Machine learning based distributed photovoltaic module failure prediction method

Abstract

This invention relates to the field of photovoltaic intelligent operation and maintenance technology, and discloses a method for predicting faults in distributed photovoltaic modules based on machine learning. The method includes: collecting DC voltage and current sequences of the modules, constructing a power time series through time-axis normalization, and decomposing it into a steady-state envelope component representing environmental changes and a transient disturbance component representing control actions using a filter; performing spectral analysis on the transient disturbance component, locking the maximum power point tracking oscillation frequency within the current window, extracting the intermodulation sideband energy distribution around this frequency and its harmonics, and calculating the bypass coupling index reflecting the nonlinearity of the module accordingly; constructing a prediction model including a long-time-series evolution extraction channel, a short-time-series waveform extraction channel, and an adaptive gating unit; dynamically weighting and fusing frequency domain structure information and time domain response information using the attention coefficient generated by the adaptive gating unit based on the bypass coupling index; and mapping the fused comprehensive representation to a health state probability.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic intelligent operation and maintenance technology, and more specifically, to a method for predicting faults in distributed photovoltaic modules based on machine learning. Background Technology

[0002] Distributed photovoltaic (PV) power generation systems have been widely deployed in industrial, commercial, and residential sectors due to their advantages such as utilizing idle rooftop resources and enabling on-site energy consumption. With the widespread application of component-level power electronic devices (MLPEs) such as micro-inverters and power optimizers, operation and maintenance (O&M) systems now possess the capability to perform high-frequency sampling of the DC voltage, current, and power of individual PV modules. Utilizing this finer-grained monitoring data to perceive the module's operating status and identify potential risks is crucial for reducing O&M costs and ensuring long-term stable power generation. However, in complex rooftop application scenarios, PV modules are often affected by fixed obstacles such as chimneys, trees, or parapet walls, creating regular shadows that change with the sun's trajectory. This causes the sub-string structure within the module to frequently switch between on and bypass states, resulting in a multi-peaked and non-smooth nonlinear output power curve. Simultaneously, the inverter's maximum power point tracking (MPPT) control algorithm continuously injects periodic disturbance signals into the system to lock into the optimal operating point. This high-frequency control disturbance and low-frequency shading envelope physically couple in nonlinear components, causing normal shading to exhibit a high degree of aliasing in electrical indicators compared to internal defects such as bypass diode failure and microcracks. Existing diagnostic methods based on time-domain waveform comparison or simple statistics often fail to effectively separate external environmental influences from internal component degradation, easily leading to misjudgments. Furthermore, existing analytical techniques typically ignore the interference of inverter control parameters (such as disturbance frequency and step size) on the frequency distribution of monitoring signals. When inverters undergo firmware upgrades, dynamic adjustments to control strategies, or the use of different brands of equipment, the signal's spectral structure drifts significantly, causing the previously trained discrimination model to fail due to the altered input distribution. How to deeply explore the intermodulation mechanism between the control loop and component structure under strong non-stationary control disturbances, extract deep fingerprint information robust to changes in control parameters, and thus achieve accurate prediction of early structural faults in components, is a key technical problem urgently needing to be solved in the field of photovoltaic intelligent operation and maintenance. Summary of the Invention

[0003] This invention provides a machine learning-based method for predicting the failure of distributed photovoltaic modules, which solves the technical problems mentioned in the background.

[0004] This invention provides a machine learning-based method for predicting faults in distributed photovoltaic modules, comprising: Acquire DC voltage and DC current sampling data of photovoltaic modules, construct power time series data, and decompose the power time series data into steady-state envelope components characterizing changes in environmental irradiance and transient disturbance components characterizing the action of control loops; Spectral feature analysis is performed on the transient disturbance component to lock the maximum power point tracking oscillation frequency within the current time window. The nonlinear response sideband energy spectrum around the maximum power point tracking oscillation frequency and its harmonics is extracted, and the bypass coupling characteristic index characterizing the degree of nonlinear action of the bypass diode is calculated based on the nonlinear response sideband energy spectrum. A fault prediction model is constructed, comprising a long-time-series evolution extraction channel, a short-time-series waveform extraction channel, and an adaptive gating unit. The adaptive gating unit is configured to receive a state vector composed of the bypass coupling feature index and generate channel attention coefficients accordingly. The channel attention coefficients are then used to dynamically and weightedly fuse the frequency domain structure features output by the long-time-series evolution extraction channel with the time domain response features output by the short-time-series waveform extraction channel. The fused features are mapped to component health status probabilities, and a fault determination result is output when the component health status probability exceeds a preset threshold.

[0005] The beneficial effects of this invention include: by deeply exploring the intermodulation sideband diffusion effect generated by maximum power point tracking control disturbances and environmental shading slow envelopes at the nonlinear inflection point of the component structure, a criterion capable of accurately decoupling external environmental interference and internal device degradation is constructed; by explicitly locking the oscillation frequency and introducing an adaptive gating mechanism, this invention effectively overcomes the problem of inconsistent signal spectrum distribution caused by inverter control parameter drift or firmware version differences, thereby achieving high robustness and high sensitivity prediction of early structural faults such as bypass diode failure in the complex dynamic shading environment of distributed rooftops. Attached Figure Description

[0006] Figure 1 This is a flowchart of the machine learning-based distributed photovoltaic module fault prediction method of the present invention; Figure 2 This is an architecture diagram of the machine learning-based distributed photovoltaic module fault prediction system of the present invention. Detailed Implementation

[0007] 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.

[0008] like Figure 1 As shown, the machine learning-based method for predicting the failure of distributed photovoltaic modules includes: Acquire DC voltage and DC current sampling data of photovoltaic modules, construct power time series data, and decompose the power time series data into steady-state envelope components characterizing changes in environmental irradiance and transient disturbance components characterizing the action of control loops; Spectral feature analysis is performed on the transient disturbance component to lock the maximum power point tracking oscillation frequency within the current time window. The nonlinear response sideband energy spectrum around the maximum power point tracking oscillation frequency and its harmonics is extracted, and the bypass coupling characteristic index characterizing the degree of nonlinear action of the bypass diode is calculated based on the nonlinear response sideband energy spectrum. A fault prediction model is constructed, comprising a long-time-series evolution extraction channel, a short-time-series waveform extraction channel, and an adaptive gating unit. The adaptive gating unit is configured to receive a state vector composed of the bypass coupling feature index and generate channel attention coefficients accordingly. The channel attention coefficients are then used to dynamically and weightedly fuse the frequency domain structure features output by the long-time-series evolution extraction channel with the time domain response features output by the short-time-series waveform extraction channel. The fused features are mapped to component health status probabilities, and a fault determination result is output when the component health status probability exceeds a preset threshold.

