A distributed photovoltaic system non-stationary signal feature extraction method and system

Abstract

The application discloses a distributed photovoltaic system non-stationary signal feature extraction method and system, relates to photovoltaic power generation, and comprises the following steps: acquiring original power signals of a distributed photovoltaic system in a preset time window and environment parameters which are time-synchronized with the original power signals; constructing a physical constraint function by using the environment parameters; performing empirical wavelet transform on the original power signals, determining a frequency band boundary by detecting local maximum values of a frequency spectrum, and decomposing the original power signals into n sub-signal components of different time scales according to the frequency band boundary; substituting each sub-signal component into the physical constraint function to calculate a physical consistency index of the i-th sub-signal component; based on the physical consistency index, setting a threshold, and dividing the n sub-signal components into a deterministic component set and a disturbance component set; and aiming at the fact that traditional VMD and other decomposition methods cannot distinguish components with different physical meanings in photovoltaic signals, leading to a lack of interpretability of the decomposition results, the application improves the interpretability.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power generation, and in particular to a method and system for extracting non-stationary signal features from distributed photovoltaic systems. Background Technology

[0002] With the advancement of global energy transition and the "dual-carbon" goal, distributed photovoltaic (PV) systems, as an important component of clean energy, are seeing increasing penetration in power systems. Distributed PV systems offer advantages such as flexible installation, local energy consumption, and reduced transmission losses, and are widely used in industrial and commercial rooftops, residential buildings, and agricultural facilities. However, the power generation of distributed PV systems is affected by various factors such as sunlight intensity, temperature, cloud cover, and equipment aging, exhibiting significant non-stationarity and stochastic characteristics, posing challenges to the safe and stable operation of the power grid. Therefore, accurate feature extraction of the power signals from distributed PV systems is crucial for fault diagnosis, power prediction, and optimized dispatching.

[0003] The power signal of a distributed photovoltaic (PV) system is a typical non-stationary signal, containing components across multiple time scales: deterministic components reflecting the variation in solar radiation, and random disturbances caused by cloud cover, equipment failures, etc. These different components have different physical origins and characteristics, requiring appropriate signal decomposition methods for separation and analysis. Traditional signal processing methods such as Fourier transform can only provide global frequency domain information and cannot capture the time-varying characteristics of the signal. While Empirical Mode Decomposition (EMD) is adaptive, it suffers from problems such as mode aliasing and endpoint effects. Variational Mode Decomposition (VMD) achieves signal decomposition by constructing a variational optimization problem, overcoming some of the shortcomings of EMD, but it is essentially still a purely mathematical optimization method.

[0004] Existing signal decomposition methods such as VMD have significant shortcomings when processing photovoltaic power signals: First, these methods only take a mathematical approach, achieving signal decomposition by optimizing the objective function, without considering the physical characteristics of the photovoltaic system and the influence of environmental factors, resulting in the lack of clear physical meaning for each modal component obtained from the decomposition; second, they cannot effectively distinguish between normal photovoltaic response components and abnormal disturbance components, making subsequent feature extraction lack specificity; third, the interpretability of the decomposition results is poor, making it difficult to establish the correspondence between the decomposed components and the actual physical process, thus limiting their application in fault diagnosis and power prediction.

[0005] In summary, the key technical challenge to improve the fault diagnosis and power prediction performance of distributed photovoltaic systems lies in how to combine the physical characteristics of photovoltaic systems to achieve physically meaningful signal decomposition and adopt differentiated feature extraction strategies for components with different physical properties. Summary of the Invention

[0006] Traditional decomposition methods such as VMD, which are purely based on mathematical optimization, cannot distinguish components with different physical meanings in photovoltaic signals, resulting in a lack of interpretability in the decomposition results. This application provides a method and system for extracting features of non-stationary signals in distributed photovoltaic systems. By introducing the concept of multi-scale decomposition with physical constraints, constructing physical constraint functions by combining environmental parameters and photovoltaic physical models, automatically classifying the decomposed signal components into deterministic and perturbation components using physical consistency indices, and adopting differentiated feature extraction strategies for components with different physical properties, the interpretability of the results is improved.

