Distributed photovoltaic performance fluctuation feature mining method

Abstract

The invention relates to the technical field of power grid operation and maintenance, in particular to a distributed photovoltaic performance fluctuation feature mining method, which comprises the following steps of: acquiring output parameters, environmental parameters and operation state parameters of a photovoltaic module to construct an original data set, judging and eliminating abnormal values by adopting a 3 sigma criterion in combination with trend consistency, filling missing values by improving an adjacent weighted completion algorithm, and obtaining a data set; obtaining a preprocessed data set; the data set is preprocessed through Z-score standardization processing so as to eliminate dimensional differences, the correlation degree of the environmental parameters and the photovoltaic output power is calculated through the Pearson's correlation coefficient, and key influence factors are determined; a standardized photovoltaic output power sequence and a key environment parameter sequence are used as input, and an initial time sequence fluctuation feature set of fluctuation indexes is constructed based on a sliding time window. According to the invention, through accurate and comprehensive mining of photovoltaic performance fluctuation characteristics, reliable data support is provided for photovoltaic power station operation and maintenance optimization, power prediction precision improvement and power grid scheduling decision.

Description

Technical Field

[0001] This application relates to the technical field of power grid operation and maintenance, and in particular to a method for mining the performance fluctuation characteristics of distributed photovoltaic systems. Background Technology

[0002] With the development of distributed photovoltaic (PV) power generation, the problem of power output fluctuation has become increasingly apparent. PV power output is affected by various environmental and equipment-related factors, exhibiting unstable characteristics. This not only affects its own operating efficiency and stability but may also adversely impact grid operation. Currently, the PV industry faces challenges related to operation and maintenance optimization and improving grid integration capacity. The key to solving this problem lies in the effective mining of PV fluctuation characteristics. Existing methods for mining PV performance fluctuation characteristics have shortcomings; the mined characteristics lack practicality and are insufficient to support engineering decision-making. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this application provides a method for mining the performance fluctuation characteristics of distributed photovoltaic systems, enabling accurate and comprehensive mining of photovoltaic performance fluctuation characteristics, and providing reliable data support for photovoltaic power plant operation and maintenance optimization, power prediction accuracy improvement, and grid dispatch decision-making.

[0004] The above-mentioned objective of this application is achieved through the following technical solution:

[0005] A method for mining the performance fluctuation characteristics of distributed photovoltaic systems includes the following steps:

[0006] The original dataset was constructed by collecting output parameters, environmental parameters and operating status parameters of photovoltaic modules. Outliers were removed by using the 3σ criterion combined with trend consistency judgment. Missing values ​​were filled by an improved nearest neighbor weighted completion algorithm to obtain a preprocessed dataset.

[0007] Z-score standardization was used to preprocess the dataset to eliminate dimensional differences. The correlation between environmental parameters and photovoltaic output power was calculated using the Pearson correlation coefficient to identify key influencing factors.

[0008] Using the standardized photovoltaic output power sequence and key environmental parameter sequence as input, an initial time series fluctuation feature set of volatility index is constructed based on a sliding time window;

[0009] The fluctuation standard deviation sequence in the initial feature set is selected as input, and the fluctuation standard deviation sequence is decomposed into three types of intrinsic mode components (high frequency, medium frequency, and low frequency) and residual trend components.

[0010] The three intrinsic mode components obtained from variational mode decomposition are combined with the initial feature set to form an expanded feature set. The covariance matrix is ​​calculated through principal component analysis and eigenvalue decomposition is performed to construct the core wave feature set.

[0011] Based on the core feature set, four types of quantitative indicators are calculated: volatility intensity, volatility frequency, volatility persistence, and volatility stability.

[0012] A dual-index verification system of "feature discrimination degree - prediction contribution degree" is adopted to output a set of fluctuation features that have passed the verification and a feature-performance correlation report.