[0009] Preferably, the DC voltage and DC current sampling data of the photovoltaic module are acquired to construct power time-series data, including: Set a fixed uniform sampling interval Construct a system containing several time nodes A unified time grid is established; each time node is calculated using a linear interpolation algorithm. DC voltage value at the location With DC current value The calculation formula is as follows: ; ; ; ; In the formula, Represents the first in the unified time grid Each time point The starting time; , These represent the original acquisition sequence at the original time. The voltage and current values; , These represent the original acquisition sequence at the original time. The voltage and current values; This indicates that the power timing data is at time node. The power value.

[0010] DC voltage sampling data is a set of discrete DC voltage samples collected from the positive and negative terminals of a photovoltaic module during continuous operation, used to reflect the instantaneous operating voltage level of the module.

[0011] DC current sampling data is a set of discrete DC current samples collected during the continuous operation of the output circuit of a photovoltaic module, used to reflect the instantaneous output current level of the module.

[0012] The original DC voltage sequence is a sequence of DC voltage samples arranged in chronological order of the original acquisition time.

[0013] The original DC current sequence is a sequence of DC current samples arranged in chronological order of the original acquisition time.

[0014] A uniform sampling interval is a fixed time difference between two adjacent time nodes when constructing a uniform time grid. It is preferably between 0.1 and 0.5 seconds. This range can usually enable the sampling frequency to reach 2 Hz to 10 Hz, which can cover the maximum power point tracking disturbance of distributed photovoltaic modules and its several harmonics, without putting too much buffering and communication pressure on the edge gateway.

[0015] A unified time grid is a set of equally spaced discrete time coordinates constructed within an observation period based on a unified sampling interval.

[0016] A time node is a discrete moment in a unified time grid, used to identify the standard time position for interpolation and power calculation.

[0017] The start time is the initial time of the unified time grid, used to determine the generation baseline for all time nodes. In engineering implementation, the start time is preferably the earliest time within the current analysis batch that simultaneously has effective voltage and effective current samples.

[0018] The original time is the actual sampling time corresponding to the sample in the original acquisition sequence, used to determine the left endpoint position of the linear interpolation.

[0019] The adjacent original time is the next valid sample time in the original acquisition sequence that is adjacent to the previous original time, and is used to determine the right endpoint position of the linear interpolation.

[0020] The aligned DC voltage values ​​are the voltage interpolation results obtained at the corresponding time nodes after the original DC voltage sequence is mapped through a unified time grid.

[0021] The aligned DC current value is the current interpolation result obtained at the corresponding time node after the original DC current sequence is mapped through a unified time grid.

[0022] The voltage value at the original time is the measured voltage sample value of the original DC voltage sequence at the corresponding original time.

[0023] The current value at the original moment is the measured current sample value of the original DC current sequence at the corresponding original moment.

[0024] The voltage values ​​at adjacent original times are the measured voltage sample values ​​of the original DC voltage sequence at adjacent original times.

[0025] The current values ​​at adjacent original times are the measured current sample values ​​of the original DC current sequence at adjacent original times.

[0026] The power value is the instantaneous power value obtained by multiplying the aligned DC voltage value and the aligned DC current value at the corresponding time node, and is used to characterize the output power level of the component at that discrete moment.

[0027] Power time series data is a discrete time series consisting of all power values ​​arranged in a uniform time grid order, used to fully characterize the dynamic changes in the output power of photovoltaic modules within the observation period.

[0028] The time node number is an integer index used to identify the sequential order of discrete time nodes in a unified time grid.

[0029] In detail, the method for setting the unified sampling interval is as follows: First, statistically analyze the actual sampling period distribution of the DC voltage channel and the DC current channel under normal sampling conditions. Then, select the more stable and shorter typical sampling period of the two as a candidate benchmark. Subsequently, determine the unified sampling interval by combining the maximum power point tracking control frequency and the edge gateway processing capability. In engineering practice, it is preferable to make the unified sampling frequency at least ten times the control frequency. For example, when the control frequency is about 0.5 Hz, the unified sampling interval can be set to 0.2 seconds to obtain a 5 Hz sampling frequency.

[0030] In detail, the method for determining the start and end boundaries of the unified time grid is as follows: the start time is defined as the earliest time in the current analysis batch that simultaneously has effective voltage and current samples, and the end time is defined as the last effective time that can still be covered by both voltage and current sequences. Then, time nodes are generated point by point according to the unified sampling interval. If processed on a daily basis, the starting point of the first effective window of the day is preferred as the start boundary, and the ending point of the last effective window of the day is preferred as the end boundary.

[0031] In detail, the boundary and missing sample handling methods of the linear interpolation algorithm are as follows: linear interpolation is only performed when the time node is between two adjacent valid original sample times; if the time node is before the start of the sequence or after the end of the sequence, no extrapolation is performed, and it is directly marked as invalid; if the interval between adjacent original samples exceeds three times the uniform sampling interval, it is considered a missing segment, and it is preferable to mark the entire segment as missing rather than interpolating across large intervals. For example, in a 0.2-second grid, if adjacent valid samples differ by 1.2 seconds, this segment should not be linearly filled in.

[0032] In detail, the sampling synchronization accuracy and abnormal sample rejection methods are as follows: At the acquisition end, the voltage and current channels use the same clock source or a periodic time synchronization mechanism to control the timestamp deviation between the two channels within one zero point of the unified sampling interval; before data is entered into the database, samples with out-of-range values, duplicate timestamp values, sudden reversed signs, and samples that clearly do not conform to the physical boundaries of the components are first rejected, and then time alignment and power calculation are performed. For example, if the voltage at a single point within a certain window suddenly jumps to three times the rated voltage, it should be marked as an abnormal bad point rather than directly participating in interpolation.

[0033] Preferably, the power time-series data is decomposed into a steady-state envelope component characterizing changes in environmental irradiance and a transient disturbance component characterizing the action of the control loop, including: The steady-state envelope component is calculated using the exponential moving average algorithm. The calculation formula is as follows: ; ; Calculate the transient disturbance components The calculation formula is as follows: ; In the formula, This represents the steady-state envelope value at the previous time point; This represents the power value at the current time point; Indicates the smoothing coefficient; This refers to the uniform sampling interval; This represents a slowly changing time constant, and the slowly changing time constant is greater than the reciprocal of the maximum power point tracking control period.

[0034] The steady-state envelope component, also known as the baseline sequence, is a slowly varying trend quantity obtained after smoothing power time-series data. It is used to characterize the overall impact of changes in environmental irradiance, temperature, and slow shading on output power.

[0035] The steady-state envelope value of the previous time node is a historical memory used to recursively calculate the steady-state envelope value of the current time node. It should be noted that in the recursive process of the exponential moving average, the steady-state envelope value of the current time node is obtained by weighted summation of the steady-state envelope value of the previous time node and the power value of the current time node.

[0036] The transient disturbance component is the residual sequence. In the subtraction calculation, the minuend is the power value at the current time point. In order to eliminate the interference of the causal phase delay caused by the exponential moving average on the residual waveform, before the subtraction operation, the obtained steady-state envelope component needs to be shifted in the historical direction on the time axis by the corresponding time constant delay. After phase alignment, the subtraction is performed. That is, the subtrahend used is the steady-state envelope value after time shift, so as to extract the rapidly changing component.