[0007] One aspect of this application provides a method for extracting non-stationary signal features from a distributed photovoltaic system, comprising: S1, acquiring the raw power signal of the distributed photovoltaic system within a preset time window. And environmental parameters synchronized with the original power signal P(t) in time, including the light intensity signal. and temperature signal S2, Constructing physical constraint functions using environmental parameters. S3, for the original power signal Perform empirical wavelet transform to obtain the original power signal. The spectrum; the frequency band boundary is determined by detecting local maxima in the spectrum, and the original power signal is then processed based on the frequency band boundary. Decomposed into n sub-signal components at different time scales S4, convert each sub-signal component Substitute into the physical constraint function Calculate the physical consistency index of the i-th sub-signal component. .

[0008] S5, based on the physical consistency index calculated in step 4 Set threshold Classify the n sub-signal components: when When, the i-th sub-signal component Classified as a set of deterministic components; when When, the i-th sub-signal component It is classified as a set of perturbation components.

[0009] S6, for each component in the deterministic component set obtained from step 5, extract its mean and linear trend slope as trend features, and extract its main frequency and spectral peak value as periodic features; for each component in the perturbation component set, extract its kurtosis and maximum rate of change as abrupt change features, and extract its variance and entropy value as random features; S7, arrange the trend features, periodic features, abrupt change features, and random features extracted in S6 in a preset order to construct a representation of the original power signal. Multiscale feature vectors.

[0010] Furthermore, physical constraint functions are constructed using environmental parameters. ,include: Where P is the power value, The rated power of the system, Standard light intensity, The average temperature coefficient, For standard test temperature, This is the illumination efficiency correction function. This refers to system availability.

[0011] Among them, physical constraint function This represents the deviation between the actual power value P and the theoretical power value. This function quantifies the degree of agreement between the signal component and the physical laws of photovoltaics by calculating the difference between the actual and theoretical power. When the signal component is close to zero, it indicates that the signal component conforms to the physical laws of photovoltaics and is a normal photovoltaic response; when... A larger value indicates that the signal component deviates from physical laws and may be caused by abnormal factors.

[0012] The illuminance efficiency correction function is used to correct for changes in photovoltaic (PV) conversion efficiency under different illuminance intensities. In practical applications, the conversion efficiency of a PV system is not constant but varies with illuminance intensity. Under low illuminance conditions, the conversion efficiency of PV cells decreases; under high illuminance conditions, the conversion efficiency may also decrease due to factors such as thermal effects. The function reflects this nonlinear relationship, ensuring the accuracy of the physical model under different lighting conditions.

[0013] System availability These are comprehensive parameters that reflect the actual operating status of a photovoltaic system.

[0014] Furthermore, the physical consistency index of the i-th sub-signal component is calculated. : The integration interval is the preset time window in step S1; the physical consistency index is... This index is used to quantitatively assess the degree of conformity between each sub-signal component and the physical laws of photovoltaics. It is obtained by calculating the squared integral of each sub-signal component under the physical constraint function. The value reflects the i-th sub-signal component The degree of deviation between the power value and the theoretical power value calculated based on the photovoltaic physics model. The smaller the value, the more it conforms to the physical laws of photovoltaics; The larger the value, the more the component deviates from the laws of physics.

[0015] By integrating over the entire time window, this index comprehensively considers the physical consistency of sub-signal components throughout the observation period, avoiding the influence of instantaneous biases. The physical consistency index is a quantitative basis for distinguishing between deterministic and perturbative components. When the threshold is reached, this component is considered to primarily reflect the normal photovoltaic response and is classified as a deterministic component; when... At that time, it was considered that the component contained a large number of abnormal disturbances and was classified as a disturbance component.

[0016] Furthermore, set a threshold. The formula is as follows: Where k is the adjustment coefficient, and its value ranges from 0.5 to 2.0. Physical consistency index The mean, Physical consistency index Standard deviation;

[0017] Furthermore, S3, for the original power signal Perform empirical wavelet transform to obtain the original power signal. The spectrum; the frequency band boundary is determined by detecting local maxima in the spectrum, and the original power signal is then processed based on the frequency band boundary. Decomposed into n sub-signal components at different time scales This includes: performing a Fourier transform on the original power signal P(t) to obtain the spectrum F(ω); and in the spectrum... By detecting local maxima in the middle, a set of maxima is obtained. Based on the set of maximum points Determine the set of frequency band boundaries ,in, , Adjacent boundaries and Define a frequency band between them; based on the set of frequency band boundaries Construct n empirical wavelet filter banks , where the i-th filter Corresponding frequency band ; the original power signal By filtering with n empirical wavelet filter banks, n sub-signal components at different time scales are obtained. ,in: Where IFFT stands for Inverse Fourier Transform. For the spectrum, Let be the i-th empirical wavelet filter.