[0013] Optionally, the output parameters of the photovoltaic module include DC voltage U, DC current I, and output power P; the environmental parameters include irradiance G, ambient temperature T, and wind speed v; and the operating status parameters include inverter efficiency η and grid connection point AC voltage Upcc.

[0014] Optional, correlation analysis: The core driving factor for photovoltaic power output fluctuations is environmental parameters. The focus is on calculating the linear correlation between three types of environmental parameters—illuminance G, ambient temperature T, and wind speed v—and photovoltaic output power P. The correlation matrix R=[rpq] is constructed using the Pearson correlation coefficient, as shown in the following formula. The Pearson correlation coefficient ranges from [-1, 1]. The closer the absolute value is to 1, the stronger the correlation between the two; the closer it is to 0, the weaker the correlation.

[0015]

[0016] In the above formula, p represents the environmental parameter dimension, m=1-3 correspond to G, T, and v respectively; q represents the output power dimension, which is fixed as the 4th column. , These are the mean values ​​of the corresponding standardized parameters. Based on engineering experience, environmental parameters with an absolute correlation value ≥ 0.7 are defined as key influencing factors. Typically, light intensity G has the highest correlation and is the core driving factor; ambient temperature T is the second highest; wind speed v has the lowest correlation and is not a key factor. Subsequent feature construction will focus on the core driving factors.

[0017] Optionally, a photovoltaic output time-series fluctuation feature set is constructed, the feature system is defined, and the coverage level, dispersion, extreme potential, changing trend, environmental correlation, fluctuation frequency, etc. are considered to comprehensively characterize the photovoltaic power output fluctuation characteristics. The output result is the initial time-series fluctuation feature set F=[fik], where k=1-6 correspond to six basic features: window mean, fluctuation standard deviation, peak factor, trend slope, environment-power coupling coefficient, and volatility index. Each feature has a clear physical meaning.

[0018] Optionally, the decomposition parameter setting and solution process, combined with the photovoltaic fluctuation characteristics, sets the high-frequency fluctuation period to 1-5 min, the medium frequency to 15-60 min, and the low frequency to several hours to one day, and optimizes the parameter decomposition mode number K=3, corresponding to the high, medium, and low scale fluctuations respectively; the penalty factor α=2000 to balance bandwidth and reconstruction error; the iteration stopping threshold ε=10 to ensure decomposition accuracy; the variational model is solved by the alternating direction multiplier algorithm, and the IMF components and center frequency are iteratively updated until the stopping condition is met, and finally outputs 3 IMF components and 1 residual trend component Res;

[0019] The output consists of three intrinsic mode components at different scales and one residual trend component, denoted as... Among them, IMF1 corresponds to high-frequency random fluctuations, IMF2 corresponds to medium-frequency periodic fluctuations, IMF3 corresponds to low-frequency trend fluctuations, and Res corresponds to long-term trends.

[0020] Optionally, step 6 involves feature dimensionality reduction and core feature extraction based on principal component analysis. The three IMF components obtained from VMD decomposition are used as new features and merged column-wise with the initial time series fluctuation feature set F to form an expanded feature set. This ensures that all features participate in the dimensionality reduction analysis;

[0021] First, calculate the expanded feature set. The covariance matrix C reflects the degree of linear correlation between features, as shown in the following formula; then, the covariance matrix C is decomposed into eigenvalues ​​to obtain 9 eigenvalues ​​and corresponding unit eigenvectors. The larger the eigenvalue, the more information the corresponding principal component contains.

[0022]

[0023] in for The mean vector is obtained by calculating the mean of each column; The feature matrix is ​​decentralized to ensure that the covariance calculation is not affected by the feature mean;

[0024] Calculate the cumulative contribution rate ηCon of each feature value. The cumulative contribution rate reflects the proportion of total information contained in the first n principal components, as shown in the following formula. Based on engineering requirements, select the first s principal components with a cumulative contribution rate ≥ 85% as core features. Multiply the expanded feature set F with the matrix formed by the first s feature vectors to obtain the core feature set and achieve feature dimensionality reduction.