[0037] The smoothing coefficient is a dimensionless coefficient in the exponential moving average that assigns weights between the steady-state envelope value of the previous time point and the current power value. Its value is between zero and one. The closer the smoothing coefficient is to one, the stronger the memory of historical trends of the steady-state envelope component, which is more conducive to suppressing rapid control disturbances; the closer the smoothing coefficient is to zero, the faster the steady-state envelope component responds to the current sample.

[0038] The slow-changing time constant is a time-scale parameter used in exponential moving averages to characterize the response speed of the steady-state envelope component. It is preferably five to twenty times the maximum power point tracking control period; for example, when the control period is one second, the slow-changing time constant is preferably five to twenty seconds. The rationale for this value is that the slow-changing time constant should be significantly larger than the control disturbance rhythm so that the steady-state envelope component primarily tracks the slow changes in the environment rather than the control ripple.

[0039] The maximum power point tracking control cycle is the time interval between one disturbance observation, incremental admittance, or other maximum power point tracking update action performed by the inverter or power optimizer.

[0040] In detail, the initial value setting method for the steady-state envelope component is as follows: at the beginning of each independent analysis segment, the first effective power value is directly used as the initial steady-state envelope component, or the arithmetic mean of the first three to five effective power values ​​is taken as the initial baseline to reduce the influence of random noise at the starting point; for example, when the power of the first effective sample on the day is 280 watts and the average of the first five samples is 276 watts, it is preferable to use 276 watts as the initial steady-state envelope component.

[0041] In detail, the update start and stop conditions for exponential moving average are as follows: the steady-state envelope component is updated only when the power time-series data is within the effective daytime generation period, the sample is not missing, and it is not marked as an anomaly; when nighttime shutdown, component offline, inverter reset, or prolonged sample loss is detected, the steady-state envelope component should be reinitialized instead of using the previous historical value. For example, if sampling resumes after a ten-minute network outage in the afternoon, the baseline should be re-established using the first effective sample after the resumption.

[0042] In detail, the handling of transient disturbance components in low-irradiance and boundary regions is as follows: When the component power is lower than 5% to 10% of the rated power, or the current is close to zero and the spectrum analysis loses its practical significance, the corresponding time period should be marked as an invalid segment and not included in the subsequent main frequency search and sideband energy calculation; for boundary samples generated at the beginning and end of the sequence due to insufficient initialization, they can be retained but should be avoided as the center of the complete analysis window when dividing the window. For example, the weak leakage current stage before early morning startup should not be included in fault determination.

[0043] Preferably, performing spectral characteristic analysis on the transient disturbance component to lock the maximum power point tracking oscillation frequency within the current time window includes: Calculate the current time window using the short-time Fourier transform algorithm power spectral density function The calculation formula is as follows: ; ; Within the preset frequency search range Internally lock the maximum power point tracking oscillation frequency The calculation formula is as follows: ; In the formula, This represents the sequence segment of the transient perturbation component within the current time window; Use the starting index of the window; The length of the window; For window functions; Frequency index; This is the result of the short-time Fourier transform; , These are the lower and upper frequency limits of the frequency search range, respectively.

[0044] The current time window is the sliding analysis window number used when performing local time-frequency analysis on power time-series data or transient disturbance components.

[0045] The power spectral density function is the energy distribution function of transient disturbance components on the frequency axis within the current time window, used to describe the strength distribution of different frequency components within this window.

[0046] The short-time Fourier transform result is a complex spectrum obtained by weighting the transient disturbance components within the current time window through a window function and projecting them onto the frequency axis. It is used to simultaneously characterize the amplitude and phase information of each frequency component.

[0047] The window start index is the starting sample position number of the current time window in the unified time grid or transient disturbance component sequence.

[0048] The window length is the number of discrete sample points contained in each time window. It is preferably between 256 and 1024 points, or a large number of zero-padding operations are performed at the end of the sequence in a shorter time window to increase the number of frequency domain interpolation points. The physical frequency resolution value corresponding to the window time span must be strictly less than the lower limit of the sideband offset frequency, so as to cover multiple perturbation periods and ensure that the discrete spectrum can truly distinguish small offsets.

[0049] A window function is a sequence of weights applied to samples within the current time window to reduce spectral leakage caused by truncation. The Hanning window, for example, strikes a good balance between main lobe width and sidelobe suppression, making it suitable for extracting the maximum power point to track the main frequency and its nearby sideband energy.

[0050] The frequency index is the frequency coordinate on the spectrum of the current time window.

[0051] The frequency search range is a preset frequency band that limits the search range of the maximum power point tracking oscillation frequency. It is preferably 0.05 Hz to 2 Hz; when the inverter control is faster, it can be extended to 5 Hz. This range should cover the possible range of the main frequency of the equipment control disturbances, while avoiding low-frequency environmental envelopes and high-frequency measurement noise as much as possible.

[0052] The lower limit frequency is the starting frequency of the frequency search interval, used to eliminate interference from excessively low frequency components on the maximum power point tracking oscillation frequency search. It is preferably 0.05 Hz or not less than twice the main peak frequency of the low-frequency environmental envelope of the day, as the maximum power point tracking perturbation generally changes faster than the slow variation in environmental irradiance.

[0053] The upper limit frequency is the termination frequency of the frequency search interval, used to eliminate false peaks caused by high-frequency noise, residual switching ripple, or communication jitter. It is preferably two hertz; for devices with faster update speeds, it can be increased to five hertz. This is to cover the upper limit of the control disturbance's main frequency while avoiding misidentifying non-target high-frequency components as the main frequency.

[0054] The maximum power point tracking oscillation frequency is the dominant frequency at the position of maximum amplitude of the power spectral density function within the frequency search interval of the current time window. It is used to characterize the dominant frequency rhythm of the current control loop applying disturbance to the component's operating point.

[0055] In detail, the current time window division method and sliding step size are as follows: taking transient perturbation components as input, adjacent samples are divided window by window according to a fixed window length, and the window number increases sequentially according to time; the overlap step size between adjacent windows is preferably half the window length to balance time continuity and computational efficiency. For example, when the window length is 128 points, the sliding step size can be 64 points.

[0056] In detail, the selection method for window length and window function is as follows: First, estimate the number of samples corresponding to each main frequency cycle based on the uniform sampling interval and the expected main frequency range, and then ensure that a window covers at least two to five main frequency cycles; on this basis, the Hanning window is preferred as the window function. For example, when the uniform sampling interval is 0.2 seconds and the main frequency is about 0.5 Hz, each cycle has about 5 points, so the window length can be 64 points or 128 points.

[0057] In detail, the normalization method and frequency resolution of the power spectral density function are as follows: First, the sampling frequency is defined as 1 divided by the uniform sampling interval, and then the discrete frequency resolution is obtained according to the window length. After obtaining the absolute square of the time-frequency transformation result as an intermediate basic value, it must be further divided by the product of the sampling frequency and the energy normalization factor of the window function to complete the final power spectral density function calculation, so that the physical dimension corresponds to the power squared per unit frequency. For example, when the uniform sampling interval is 0.2 seconds and the window length is 128 points, the frequency resolution is approximately 0.039 Hz.