[0018] Among these methods, the Empirical Wavelet Transform (EWT) adaptively determines the frequency band division by detecting the characteristics of the signal spectrum, and then constructs a corresponding wavelet filter bank to decompose the signal. Unlike traditional wavelet transforms that use predefined wavelet bases, EWT adaptively constructs wavelet filters based on the spectral characteristics of the signal itself, making it more suitable for processing non-stationary signals such as photovoltaic power signals. This method can decompose the original power signal into sub-signal components with different time-scale characteristics.

[0019] Frequency band boundaries are the set of dividing points that divide the signal spectrum into different frequency bands. In this scheme, frequency band boundaries are determined by detecting local maxima in the spectrum, and are used to define the frequency range corresponding to each sub-signal component. Specifically, Indicates the initial boundary. The termination boundary (normalized frequency) is represented by two adjacent boundaries. and They form a frequency band Each frequency band corresponds to an eigenmode of the signal, reflecting a physical process at a specific time scale. The adaptive determination of frequency band boundaries is a key feature that distinguishes EWT from traditional wavelet transform.

[0020] Local maxima refer to points in the spectrum The frequency locations where the amplitude is greater than other points in the neighborhood are considered local maxima. In this scheme, these maxima reflect the energy concentration areas of different frequency components in the signal, representing the center frequencies of different time-scale characteristics in the photovoltaic power signal. By detecting these local maxima, the main frequency components of the signal can be identified, providing a basis for subsequent frequency band division. The valley locations between these maxima are usually selected as frequency band boundaries, thereby achieving adaptive division of the signal spectrum. In photovoltaic signals, these maxima may correspond to characteristic frequencies of different physical processes such as diurnal variations, cloud disturbances, and equipment oscillations.

[0021] Furthermore, for each component in the deterministic component set, its mean and linear trend slope are extracted as trend features, including: for the j-th sub-signal component in the deterministic component set. Iterate through all sampling points within the preset time window and extract the mean of the corresponding sub-signal; for the sub-signal components All sampling points are linearly fitted using the least squares method to obtain the slope of the linear trend that reflects the changing trend.

[0022] Furthermore, for each component in the deterministic component set, its dominant frequency and spectral peak value are extracted as periodic features, including: for sub-signal components The frequency domain transformation is performed to obtain the corresponding spectrum; peak detection is performed on the spectrum to identify the frequency point with the largest spectral amplitude, which is the main frequency that reflects the periodicity; the amplitude corresponding to the main frequency is extracted as the spectral peak value that reflects the periodicity intensity.

[0023] Furthermore, for each component in the perturbation component set, its kurtosis and maximum rate of change are extracted as abrupt change features, including: for the j-th sub-signal component in the perturbation component set. Calculate its higher-order statistical characteristics and extract the kurtosis value, which reflects the sharpness of the signal; for sub-signal components Differential processing is performed to calculate the rate of change between adjacent sampling points, and the maximum value of the rate of change is obtained as the maximum rate of change reflecting the degree of signal abrupt change.

[0024] Furthermore, for each component in the perturbation component set, its variance and entropy values ​​are extracted as random features, including: for sub-signal components Calculate the corresponding statistical discrete features and extract the variance reflecting the degree of signal fluctuation; for sub-signal components Information entropy analysis is performed to calculate the sample entropy value as a measure of signal complexity and randomness.

[0025] Another aspect of this application provides a system for extracting non-stationary signal features from a distributed photovoltaic system, comprising: a data acquisition module for acquiring the raw power signal of the distributed photovoltaic system within a preset time window. And environmental parameters synchronized with the original power signal P(t) in time, including the light intensity signal. and temperature signal The physical constraint module utilizes environmental parameters to construct physical constraint functions. ,in: The signal decomposition module decomposes the original power signal. Perform empirical wavelet transform to obtain the original power signal. The spectrum; the frequency band boundary is determined by detecting local maxima in the spectrum, and the original power signal is then processed based on the frequency band boundary. Decomposed into n sub-signal components at different time scales .