[0025]

[0026] The output is the core fluctuation feature set F after dimensionality reduction, with a dimension of m×s (s≤9), a cumulative contribution rate of ≥85%, and retains the core information of the original features;

[0027] Optionally, a dual-index verification method of "feature discrimination" and "prediction contribution" is used. Feature discrimination is achieved by the difference between features under fault and normal operating conditions, as shown in the following formula; prediction contribution is achieved by the error reduction rate of feature participation in power prediction, as shown in the following formula:

[0028]

[0029]

[0030] in and σ represents the mean of feature S under fault and normal operating conditions, respectively; σ is the corresponding standard deviation; MAE is the mean absolute error of power prediction without core features, and MAE is the prediction error after adding core features. Feature effectiveness is achieved when D ≥ 1.5 and E ≥ 10%.

[0031] In summary, this application has the following beneficial technical effects:

[0032] This invention adopts a two-layer anomaly detection mechanism of "3σ criterion + trend consistency judgment". Compared with the traditional single criterion, it can reduce the misjudgment rate of outliers and effectively avoid misjudging normal fluctuations such as rapid cloud cover as anomalies. The improved proximity weighted completion algorithm reduces the completion error of continuous missing data through the design of time distance decay weight, laying a highly reliable data foundation for subsequent feature mining.

[0033] By combining time-series sliding windows and VMD decomposition, multi-scale feature extraction of photovoltaic fluctuations can be achieved, which improves the comprehensiveness of fluctuation characterization compared with traditional single feature methods. By quantifying the dynamic correlation between environmental factors and photovoltaic output into feature indicators, the matching degree between features and fluctuation physical mechanisms is improved, which can accurately identify hidden faults such as module dust accumulation and inverter efficiency degradation. The introduction of VMD algorithm to replace traditional EMD decomposition solves the mode mixing problem and improves the accuracy of multi-scale fluctuation separation.

[0034] By applying PCA dimensionality reduction technology to reduce the dimensionality of the core feature set, eliminate redundant features, shorten the training time of subsequent power prediction models, and reduce the deployment cost of edge computing devices, it is more suitable for the "decentralized and miniaturized" characteristics of distributed photovoltaics. The core feature set has been verified by dual indicators of discrimination and contribution, and has extremely strong engineering robustness. The feature effectiveness remains stable under different weather scenarios such as sunny, cloudy, and overcast conditions, solving the problems of poor scenario adaptability and high redundancy of traditional feature sets, and providing the possibility for real-time analysis at the edge.

[0035] The quantitative indicators of fluctuation characteristics such as intensity, frequency, and persistence can be directly connected to the photovoltaic operation and maintenance platform to improve the accuracy of component fault early warning and help achieve early fault detection and early operation and maintenance intervention. As a feature input for the power prediction model, it can reduce power prediction error and reduce the waste of reserve capacity caused by prediction deviation. The obtained fluctuation characteristic analysis can support the investment evaluation of photovoltaic projects and can be directly applied to the optimization of photovoltaic power plant operation and maintenance strategies, providing data support for the healthy development of the industry. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating one embodiment of this application;

[0037] Figure 2 This is a schematic diagram of a real-time meteorological data acquisition system for photovoltaic power generation according to one embodiment of this application;

[0038] Figure 3 This is a schematic diagram showing the variation patterns of surface irradiance and extraterrestrial irradiance corresponding to four weather types in one embodiment of this application.

[0039] Figure 4 This is a schematic histogram showing the correlation coefficient distribution between photovoltaic daily power generation and influencing factors according to an embodiment of this application. Detailed Implementation

[0040] The present application will be further described in detail below with reference to the accompanying drawings.