[0058] In detail, the frequency search interval is set as follows: first read the maximum power point tracking update frequency range in the inverter or power optimizer control strategy description as the initial search interval, then use a period of health history data to calibrate the common distribution position of the main peak, and finally leave a margin of one to two frequency resolution units on both sides; if there is no equipment description, the interval can be back-calculated by statistical analysis of the main peak of the power spectrum during the stable power generation period during the day.

[0059] In detail, the rules for determining the main frequency in the case of main peak competition are as follows: When there are multiple local peaks with similar amplitudes within the frequency search interval, the peak closest to the main frequency locked in the previous window is selected first; if the current window is the first window, the peak with the largest amplitude and closer to the device's nominal control frequency is selected first; if the candidate peaks still cannot be distinguished, the median smoothed result of three adjacent windows can be used as the final main frequency. For example, when two peaks of 0.4 Hz and 0.44 Hz coexist, the one closer to 0.42 Hz in the previous window can be retained first.

[0060] Preferably, extracting the energy spectrum of the nonlinear response sidebands around the tracking oscillation frequency of the maximum power point and its harmonics includes: For harmonic order Calculate the energy of the center frequency band Energy on the left side and the right side carries energy The calculation formula is as follows: ; ; ; In the formula, This indicates the maximum power point tracking oscillation frequency; This represents half of the central bandwidth parameter; , These represent the lower limit of the sideband offset frequency and the upper limit of the sideband offset frequency, respectively. This represents half of the sideband width parameter; This represents the power spectral density function.

[0061] The nonlinear response sideband energy spectrum is a set of features formed by the combination of the center band energy, the left band energy, and the right band energy according to the harmonic order around the tracking oscillation frequency of the maximum power point and its harmonic center frequency. It is used to characterize the nonlinear spectral diffusion characteristics of the component under the coupling effect of control disturbance and slow-changing environment.

[0062] Harmonic order is an order index that assigns integer multiples of the center frequency when the maximum power point tracking oscillation frequency is taken as the fundamental frequency. The higher the harmonic order, the more likely it is to contain more detailed nonlinear modulation information, but it is also more susceptible to noise and spectral leakage.

[0063] The center frequency, for a given harmonic order, is the frequency position obtained by multiplying the maximum power point tracking oscillation frequency by an integer multiple. It serves as a common reference point for the center frequency band, the left-side frequency band, and the right-side frequency band. The center frequency reflects the theoretical location of the dominant frequency of the control disturbance and its harmonics in the frequency spectrum.

[0064] The center bandwidth parameter is the total bandwidth parameter used when defining the center frequency band around the center frequency. It is preferably the bandwidth corresponding to two to four discrete frequency resolution units, or 0.05 to 0.2 times the current maximum power point tracking oscillation frequency. The center bandwidth must cover the natural drift of the main peak in adjacent windows, but it should not be so wide as to swallow up the energy of neighboring sidebands.

[0065] Half of the center bandwidth parameter is the half-bandwidth used when constructing the upper and lower boundaries of the center frequency band, which is used to extend the center frequency into a left-right symmetrical integration interval.

[0066] The center frequency band is a frequency integration interval centered on the center frequency and formed by offset half-bandwidths above and below it. It is used to statistically analyze the total energy within the band containing the main frequency and its harmonic peaks.

[0067] The center band energy is the energy value obtained by integrating the power spectral density function over the frequency within the center band, and is used to characterize the main peak energy intensity near the center frequency of the corresponding order.

[0068] The lower limit of the sideband offset frequency is the small offset of the sideband from the center frequency when constructing the sideband relative to the center frequency. Its value must be strictly greater than the frequency resolution value of the current time window, preferably 0.05 Hz to 0.1 Hz. This range can usually avoid the tail of the central main peak, while retaining the first-layer sideband diffusion information caused by slow-varying environmental modulation.

[0069] The upper limit of the sideband offset frequency is the maximum offset of the sideband from the center frequency when constructing the sideband relative to the center frequency. It is preferably between 0.05 Hz and 0.2 Hz, and should be greater than the lower limit of the sideband offset frequency. This range typically covers the main sideband diffusion region caused by blocking envelopes, bypass switching, and local nonlinear coupling.

[0070] The sideband width parameter is used to determine the total width of the left and right sideband frequency ranges. It is preferably the bandwidth corresponding to two to four discrete frequency resolution units, or between 0.02 Hz and 0.1 Hz. The sideband width must be sufficient to cover sideband drift while avoiding the incorporation of irrelevant spectral regions into the sideband energy.

[0071] Half of the sideband width parameter is the halfband width used when constructing the upper and lower boundaries of the left and right sidebands, which is used to form an integral interval of finite width over the offset interval defined by the upper and lower offset limits.

[0072] The left-side band range is an integral interval located to the left of the center frequency, defined by the lower limit of the side-band offset frequency, the upper limit of the side-band offset frequency, and the half-bandwidth. It is used to statistically analyze the energy spread in the sideband below the center frequency. This range can characterize the energy distribution of nonlinear modulation in the negative offset direction.

[0073] The right-side band range is an integral interval located to the right of the center frequency, defined by the lower limit of the side-band offset frequency, the upper limit of the side-band offset frequency, and the half-bandwidth. It is used to statistically analyze the energy spread in the sideband above the center frequency. This range can characterize the energy distribution of nonlinear modulation in the positive offset direction.

[0074] The left-side band energy is the energy value obtained by integrating the power spectral density function over the frequency range of the left-side band. It is used to quantify the degree of sideband diffusion of the corresponding harmonic to the left of the center frequency. The larger this parameter is, the more obvious the spectral diffusion caused by environmental envelope and structural nonlinear coupling is.

[0075] The right-side band energy is the energy value obtained by integrating the power spectral density function over the frequency range of the right-side band. It is used to quantify the degree of sideband diffusion of the corresponding harmonic order to the right of the center frequency. The left-side band energy and the right-side band energy together determine the energy distribution pattern of this harmonic order.

[0076] In detail, the range of harmonic orders is as follows: starting from the first order and increasing, with the total number of harmonic orders as the upper limit; in engineering implementation, the first to third orders are preferred, and can be extended to the fifth order when the noise level is low. If the frequency band corresponding to higher orders is close to the Nyquist upper limit or the signal-to-noise ratio drops significantly, then it should not be included in the analysis.