[0026] The physical consistency calculation module calculates each sub-signal component. Substitute into physical constraint function Calculate the physical consistency index of the i-th sub-signal component. ,in: Component classification module, based on physical consistency index Set threshold Classify the n sub-signal components: when When, the i-th sub-signal component Classified as a set of deterministic components; when When, the i-th sub-signal component Classified as a set of perturbation components; among which the threshold .

[0027] The feature extraction module includes: a deterministic feature extraction unit, which extracts the mean and linear trend slope as trend features, and the dominant frequency and spectral peak value as periodic features for each component in the deterministic component set; a perturbation feature extraction unit, which extracts the kurtosis and maximum rate of change as abrupt change features, and the variance and entropy value as random features for each component in the perturbation component set; and a feature vector construction module, which arranges the trend features, periodic features, abrupt change features, and random features in a preset order to construct a representation of the original power signal. (Multi-scale feature vectors)

[0028] Compared to existing technologies, the advantages of this application are:

[0029] Traditional decomposition methods such as VMD are purely based on mathematical optimization and cannot distinguish components with different physical meanings in photovoltaic signals. This results in a lack of interpretability in the decomposition results, making them difficult to use for subsequent fault diagnosis or power prediction.

[0030] This application first obtains the original power signal and time synchronization environmental parameters of the photovoltaic system, and constructs a physical constraint function based on the photovoltaic power generation physical model. Then, it uses empirical wavelet transform to decompose the original power signal into multi-scale components, obtaining sub-signal components at different time scales. By calculating the physical consistency index of each sub-signal component, the sub-signal components are automatically classified into a set of deterministic components and a set of perturbation components. Finally, it adopts a differentiated feature extraction strategy for the component sets with different physical properties, extracting trend features and periodic features for deterministic components, and extracting mutation features and random features for perturbation components, thus constructing a multi-scale feature vector.

[0031] This application can automatically identify and separate deterministic components from normal photovoltaic responses and disturbance components caused by abnormal factors; it provides more targeted and physically meaningful feature inputs for subsequent fault diagnosis and power prediction, thereby improving the intelligence level of photovoltaic system operation and maintenance. Attached Figure Description

[0032] This application will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0033] Figure 1This is an exemplary flowchart of a method for extracting non-stationary signal features from a distributed photovoltaic system according to some embodiments of this application;

[0034] Figure 2 This is an exemplary flowchart illustrating the extraction of sub-signal components at different time scales according to some embodiments of this application;

[0035] Figure 3 This is an exemplary flowchart illustrating the extraction of features from a deterministic set of components according to some embodiments of this application;

[0036] Figure 4 This is an exemplary flowchart illustrating the extraction of features from a set of perturbation components according to some embodiments of this application;

[0037] Figure 5 This is an exemplary block diagram of a non-stationary signal feature extraction system for a distributed photovoltaic system, as shown in some embodiments of this application. Detailed Implementation

[0038] The methods and systems provided in the embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0039] like Figure 1 As shown, S1 acquires the raw power signal of the distributed photovoltaic system within a preset time window. And environmental parameters synchronized with the original power signal P(t) in time, including the light intensity signal. and temperature signal S2, Constructing physical constraint functions using environmental parameters. S3, for the original power signal Perform empirical wavelet transform to obtain the original power signal. The spectrum; the frequency band boundary is determined by detecting local maxima in the spectrum, and the original power signal is then processed based on the frequency band boundary. Decomposed into n sub-signal components at different time scales S4, convert each sub-signal component Substitute into the physical constraint function Calculate the physical consistency index of the i-th sub-signal component. S5, based on the physical consistency index calculated in step 4. Set threshold Classify the n sub-signal components: when When, the i-th sub-signal component Classified as a set of deterministic components; when When, the i-th sub-signal component The components are categorized into a set of perturbation components; S6, for each component in the deterministic component set obtained in step 5, extract its mean and linear trend slope as trend features, and extract its main frequency and spectral peak value as periodic features; for each component in the perturbation component set, extract its kurtosis and maximum rate of change as abrupt change features, and extract its variance and entropy value as random features; S7, arrange the trend features, periodic features, abrupt change features, and random features extracted in S6 in a preset order to construct a representation of the original power signal. Multiscale feature vectors.