[0041] Distributed photovoltaic (PV) power generation is significantly affected by natural and equipment factors such as sunlight intensity, ambient temperature, cloud cover, and module aging, exhibiting strong volatility, intermittency, and randomness. This performance fluctuation not only impacts the power generation efficiency and operational stability of the PV power plant itself but also impacts the power balance, voltage quality, and relay protection of the distribution network, potentially leading to safety issues such as distribution network frequency fluctuations and line overloads. The PV industry faces the need for cost reduction and efficiency improvement in operation and maintenance, as well as grid safety and integration. Accurate fluctuation feature mining enables early fault detection and reduces operation and maintenance costs; quantifying the impact of PV fluctuations on the distribution network enhances the grid's capacity to accommodate PV. To meet these needs, constructing a feature set that matches the physical mechanism of PV fluctuations and achieving accurate mining of distributed PV performance fluctuation characteristics is crucial.

[0042] In traditional distributed photovoltaic (PV) performance fluctuation feature mining, monitoring data often faces the combined problem of sudden noise and continuous missing data. Existing preprocessing methods only use simple filtering or mean completion, failing to design solutions for the "sudden noise-continuous missing data" composite problem in PV monitoring data, thus failing to ensure data reliability. The feature construction process relies on the statistical characteristics of output data, neglecting the coupling relationship between environmental parameters and PV output power, resulting in features lacking physical meaning. Fluctuation decomposition using algorithms such as wavelet decomposition suffers from mode mixing and boundary effects, distorting short-term fluctuation features. Furthermore, the lack of a verification mechanism to correlate extracted features with PV performance renders the features insufficiently practical and unable to support engineering decisions. Specifically, insufficient data reliability directly affects the accuracy of fluctuation features; feature construction ignoring the driving role of environmental factors leads to a disconnect between features and physical mechanisms; boundary effects of decomposition algorithms distort short-term fluctuation components; and the lack of a verification mechanism prevents features from effectively correlating with actual performance status.

[0043] For example, during the operation of a distributed photovoltaic power station, when a rapid cloud movement causes a sudden change in sunlight intensity, the photovoltaic output power drops sharply. Simultaneously, sensor malfunctions result in continuous missing ambient temperature data. In this situation, traditional data preprocessing methods use simple linear interpolation to fill in the missing values, leading to inconsistencies between the completed data and the actual trend. Feature construction is based solely on statistical indicators of the power sequence, ignoring the driving effect of temperature changes on power fluctuations. When using wavelet decomposition to extract fluctuation features, boundary effects distort short-term fluctuation components, failing to accurately reflect the characteristics of the sudden power drop. The resulting feature set fails to pass the fault condition verification, preventing maintenance personnel from timely identifying potential faults. Furthermore, in this scenario, deficiencies in data preprocessing introduce noise interference into the original dataset, the correlation between environmental parameters and power is not quantified, and the initial time-series fluctuation feature set fails to reflect the environment-power coupling relationship. The distortion during the decomposition process further amplifies the feature error, making it difficult for the feature set to accurately characterize the photovoltaic performance fluctuation state.

[0044] If the above problems are not addressed, distorted fluctuation characteristics will lead to misjudgments of photovoltaic performance status, potentially causing missed or false fault reports and increasing operation and maintenance risks. The lack of correlation verification between features and photovoltaic performance means that extracted features cannot effectively guide power prediction and grid dispatch, reducing the distribution network's capacity to accommodate photovoltaics. In the long term, insufficient data reliability and poor feature usability will affect the operational stability of photovoltaic power plants, continuously impacting the power balance and voltage quality of the distribution network, and may even trigger relay protection malfunctions, threatening grid security. Therefore, it is urgent to construct a feature set that matches the physical mechanism of photovoltaic fluctuations to improve data reliability, enhance feature comprehensiveness, and ensure engineering practicality.