[0077] In detail, the setting methods for the center bandwidth parameter, the lower limit of the sideband offset frequency, the upper limit of the sideband offset frequency, and the sideband width parameter are as follows: First, set the minimum resolvable bandwidth based on the discrete frequency resolution of the power spectral density function, and then calibrate it in conjunction with the main peak width and sideband diffusion range in the healthy sample. When assigning parameters, strict spatial isolation constraints must be applied, that is, the lower limit of the sideband offset frequency minus half of the sideband width parameter must be strictly greater than half of the center bandwidth parameter, so as to completely eliminate the physical overlap or frequency inversion between the sideband integration interval and the center frequency band; the center bandwidth preferably covers the main peak and its slight drift region, the sideband offset region preferably covers the main diffusion band formed by slow envelope modulation, and the sideband width preferably covers one or two frequency resolution units near each side diffusion peak.

[0078] In detail, the handling methods for frequency band overlap, boundary crossing, and adjacent harmonic interference are as follows: When the left band, right band, or center band overlaps with the vicinity of zero frequency, the upper limit of Nyquist, or adjacent harmonic bands, boundary clipping should be performed first, and then the energy should be calculated based on the remaining effective frequency band; if the effective bandwidth after clipping is less than one frequency resolution unit, the harmonic order should be marked as invalid and not participate in the bypass coupling characteristic index calculation of the current window.

[0079] In detail, the discrete implementation of the frequency band integral is as follows: the continuous frequency boundary is rounded inward or outward to the nearest valid discrete frequency point index position, and rewritten as the sum of the power spectral density functions of discrete frequency points within the target frequency band multiplied by the frequency resolution. This avoids the energy calculation at the boundary position being missed, thus ensuring consistency with the meaning of the continuous integral. For example, if a center frequency band covers four discrete frequency points, namely three, four, five, and four, and the frequency resolution is 0.04 Hz, then the energy of the center frequency band is the sum of the above four values ​​multiplied by 0.04.

[0080] In detail, the output organization of the nonlinear response sideband energy spectrum is as follows: the center band energy, left band energy, and right band energy of each harmonic order are arranged from low to high to form a fixed-order feature vector or dictionary structure, so as to calculate the energy distribution probability, sideband diffusion entropy value, and bypass coupling characteristic index in the subsequent calculation; for example, when the total number of harmonic orders is three, it can be output in the order of first-order center, first-order left, first-order right, second-order center, second-order left, second-order right, third-order center, third-order left, and third-order right.

[0081] Preferably, the bypass coupling characteristic index, which characterizes the degree of nonlinear operation of the bypass diode, is calculated based on the nonlinear response sideband energy spectrum, including: For harmonic order Calculate the energy distribution probability The calculation formula is as follows: ; Calculate the sideband diffusion entropy value for this harmonic order. The calculation formula is as follows: ; Calculate the bypass coupling characteristic index The calculation formula is as follows: ; In the formula, , , These are the center frequency band energy, right side band energy, and left side band energy, respectively. For the first Weighting coefficients for first harmonics; This represents the total number of harmonic orders.

[0082] Total energy is the sum of the energy in the center band, the left band, and the right band for the same harmonic order. It is used to convert the absolute energy of different frequency bands into comparable relative distribution results. The larger the total energy, the stronger the overall spectral activity of that harmonic order.

[0083] The energy distribution probability is a dimensionless probability value obtained by representing the proportion of energy in a certain frequency band to the total energy under a given energy category for a corresponding harmonic order. It is used to describe the relative energy distribution pattern between the center band and the left and right side bands. The value of this parameter ranges between zero and one, and the sum of the probabilities of the three categories under the same harmonic order is one.

[0084] The energy category index is a discrete identifier that distinguishes three types of energy locations: the center frequency band, the left side band, and the right side band. Zero represents the center frequency band, a negative sign represents the left side band, and a positive sign represents the right side band.

[0085] The sideband diffusion entropy is a dimensionless quantity obtained by calculating the entropy of the energy distribution probability of the corresponding harmonic order. It is used to characterize the degree of dispersion and diffusion of energy distribution in the center band and the left and right sidebands. The larger the entropy value, the more dispersed the energy is in the sidebands, which usually corresponds to stronger nonlinear coupling or more complex bypass switching behavior.

[0086] The bypass coupling characteristic index is a comprehensive index obtained by weighting and summing the sideband diffusion entropy values ​​of multiple harmonic orders according to weighting coefficients. It is used to characterize the degree of nonlinear action of the bypass diode and the strength of its influence on the spectral structure.

[0087] The weighting coefficients for corresponding harmonic orders are dimensionless weights used to control the contribution of different harmonic orders to the bypass coupling characteristic index. Preferably, the weights are normalized and decrease in order; for example, when the total number of harmonic orders is three, weights of 0.5, 0.3, and 0.2 are preferred. Lower-order harmonics are generally more stable and have a higher signal-to-noise ratio, therefore they should be given greater weight.

[0088] The total number of harmonic orders is the upper limit of the harmonic orders used in the calculation of the bypass coupling characteristic index. Ideally, it should be three to five. In distributed photovoltaic module operation data, excessively high orders are often susceptible to noise, spectral leakage, and equipment variations, while three to five orders are often sufficient to cover the main nonlinear diffusion information.

[0089] In detail, the protection method for energy allocation probability when the total energy is close to zero or the signal-to-noise ratio is as follows: Before calculating the energy allocation probability, the estimated value of the background white noise floor of each frequency band is first uniformly deducted from the energy of each frequency band; then it is determined whether the sum of the energy of the center frequency band, the energy of the left band, and the energy of the right band is less than the preset minimum energy threshold; if it is less than or indicates that the current distribution is completely dominated by environmental noise and exhibits false uniformity, then this step is skipped directly and its contribution is removed during subsequent weight normalization to avoid false alarms caused by white noise traps and false high entropy values.

[0090] In detail, the calculation method for the side-band diffusion entropy value when the probability is zero is as follows: In the entropy calculation, the convention of treating zero as zero when multiplied by the logarithm is adopted, or the lower limit of the probability of each energy distribution is truncated to a very small positive number before the calculation to avoid infinite or undefined results in the numerical calculation; for example, the probability less than 10 to the power of negative 6 can be uniformly replaced with 10 to the power of negative 6 before calculating the logarithm.

[0091] In detail, the setting and normalization method of the weighting coefficients is as follows: first, initial weights are given based on the distinguishing ability of each harmonic order in the health data and fault data; then, all weights are normalized by summing them to one. If there are no prior statistical results, empirical weights that decrease with harmonic order can be used. In implementation, it is preferable to make the weight of lower-order harmonics greater than that of higher-order harmonics to improve robustness.

[0092] In detail, the method for determining the total number of harmonic orders is as follows: first, determine the highest observable harmonic order based on the unified sampling interval and the current dominant frequency, and then select the final value by combining the signal-to-noise ratio, frequency band overlap and health fault differentiation capability of each harmonic; generally, in the case of distributed photovoltaic modules, three to five are preferred. If the higher-order harmonics are too weak or the frequency band is out of bounds, the number should be reduced appropriately.