[0040] Specifically, in a distributed photovoltaic system, the following devices are used to collect the required data:

[0041] Power data acquisition: The AC power value P(t) output by the photovoltaic inverter is read in real time through the RS485 communication interface or Ethernet interface. The acquisition frequency is set to 1Hz, that is, 1 data point is acquired per second to ensure that rapid changes in power can be captured.

[0042] Irradiance acquisition: A calibrated radiometer is installed near the photovoltaic array to measure the total horizontal irradiance I(t), in W / m². The radiometer transmits data to the data acquisition unit via a 4–20 mA current signal or Modbus protocol.

[0043] Temperature data acquisition: A PT100 temperature sensor is installed on the backsheet of the photovoltaic module to measure the module's operating temperature T (t), in °C. The temperature signal is converted into a standard signal by a temperature transmitter and then connected to the data acquisition system.

[0044] Depending on the application requirements, the time window can be flexibly set: Intraday analysis: Set a 24-hour time window from 0:00 to 23:59 of the day to analyze daily cycle characteristics; Short-term analysis: Set a 1-4 hour time window to analyze rapid changes such as cloud cover; Long-term analysis: Set a 7-day or 30-day time window to analyze seasonal trends.

[0045] To ensure data time consistency, the following synchronization scheme is adopted: all acquisition devices synchronize with the same time server via NTP (Network Time Protocol), achieving millisecond-level time accuracy; each data point is stamped with a unified timestamp format during data acquisition, which is "YYYY-MM-DDHH:MM:SS.fff"; a data caching mechanism is set up so that when data from a certain sensor arrives with a delay, the data is aligned according to the timestamp.

[0046] Real-time quality checks are performed during data acquisition: Power signal range check: 0≤P(t)≤1.1×P_rated, exceeding the range is marked as abnormal; Illuminance range check: 0≤I(t)≤1200W / m², automatically zeroed at night; Temperature signal range check: -20°C≤T(t)≤85°C, exceeding the range is considered a sensor fault.

[0047] S2, Constructing physical constraint functions using environmental parameters , Where P is the power value, Where I is the system's rated power and I is the illuminance value. Standard light intensity, Here, T represents the average temperature coefficient, and T is the temperature value. For standard test temperature, This is the illumination efficiency correction function. For system availability; This is the sum of the rated power of all photovoltaic strings under the inverter, in W. ; The value range is: 0.003 to 0.005 / °C for monocrystalline silicon modules; ; Where α is the light saturation coefficient, and the value of α ranges from 0.05 to 0.08; The value ranges from 0.75 to 0.95.

[0048] In particular, traditional signal decomposition methods such as VMD rely entirely on mathematical optimization objectives (e.g., minimizing the sum of the bandwidths of each mode) for their decomposition criteria, failing to determine what physical process each component represents. This scheme, however, constructs a physical constraint function... A clear physical criterion was established: any power component that conforms to the physical laws of photovoltaic power generation should have a value close to zero under this constraint function.

[0049] This application transforms the abstract mathematical decomposition problem into a discrimination problem with clear physical meaning. When a certain sub-signal component After substituting the constraint function, if This indicates that the variation pattern of this component is highly consistent with the changes in environmental parameters (light and temperature), which conforms to the photoelectric conversion physical mechanism of photovoltaic cells, and therefore can be identified as a "normal photovoltaic response"; conversely, if A large deviation indicates that the component contains elements that cannot be explained by environmental factors, such as equipment failure, power grid disturbances, and other abnormal factors.

[0050] like Figure 2 As shown, in S3, the original power signal... Performing empirical wavelet transform includes: S31, performing Fourier transform on the original power signal obtained in S1 to obtain a spectrum ; S32, detecting local maximum points in the spectrum to obtain a set of maximum points ; S33, determining a set of frequency band boundaries based on the set of maximum points , where , , an adjacent boundary and define a frequency band; S34, constructing n empirical mode functions according to the set of frequency band boundaries, where the i-th empirical mode function corresponds to the frequency band , and is defined as:

[0051] When i = 1:

[0052] ;

[0053] When 1 < i < n:

[0054] ; where, τ ∈ (0, 1) is the transition band width parameter, is the transition function;

[0055] S35, performing convolution operation on the original power signal and n empirical mode functions to obtain n sub-signal components at different time scales : ; where, IFFT represents inverse Fourier transform, is the spectrum obtained in S31, is the i-th empirical mode function.