[0045] To address this, this application proposes a method for mining the performance fluctuation characteristics of distributed photovoltaic systems, comprising the following steps:

[0046] S101: Collect the output parameters, environmental parameters, and operating status parameters of the photovoltaic module to construct the original dataset;

[0047] S102, preprocess the original dataset to obtain a preprocessed dataset;

[0048] S103, calculate the correlation between environmental parameters and photovoltaic output power to identify key influencing factors;

[0049] S104, using the photovoltaic output power sequence and key environmental parameter sequence as input, constructs the initial time series volatility feature set of volatility index based on the sliding event window;

[0050] S105, take the fluctuation standard deviation sequence of the initial feature set as input, and decompose the fluctuation standard deviation sequence into three types of intrinsic mode components (high frequency, medium frequency, and low frequency) and residual trend components through variational mode decomposition;

[0051] S106, the obtained three types of intrinsic mode components are combined with the initial feature set to form an expanded feature set;

[0052] S107, based on principal component analysis, performs feature dimensionality reduction and core feature extraction on the expanded feature set;

[0053] S108 calculates four types of quantitative indicators based on the core feature set: volatility intensity, volatility frequency, volatility persistence, and volatility stability.

[0054] S109, Verify the effectiveness of the features;

[0055] S110 outputs a verified set of fluctuation characteristics and a report on the correlation between these characteristics and photovoltaic performance.

[0056] For ease of understanding, the following explains some key terms in this embodiment:

[0057] Raw dataset: refers to the raw data collection directly collected from photovoltaic modules, environmental monitoring equipment, and operation monitoring systems without any processing, and usually contains noise, outliers, and missing values;

[0058] Correlation degree: an indicator that measures the strength and direction of the interdependence between two or more variables; in this method, it is used to assess the degree of correlation between environmental parameters and photovoltaic output power.

[0059] Sliding event window: a data processing technique that defines a fixed-size time window and slides it across time series data to extract local features or perform local analysis within each window;

[0060] Volatility indicators are used to quantify the degree of variation and instability of photovoltaic output power over a specific period of time, such as standard deviation and peak factor.

[0061] Variational Mode Decomposition (VMD) is an adaptive, non-recursive signal decomposition method that can decompose complex non-stationary signals into a series of intrinsic mode components with different center frequencies, effectively avoiding the mode aliasing problem.

[0062] Intrinsic Mode Components (IMFs) are components with specific frequency and amplitude characteristics obtained from VMD decomposition, representing the fluctuation patterns of the original signal at different time scales.

[0063] Principal Component Analysis (PCA) is a multivariate statistical analysis method that uses orthogonal transformation to convert a set of potentially related variables into a set of linearly unrelated variables, i.e., principal components, thereby achieving data dimensionality reduction and feature extraction.

[0064] Core feature set: refers to the representative subset of features that retain most of the information of the original data after feature dimensionality reduction and extraction, and is used for subsequent quantitative index calculation;

[0065] The following section will provide further details based on specific usage scenarios.

[0066] First, step S101 is executed to collect the output parameters, environmental parameters, and operating status parameters of the photovoltaic modules to construct the raw dataset. In order to comprehensively reflect the operating status and external influences of the distributed photovoltaic system, data needs to be collected from multiple dimensions. Output parameters may include the DC voltage and DC current of the photovoltaic modules, which can be directly obtained by sensors installed at the output end of the photovoltaic modules. Environmental parameters may include light intensity and ambient temperature, which can be measured by weather stations or independent sensors. Operating status parameters may include the efficiency of the inverter and the AC voltage at the grid connection point, which can be obtained by the monitoring system built into the inverter or the monitoring equipment on the grid side. By aligning and integrating these data from different sources and of different types in the time dimension, the raw dataset can be constructed. For example, synchronous sampling can be performed at fixed time intervals (such as every minute or every five minutes), and all collected data can be stored in a unified database or file.