[0093] In detail, the numerical normalization method for the bypass coupling characteristic index is as follows: after calculating the weighted sum, it can be further divided by the theoretical maximum entropy value or linearly normalized using the minimum and maximum values ​​on the training set, so that the bypass coupling characteristic index falls into a stable and comparable range; if the theoretical maximum entropy value is used for normalization, in order to eliminate the proportional differences caused by different logarithmic bases, the three-class maximum entropy value under the logarithmic base used in the calculation should be strictly used as the normalization scaling benchmark, so as to ensure that the values ​​between different devices and different dates are comparable.

[0094] Preferably, a fault prediction model is constructed, comprising a long-time-series evolution extraction channel, a short-time-series waveform extraction channel, and an adaptive gating unit, including: Build the current time window state vector The formula is constructed as follows: ; Generate the channel attention coefficient The calculation formula is as follows: Calculate the fusion feature vector The calculation formula is as follows: ; In the formula, The bypass coupling characteristic index; , They are respectively The first-order difference and the second-order difference; The percentage of energy in the side band; The maximum power point is used to track the oscillation frequency; The centroid of the slow envelope frequency; Use the Sigmoid activation function; , , These are trainable parameters; The frequency domain structure features; This refers to the time-domain response characteristics.

[0095] The long-term evolution extraction channel uses a sequence of state vectors from multiple consecutive time windows as input to extract a feature extraction subnetwork that captures cross-window evolution patterns and frequency domain structural changes. This channel focuses on capturing the evolutionary trajectory of the side-channel coupling feature index, sideband energy proportion, and main frequency changes over a longer time scale.

[0096] The short-time waveform extraction channel uses the transient perturbation component sequence within the current time window as input to extract the feature extraction subnetwork of fine-grained time-domain response morphology within the current window.

[0097] The adaptive gating unit is a control unit that generates channel attention coefficients based on the state vector and dynamically weights and fuses frequency domain structural features and time domain response features.

[0098] The state vector is a feature vector formed by sequentially splicing the bypass coupling characteristic index, the first difference of the bypass coupling characteristic index, the second difference of the bypass coupling characteristic index, the sideband energy ratio, the maximum power point tracking oscillation frequency, and the centroid of the slow envelope frequency in a fixed order. Each component occupies an independent and different dimensional position in the vector.

[0099] The first-order difference is the change in the bypass coupling characteristic index between two adjacent time windows, used to characterize the increasing or decreasing trend of bypass coupling strength on the window-level time scale.

[0100] The second-order difference is the change in the first-order difference between adjacent time windows, used to characterize the acceleration, deceleration, or inflection point behavior of the bypass coupling characteristic exponent.

[0101] Sideband energy percentage is a dimensionless feature used to characterize the proportion of sideband energy in the total energy of the center band and the left and right sidebands.

[0102] The slow envelope frequency centroid is a representative frequency obtained by energy weighting the steady-state envelope component or its low-frequency spectrum within a preset low-frequency range. It is used to characterize the rhythm center of slow-varying and regular shading changes in environmental irradiance.

[0103] A state vector sequence is a sequence formed by arranging state vectors corresponding to multiple consecutive time windows in chronological order. It is used to provide cross-window evolution information for long-term evolution extraction channels. The longer the state vector sequence, the easier it is for the model to capture slowly accumulating fault symptoms, but it also increases the complexity of training and inference.

[0104] Frequency domain structural features are feature representations output after the long-time evolution extraction channel models the state vector sequence, and are used to summarize the frequency domain structural evolution law across the window.

[0105] The time-domain response feature is the feature representation output by the short-time waveform extraction channel after modeling the transient disturbance component sequence within the current time window, which is used to summarize the instantaneous morphological information of the waveform within the window.

[0106] The channel attention coefficient is a dimensionless weight vector between zero and one, generated by the adaptive gating unit based on the state vector. It is used to control the respective weights of frequency domain structural features and time domain response features when fused in each specific feature dimension. The closer the channel attention coefficient is to one, the more the current window relies on frequency domain structural features; the closer the channel attention coefficient is to zero, the more the current window relies on time domain response features.

[0107] The fused feature vector is a comprehensive feature representation obtained by adding the first weighted result and the second weighted result, which is used to simultaneously carry frequency domain structure information and time domain response information.

[0108] The trainable weight parameter is the weight matrix used in the adaptive gating unit to perform linear transformations on the state vector. It is used to learn the multi-dimensional influence of each state component on the channel attention coefficient, ensuring that the output channel attention coefficient maintains the same dimension as the feature to be fused, thus enabling fine-grained element-wise dynamic weighting during fusion. During training, this parameter is continuously updated with feedback from sample labels and the loss function to form an adaptive weighting pattern for different operating conditions.

[0109] Trainable bias parameters are bias parameters used in conjunction with trainable weight parameters in adaptive gating units to adjust the overall offset level of channel attention coefficients.

[0110] Dimensional mapping trainable parameters are linear transformation parameters that map time-domain response features to a space with the same dimensionality as frequency-domain structural features.

[0111] In detail, the network structure and input length of the long-term evolution extraction channel are as follows: preferably, a recurrent structure or a temporal convolutional structure that can handle sequential dependencies is adopted, and the input is a sequence of state vectors composed of multiple consecutive time windows, with the sequence length preferably being eight to thirty-two; for example, a two-layer gated recurrent structure can be adopted, with each layer having a hidden dimension of sixty-four, thereby balancing expressive power and edge deployment cost.

[0112] In detail, the network structure and input preprocessing method for the short-time waveform extraction channel are as follows: a one-dimensional convolutional feature extraction structure or a shallow temporal convolutional residual structure is preferred, and the input is the transient perturbation component sequence within the current time window; before input, mean removal, amplitude normalization and abnormal peak pruning should be performed to reduce the impact of dimensional differences on training.

[0113] In detail, the calculation methods for the first-order and second-order differences are as follows: In order to suppress the sharp amplification of high-frequency random noise by the difference operation, the bypass coupling characteristic index sequence composed of multiple time windows is first smoothed by the moving average. Then, the first-order difference is preferably defined as the smoothed bypass coupling characteristic index of the current window minus the smoothed value of the previous window, and the second-order difference is defined as the first-order difference of the current window minus the first-order difference of the previous window.

[0114] In detail, the sideband energy ratio is the sum of all left and right sideband energies involved in the calculation divided by the sum of the center band energy and the left and right sideband energies.

[0115] In detail, the calculation method for the slow envelope frequency centroid is as follows: First, calculate the low-frequency power spectrum of the steady-state envelope component within a preset low-frequency range. Then, use frequency as the weight and spectral energy as the weighting coefficient to obtain the frequency centroid. If a discrete implementation is used, the discrete frequency points within the low-frequency band should be multiplied by their corresponding energies, summed, and then divided by the total low-frequency energy. The slow envelope frequency centroid obtained in this way can represent the slow-changing environmental rhythm within that window on that day.

[0116] In detail, the linear transformation implementation of the adaptive gating unit is as follows: In order to eliminate the influence of the inconsistency of the physical units of each feature, the terms of the state vector are first standardized into dimensionless form according to the mean and standard deviation of the training set. Then, the dimension mapping weight matrix is ​​used as a trainable weight parameter, which together with the trainable bias parameter completes the multi-dimensional linear mapping. Finally, the channel attention coefficients between zero and one with corresponding dimensions are obtained through the monotonically compressed activation function.