[0056] Particularly, different from traditional methods such as VMD that require preset decomposition levels, this application uses EWT to adaptively determine the decomposition structure by detecting local maximum points of the spectrum . Different physical processes in a photovoltaic system naturally have different characteristic frequencies: daily periodic variations are concentrated in the low frequency band, power fluctuations caused by cloud shading are distributed in the middle frequency band, and equipment failures or grid disturbances may generate high frequency oscillations. These physical processes at different time scales are manifested as maximum points of energy aggregation in the spectrum.

[0057] This application realizes guiding the decomposition process by the physical characteristics of the signal itself rather than artificially setting decomposition parameters by detecting these maximum points and accordingly determining the frequency band boundaries . This adaptability ensures that each decomposed sub-signal component Each corresponds to a real physical process in the photovoltaic system, thus avoiding the mode mixing problem caused by forced decomposition.

[0058] Furthermore, the empirical mode function in this application Adopted based on The smooth transition band conforms to the spectral characteristics of actual physical processes—real physical processes are not completely separated, but rather have a certain degree of spectral overlap. This design makes the decomposition results closer to the physical essence of photovoltaic signals.

[0059] S4, each sub-signal component Substitute into the physical constraint function Calculate the physical consistency index of the i-th sub-signal component. , The integration interval is the preset time window in step S1;

[0060] S5, based on the physical consistency index calculated in step 4 Set threshold Classify the n sub-signal components: when When, the i-th sub-signal component Classified as a set of deterministic components; when When, the i-th sub-signal component Classified as a set of disturbance components;

[0061] Set threshold The formula is as follows: Where k is the adjustment coefficient, and its value ranges from 0.5 to 2.0. Physical consistency index The mean, Physical consistency index Standard deviation;

[0062] In particular, traditional signal processing methods, after obtaining the decomposition results, often rely on human experience or simple energy proportions to determine the properties of each component. This method is highly subjective and lacks theoretical basis. This scheme, however, introduces a physical consistency index... They creatively established a quantitative physical discrimination criterion.

[0063] Each sub-signal component Whether it conforms to the physical laws of photovoltaics can be transformed into a calculable numerical index. This is achieved by calculating the cumulative squared error of the components deviating from the physical model over the entire time window. This directly reflects the degree of "physical rationality" of the component. This quantitative assessment based on physical mechanisms makes the originally vague classification problem clear and explicit.

[0064] Furthermore, the adaptive threshold setting method adopted in this application It fully considers the differences between various photovoltaic systems and operating conditions. This is achieved by using the statistical characteristics (mean) of the physical consistency index. and standard deviation This design uses a fixed threshold to determine the classification threshold, avoiding potential misjudgments. This allows the classification criteria to be automatically adjusted based on the physical characteristics of the specific signal, truly realizing "letting physical laws guide classification decisions."

[0065] S6, such as Figure 3 As shown, for the j-th sub-signal component in the deterministic component set... Calculate its mean: Where N is the number of sampling points within the preset time window, This refers to the k-th sampling time.

[0066] Fitting linear functions using the least squares method Extract the slope of the linear trend : The summation is performed on k from 1 to N.

[0067] For the j-th sub-signal component in the deterministic component set Perform Discrete Fourier Transform: , where FFT stands for Fast Fourier Transform.

[0068] In the spectrum Find the frequency point with the largest amplitude to obtain the main frequency. and the corresponding spectral peak : , The search range of f is , The sampling frequency.

[0069] like Figure 4 As shown, for the j-th sub-signal component in the set of disturbance components... Calculate its kurtosis: ,in, for The mean, for The standard deviation of E[*], where E[*] represents the expected value;

[0070] Calculate the maximum rate of change: The derivative is calculated using the finite difference method: Δt is the sampling time interval.

[0071] For the j-th sub-signal component in the set of disturbance components Calculate its variance: ,in, for The mean of the samples, where N is the number of sampling points;

[0072] Calculate sample entropy SampEn: Set embedding dimension m=2, tolerance Count the number of template vector matches of length m. Matching number with template vector of length m+1 Sample entropy The criteria for template matching is that the maximum difference between vectors is less than the tolerance r.

