intermediate risk quantity.

9. A new energy power generation safety monitoring system based on artificial intelligence according to claim 8, characterized in that, Configure two trainable parameters for the monotonic output head: the gain parameter and the offset parameter. Apply the softplus function to the gain parameter to obtain a non-negative gain; multiply the intermediate risk amount by the non-negative gain, and add the result of the multiplication to the offset parameter to obtain the intermediate value before the output; The risk index is obtained by applying the Sigmoid function to the intermediate value before the output. The loss function of the risk prediction model is set as the cross-entropy loss function. The labels of the historical training data are obtained offline. These labels are indicators of whether grid connection oscillation or flicker of new energy power generation has occurred within a preset time after the calculation of the historical risk index. The value of the label is 0 or 1.

10. The new energy power generation safety monitoring system based on artificial intelligence according to claim 1, characterized in that, Subtract the freezing potential from 1 to obtain the weight; determine if the risk index at the current moment is greater than the upper quantile baseline at the previous moment, then the indicator function result is 1, otherwise the indicator function result is 0; first subtract the quantile position from 1, then subtract the difference from the indicator function result, multiply by the step size, and then multiply by the weight to obtain the upper quantile baseline update amount. Add the upper quantile baseline update to the upper quantile baseline of the previous moment to obtain the upper quantile baseline of the current moment; The deviation is obtained by subtracting the upper quantile baseline of the previous time from the current risk index and taking the absolute value of the difference. The mean absolute deviation index moving estimate of the current time is obtained by subtracting the forgetting factor from 1 and multiplying it by the mean absolute deviation index moving estimate of the previous time, adding the forgetting factor multiplied by the weight and then multiplying by the deviation. The adaptive threshold is obtained by adding the safety margin coefficient to the upper quantile baseline of the current time and multiplying it by the mean absolute deviation index moving estimate of the current time.

Images