[0117] In detail, the dimensional alignment method between frequency domain structural features and time domain response features is as follows: before fusion, it is agreed that the output dimensions of the two are the same, or the time domain response features are mapped to the same dimension as the frequency domain structural features using dimensional mapping trainable parameters; for example, when the frequency domain structural features output 64 dimensions, the dimensional mapping parameters can be used to map the original time domain features to 64 dimensions, and then added to the frequency domain structural features bit by bit.

[0118] In detail, the training constraints for the fused feature vectors are as follows: during the model training phase, the network containing the fused feature vectors is subjected to binary cross-entropy loss, and combined with class balancing sampling, weight decay, inactivation regularization and early stopping strategies to avoid overfitting caused by the scarcity of faulty samples; if the on-site sample classes are extremely unbalanced, abnormal samples should be given higher loss weights.

[0119] Preferably, the fused features are mapped to component health state probabilities, and a fault determination result is output when the component health state probability exceeds a preset threshold, including: Calculate the current time window Window-level anomaly probability The calculation formula is as follows: ; Calculate the daily anomaly score of the component The calculation formula is as follows: ; Based on the daily anomaly score of the component Output the fault determination result The decision logic is as follows: ; In the formula, The fused feature vector; , These are the weight and bias parameters for a linear classifier; Use the Sigmoid activation function; This represents the total number of valid time windows for the day. This is the set of valid time windows for the current day; The preset threshold is defined as follows.

[0120] The window-level anomaly probability is the anomaly probability value corresponding to the current time window obtained by mapping the fused feature vector through a linear classifier and compressing it through a non-linear activation function. It is used to characterize the probability that a component will behave abnormally within the window. The value of this parameter ranges between zero and one.

[0121] All valid time windows for the day are the set of all analysis windows retained after validity screening of the photovoltaic modules to be tested within a certain natural day, and are used to participate in the statistical analysis of the module's abnormality score for the day.

[0122] The total number of valid time windows on a given day is the number of windows that pass the validity screening within a calendar day and are ultimately included in the daily statistics. This number is used to calculate the arithmetic average of the window-level anomaly probabilities.

[0123] The set of valid time windows for a given day is the set of all valid window numbers or valid window indices within a calendar day.

[0124] The component's daily anomaly score, or component health probability, is the arithmetic mean of the window-level anomaly probabilities across all valid time windows for that day. It characterizes the overall probability level of the component being in an abnormal state on that calendar day. This parameter ranges between zero and one; a higher value indicates a greater likelihood of the component experiencing a fault or significant anomaly.

[0125] The preset threshold is used to determine whether a component's daily anomaly score is normal or an anomaly. It is preferably between 0.6 and 0.8 to comprehensively balance the risk of missed reports and the cost of false alarms in the verification data. When the operational scenario emphasizes early warning, it can be appropriately lowered; when the emphasis is on reducing false alarms, it can be appropriately raised.

[0126] The fault determination result is a binary determination output obtained by comparing the component's daily anomaly score with a preset threshold. It is used to clearly indicate to the operation and maintenance system whether the component should be marked as abnormal or normal on the current date.

[0127] A linear classifier is a prediction unit that performs a linear mapping on the fused feature vector and outputs a classification value. It is used to compress multi-dimensional fused features into a single anomaly score. By connecting a non-linear activation function after the linear classifier, the unbounded discrimination value can be converted into a probability quantity in the range of zero to one.

[0128] The weight parameters of a linear classifier are trainable weight vectors that act on each component of the fused feature vector, used to learn the importance of different fused features for anomaly detection. During training, these parameters are continuously adjusted based on the fault label to improve the ability of window-level anomaly probabilities to distinguish real faults.

[0129] The bias parameter of a linear classifier is a trainable bias term that works in conjunction with the weight parameters to adjust the overall discriminant surface's shift position. This parameter allows the model to maintain an appropriate classification threshold even when the overall value of the fused feature vector is low or high.

[0130] In detail, the training method and labeling rules for the linear classifier are as follows: the fused feature vector is used as input, and the component health label or window-level manual verification label corresponding to the date to which the window belongs is used as the supervision label. The weight parameters and bias parameters of the linear classifier are trained by supervised learning. The label should at least distinguish between normal and abnormal categories. If it is necessary to support multiple fault types, a binary anomaly detector can be trained first and then expanded into a multi-classifier.

[0131] In detail, the rules for determining the effective time window are as follows: the normal and stable operation time window and the abnormal fluctuation time window are all recorded as effective time windows, where the power is greater than the minimum power threshold, the data missing ratio is lower than the preset upper limit, and the main frequency search is successful, in order to maintain the objectivity and integrity of the statistical base and prevent statistical evaluation bias; nighttime windows, component offline windows, inverter start-stop transition windows, and windows with severe sample missing should all be removed; for example, if the number of missing samples in a window exceeds 20%, it can be directly judged as invalid.

[0132] In detail, the method for determining the preset threshold is as follows: different thresholds are validated on historical labeled data, and the optimal threshold is selected after comprehensively comparing the false negative rate, false positive rate, and operation and maintenance cost; preferably, candidate thresholds are obtained first through grid search on the development set, and then the final threshold is confirmed on the independent validation set. If the system focuses on early warning, the threshold can be set appropriately low; if the focus is on reducing false positives, the threshold can be set appropriately high.

[0133] In detail, the method for determining when the number of valid time windows is insufficient on a given day is as follows: First, set a minimum threshold for the number of valid windows. When the total number of valid windows on a given day is less than this threshold, do not directly output "normal" or "abnormal". Instead, output "data insufficient" and wait for more windows to be added. For example, if at least twelve valid windows are required per day, but only five valid windows are obtained on a given day, the daily fault conclusion should be suspended to avoid statistical instability.

[0134] In detail, the timing of the fault determination results output and the rules for continuous alarms are as follows: the determination results for the day are output after the statistics of all valid time windows for the day are completed and the abnormal score of the component for the day is obtained. If it is used for actual operation and maintenance alarms, it is preferable to add a continuous day confirmation mechanism. For example, it is upgraded to a strong alarm only after exceeding the preset threshold for two or three consecutive days, while exceeding the threshold on a single day can be used as an early warning to reduce false alarms caused by occasional disturbances.

[0135] like Figure 2 As shown, Figure 2 The diagram illustrates an end-to-end deployment architecture for distributed photovoltaic (PV) module fault prediction / early warning: The PV module in the upper left outputs DC power, which is connected to the grid via an inverter (connected to the grid tower). Simultaneously, signals are drawn from the module side and fed into the DC measurement unit in the lower left corner for voltage (V) and current (A) acquisition. The acquired DC operating data is sent to the edge gateway in the middle (responsible for aggregation, communication, and uploading), and then uploaded to the cloud analysis platform on the right. Internally, the cloud platform sequentially performs time-series processing (such as time alignment, resampling, power sequence construction and denoising), time-frequency feature extraction (such as time-frequency analysis of power disturbances and tracking the main frequency), and machine learning prediction (using extracted features to assess / classify fault probability). The final prediction results are output in two ways: first, a fault warning is pushed to maintenance personnel in the upper right corner (mobile phone notification); second, a visualization is displayed in the monitoring module in the lower right corner (alarm status and trend chart).