[0073] S7 arranges the trend features, periodic features, mutation features and random features extracted in S6 in a preset order to construct a multi-scale feature vector representing the original power signal P(t).

[0074] like Figure 5 As shown, a non-stationary signal feature extraction system for a distributed photovoltaic system includes: a data acquisition module for acquiring the raw power signal of the distributed photovoltaic system within a preset time window. And environmental parameters synchronized with the original power signal P(t) in time, including the light intensity signal. and temperature signal The physical constraint module utilizes environmental parameters to construct physical constraint functions. ,in: The signal decomposition module decomposes the original power signal. Perform empirical wavelet transform to obtain the original power signal. The spectrum; the frequency band boundary is determined by detecting local maxima in the spectrum, and the original power signal is then processed based on the frequency band boundary. Decomposed into n sub-signal components at different time scales .

[0075] The physical consistency calculation module calculates each sub-signal component. Substitute into physical constraint function Calculate the physical consistency index of the i-th sub-signal component. ,in: Component classification module, based on physical consistency index Set threshold Classify the n sub-signal components: when When, the i-th sub-signal component Classified as a set of deterministic components; when When, the i-th sub-signal component Classified as a set of perturbation components; among which the threshold .

[0076] The feature extraction module includes: a deterministic feature extraction unit, which extracts the mean and linear trend slope as trend features, and the dominant frequency and spectral peak value as periodic features for each component in the deterministic component set; a perturbation feature extraction unit, which extracts the kurtosis and maximum rate of change as abrupt change features, and the variance and entropy value as random features for each component in the perturbation component set; and a feature vector construction module, which arranges the trend features, periodic features, abrupt change features, and random features in a preset order to construct a representation of the original power signal. (Multi-scale feature vectors)

[0077] The foregoing illustrative description of the present application and its embodiments is not restrictive and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. The accompanying drawings are only one embodiment of the present application, and the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present application, such designs should fall within the scope of protection of this application. Furthermore, the word "comprising" does not exclude other elements or steps, and the word "a" preceding an element does not exclude the inclusion of "a plurality" of that element. Terms such as "first," "second," etc., are used to indicate names and do not indicate any specific order.

Claims

1. A method for distributed photovoltaic system non-stationary signal feature extraction, characterized in that, Comprise: S1, obtaining original power signals of the distributed photovoltaic system in a preset time window , and environment parameters time-synchronized with the original power signals , the environment parameters comprising an illumination intensity signal and a temperature signal ; S2, constructing a physical constraint function using the environmental parameters ; S3, performing empirical wavelet transform on the original power signal to obtain a spectrum of the original power signal S4, performing empirical wavelet transform on the spectrum of the original power signal to obtain a spectrum of the spectrum of the original power signal S5, performing empirical wavelet transform on the spectrum of the spectrum The band boundaries are determined by detecting local maxima of the spectrum, and the original power signal is decomposed into n different time-scale sub-signal components according to the band boundaries ;​ S4, calculating a physical consistency index of the i-th sub-signal component , substituting into a physical constraint function , calculating a physical consistency index of the i-th sub-signal component ; S5, based on the physical consistency index calculated in step 4 , setting a threshold value , classifying the n sub-signal components: When the ith sub-signal component is classified as a deterministic component set; When the ith sub-signal component is classified as a set of disturbance components; S6, for each component in the certainty component set obtained in step 5, extract its mean value, linear trend slope as trend feature, extract its main frequency, spectrum peak as period characteristic; For each component in the disturbance component set, extract its kurtosis, maximum change rate as mutation feature, extract its variance, entropy value as random characteristic; S7, arrange the trend feature, the periodic feature, the mutation feature and the random feature extracted in S6 according to a preset order to construct a multi-scale feature vector representing the original power signal ; Utilizing environmental parameters to construct physical constraint functions including: wherein P is a power value, is a system rated power, is a standard light intensity, is an average temperature coefficient, is a standard test temperature, is a light efficiency correction function, is a system availability; calculating a physical consistency indicator for the i-th sub-signal component : Wherein, the integral interval is the preset time window in step S1.