[0136] 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 machine learning-based method for predicting faults in distributed photovoltaic modules, characterized in that, include: Acquire DC voltage and DC current sampling data of photovoltaic modules, construct power time series data, and decompose the power time series data into steady-state envelope components characterizing changes in environmental irradiance and transient disturbance components characterizing the action of control loops; Spectral feature analysis is performed on the transient disturbance component to lock the maximum power point tracking oscillation frequency within the current time window. The nonlinear response sideband energy spectrum around the maximum power point tracking oscillation frequency and its harmonics is extracted, and the bypass coupling characteristic index characterizing the degree of nonlinear action of the bypass diode is calculated based on the nonlinear response sideband energy spectrum. A fault prediction model is constructed, comprising a long-time-series evolution extraction channel, a short-time-series waveform extraction channel, and an adaptive gating unit. The adaptive gating unit is configured to receive a state vector composed of the bypass coupling feature index and generate channel attention coefficients accordingly. The channel attention coefficients are then used to dynamically and weightedly fuse the frequency domain structure features output by the long-time-series evolution extraction channel with the time domain response features output by the short-time-series waveform extraction channel. The fused features are mapped to component health status probabilities, and a fault determination result is output when the component health status probability exceeds a preset threshold.

2. The machine learning-based distributed photovoltaic module fault prediction method according to claim 1, characterized in that, Acquire DC voltage and DC current sampling data of photovoltaic modules to construct power time-series data, including: A fixed, uniform sampling interval is set, and a uniform time grid is constructed based on the uniform sampling interval. The original DC voltage sequence and the original DC current sequence are mapped to each time node of the uniform time grid using a linear interpolation algorithm to obtain aligned DC voltage and DC current values. The aligned DC voltage and DC current values ​​at the same time node are multiplied to obtain the power value at that time node. The power time series data is composed of the power values ​​at all time nodes.

3. The machine learning-based distributed photovoltaic module fault prediction method according to claim 2, characterized in that, The power time-series data is decomposed into a steady-state envelope component characterizing changes in environmental irradiance and a transient disturbance component characterizing the action of the control loop, including: The power time series data is smoothed using an exponential moving average algorithm to calculate a baseline sequence reflecting the slow variation trend of power. This baseline sequence is used as the steady-state envelope component. The difference between the power time series data and the steady-state envelope component at corresponding time nodes is calculated to obtain a residual sequence after removing the slow variation trend. This residual sequence is used as the transient disturbance component.

4. The machine learning-based distributed photovoltaic module fault prediction method according to claim 3, characterized in that, Perform spectral characteristic analysis on the transient disturbance components, and lock the maximum power point tracking oscillation frequency within the current time window, including: The transient disturbance component is subjected to time-frequency transformation using a short-time Fourier transform algorithm to calculate the power spectral density function of the current time window; within a preset frequency search interval, the frequency point corresponding to the maximum amplitude of the power spectral density function is searched, and this frequency point is used as the maximum power point tracking oscillation frequency within the current time window.

5. The machine learning-based distributed photovoltaic module fault prediction method according to claim 4, characterized in that, Extracting the nonlinear response sideband energy spectrum around the tracking oscillation frequency of the maximum power point and its harmonics includes: For each set harmonic order, the center frequency is defined as an integer multiple of the maximum power point tracking oscillation frequency. Based on preset center bandwidth parameters, the center frequency band range around the center frequency is determined, and the power spectral density function is integrated within this center frequency band range to obtain the center frequency band energy. Based on preset sideband offset frequency lower limit, sideband offset frequency upper limit, and sideband width parameters, the left sideband frequency band range located to the left of the center frequency and the right sideband frequency band range located to the right of the center frequency are determined respectively. The power spectral density function is integrated within the left sideband frequency band range and the right sideband frequency band range respectively to obtain the left sideband energy and the right sideband energy. The center frequency band energy, left sideband energy, and right sideband energy corresponding to each harmonic order are combined to form the nonlinear response sideband energy spectrum.

6. The machine learning-based distributed photovoltaic module fault prediction method according to claim 5, characterized in that, The bypass coupling characteristic index, which characterizes the degree of nonlinear operation of the bypass diode, is calculated based on the nonlinear response sideband energy spectrum, including: For each harmonic order, the corresponding center band energy, left band energy, and right band energy are summed to obtain the total energy. The ratio of each part of the energy to the total energy is calculated to obtain the corresponding energy allocation probability. The sideband diffusion entropy value of the harmonic order is calculated based on the energy allocation probability. The sideband diffusion entropy values ​​of all harmonic orders are weighted and summed to obtain the bypass coupling characteristic index.

7. The machine learning-based distributed photovoltaic module fault prediction method according to claim 6, characterized in that, A fault prediction model is constructed, comprising a long-time-series evolution extraction channel, a short-time-series waveform extraction channel, and an adaptive gating unit, including: A state vector is constructed, comprising the bypass coupling characteristic index and its first-order difference, second-order difference, sideband energy ratio, maximum power point tracking oscillation frequency, and slow envelope frequency centroid. The state vector sequence of multiple consecutive time windows is used as input to the long-term evolution extraction channel to extract the frequency domain structural features. The transient disturbance component sequence within the current time window is used as input to the short-term waveform extraction channel to extract the time domain response features. The state vector is input to the adaptive gating unit, and after linear transformation and activation function processing, the channel attention coefficients with values ​​between zero and one are generated. The frequency domain structural features are weighted using the channel attention coefficients to obtain a first weighted result, and the time domain response features are weighted using the difference between one and the channel attention coefficients to obtain a second weighted result. The first weighted result and the second weighted result are added together to obtain a fused feature vector.

8. The machine learning-based distributed photovoltaic module fault prediction method according to claim 7, characterized in that, The fused features are mapped to component health state probabilities, and a fault determination result is output when the component health state probability exceeds a preset threshold, including: The fused feature vector is mapped using a linear classifier and processed by a nonlinear activation function to obtain the window-level anomaly probability of the current time window. The window-level anomaly probabilities of the photovoltaic module under test in all valid time windows of the day are obtained, and the arithmetic mean of the window-level anomaly probabilities of all valid time windows is calculated. This arithmetic mean is used as the module's anomaly score for the day. The module's anomaly score for the day is compared with a preset threshold as the module's health status probability: if the module's anomaly score for the day is greater than or equal to the preset threshold, the fault determination result is output as abnormal; if the module's anomaly score for the day is less than the preset threshold, the fault determination result is output as normal.

Images