2. The distributed photovoltaic system non-stationary signal feature extraction method according to claim 1, characterized in that: Setting a threshold by the following equation: wherein k is a regulation coefficient, and the value range is 0.5-2.0; is the mean value of the physical consistency index is the mean value of the physical consistency index is the standard deviation of the physical consistency index is the standard deviation of the physical consistency index 3. The distributed photovoltaic system non-stationary signal feature extraction method according to claim 1 or 2, characterized in that: S3, decomposing the original power signal into n different time scale sub-signal components according to the frequency band boundary , comprising:​ The Fourier transform of the original power signal P(t) is performed to obtain a frequency spectrum ; In the spectrum local maximum points are detected to obtain a maximum point set ; based on a set of maximum points , determining a set of band boundaries wherein, , , defining a band between adjacent boundaries and ​ According to a frequency band boundary set , constructing n empirical wavelet filter banks where the i-th filter corresponds to a frequency band ; filtering the original power signal filtering through n empirical wavelet filter banks to obtain n sub-signal components of different time scales wherein: where IFFT denotes an inverse Fourier transform, is the spectrum, is the i-th empirical wavelet filter.

4. The distributed photovoltaic system non-stationary signal feature extraction method according to claim 3, characterized in that: For each component in the certainty component set, extract its mean value, linear trend slope as trend feature, comprising: the jth sub-signal component in the deterministic component set traverse all sampling points in the preset time window to extract the mean value of the corresponding sub-signal to the sub-signal components All the sampling points of the sub-signal components are linearly fitted by the least square method to obtain a linear trend slope reflecting the trend of change.

5. The distributed photovoltaic system non-stationary signal feature extraction method according to claim 4, characterized in that: For each component in the certainty component set, extract its main frequency, spectrum peak as period characteristic, comprising: sub-signal components performing a frequency domain transform to obtain corresponding frequency spectra; 6. The distributed photovoltaic system non-stationary signal feature extraction method according to claim 3, characterized in that: For each component in the disturbance component set, extract its kurtosis, maximum change rate as mutation feature, comprising:

7. The distributed photovoltaic system non-stationary signal feature extraction method according to claim 6, characterized in that: For each component in the disturbance component set, extract its variance, entropy value as random characteristic, comprising: for j = 1, 2,..., N Calculate the high order statistical characteristics, and extract the kurtosis value reflecting the sharpness of the signal sub-signal components The differential processing is performed to calculate the change rate between adjacent sampling points, and the maximum value of the change rate is obtained as the maximum change rate reflecting the signal mutation degree. Comprise: The feature extraction module comprises: sub-signal components a corresponding statistical dispersion feature is calculated, and a variance reflecting the fluctuation degree of the signal is extracted; sub-signal components Information entropy analysis is performed to calculate sample entropy value as a measure of signal complexity and randomness.

8. A distributed photovoltaic system non-stationary signal feature extraction system for implementing the method of any one of claims 1 to 7, characterized in that, The certainty feature extraction unit, for each component in the certainty component set, extract its mean value, linear trend slope as trend feature, extract its main frequency, spectrum peak as period characteristic; A data acquisition module acquires original power signals of the distributed photovoltaic system within a preset time window , and environment parameters time-synchronized with the original power signals P(t), the environment parameters including an illumination intensity signal and a temperature signal ; a physical constraint module constructs a physical constraint function using the environmental parameters wherein: ； a signal decomposition module, which performs an empirical wavelet transform on the original power signal to obtain a frequency spectrum of the original power signal, determines band boundaries by detecting local maxima of the frequency spectrum, and decomposes the original power signal into n sub-signal components of different time scales according to the band boundaries, ​​​ a physical consistency calculation module, which calculates a physical consistency index for each sub-signal component substituting the physical constraint function , calculates a physical consistency index for the i-th sub-signal component wherein: ； component classification module, based on a physical consistency index , setting a threshold value , classifying n sub-signal components: when , the i-th sub-signal component is classified into a deterministic component set; when , the i-th sub-signal component is classified into a disturbance component set; wherein the threshold value ; The disturbance feature extraction unit, for each component in the disturbance component set, extract its kurtosis, maximum change rate as mutation feature, extract its variance, entropy value as random characteristic; ​ ​ The feature vector construction module arranges the trend feature, the periodic feature, the mutation feature and the random feature according to a preset order to construct a multi-scale feature vector representing the original power signal .

Images