[0067] Then, step S102 is executed to preprocess the original dataset to obtain a preprocessed dataset. The original dataset usually contains noise, outliers, and missing values, and using it directly will affect the accuracy of subsequent analysis. Preprocessing aims to improve data quality. For example, for outliers, statistical methods can be used, such as setting a fixed threshold and marking data points that exceed the threshold range as outliers and removing them. For missing values, simple interpolation methods, such as linear interpolation or forward imputation, can be used to estimate and fill in the missing data based on the values ​​of adjacent data points. After these processes, a relatively clean, complete, and consistent preprocessed dataset can be obtained.

[0068] In step S103, the correlation between environmental parameters and photovoltaic output power is calculated to identify key influencing factors. Photovoltaic output power is affected by various environmental factors, and identifying key influencing factors helps to more accurately understand the fluctuation mechanism. The correlation can be calculated using various statistical methods. For example, a simple correlation coefficient (such as Spearman's correlation coefficient) can be used to measure the linear or nonlinear relationship between each environmental parameter and photovoltaic output power. Based on the calculated correlation values, the environmental parameters can be sorted, and according to a preset correlation threshold, the environmental parameters that have the most significant impact on photovoltaic output power can be selected as key influencing factors.

[0069] In step S104, in order to capture the dynamic characteristics of photovoltaic performance fluctuations, features need to be extracted from the time series data. Taking the photovoltaic output power sequence and key environmental parameter sequence as input, an initial time series fluctuation feature set of volatility indicators is constructed based on a sliding event window. The sliding event window is an effective method that extracts local information by moving a window of a fixed length over the time series. Within each window, various volatility indicators can be calculated. For example, the average, maximum, and minimum values ​​of photovoltaic output power within the window, as well as a simple fluctuation range (maximum value minus minimum value), can be calculated. These indicators together constitute the initial time series fluctuation feature set, reflecting the basic fluctuation characteristics of photovoltaic output power in different time periods.

[0070] Next, step S105 is executed. The fluctuation standard deviation sequence can reflect the dispersion of photovoltaic output power, but it may contain fluctuation modes of multiple time scales. The fluctuation standard deviation sequence in the initial feature set is taken as input, and the fluctuation standard deviation sequence is decomposed into three types of intrinsic mode components (high frequency, mid frequency, and low frequency) and residual trend components through variational mode decomposition. In order to analyze these modes more finely, signal decomposition techniques can be used. For example, empirical mode decomposition (EMD) or wavelet decomposition can be used to decompose the fluctuation standard deviation sequence into different frequency components. The components obtained by these decompositions can represent the fluctuation modes of high frequency, mid frequency, and low frequency, as well as a residual component that reflects the overall trend.

[0071] In step S106, the three types of intrinsic mode components obtained are combined with the initial feature set to form an expanded feature set. The intrinsic mode components obtained through signal decomposition contain deeper fluctuation information. Combining them with the initial feature set can enrich the dimension of the features and provide a more comprehensive fluctuation description. The merging process can directly add the high-frequency, mid-frequency, and low-frequency component sequences obtained from the decomposition as new feature dimensions to the original initial feature set. For example, if the initial feature set contains N features and the decomposition yields 3 intrinsic mode components, the expanded feature set will contain N+3 features.

[0072] In step S107, the expanded feature set may contain redundant information or highly correlated features, which increases computational complexity and may affect the model's generalization ability. Principal component analysis is used to perform feature dimensionality reduction and core feature extraction on the expanded feature set. Feature dimensionality reduction aims to reduce the number of features while retaining as much original information as possible. For example, methods such as linear discriminant analysis (LDA) or independent component analysis (ICA) can be used to map the high-dimensional feature space to a low-dimensional space. In this way, a few core features that can represent the main fluctuation patterns can be extracted from the expanded feature set.

[0073] Then, step S108 is executed. The core feature set is a refined representation of photovoltaic performance fluctuations and can be used to calculate more interpretable quantitative indicators. Based on the core feature set, four types of quantitative indicators are calculated: fluctuation intensity, fluctuation frequency, fluctuation duration, and fluctuation stability. For example, fluctuation intensity can be measured by the variance or energy of the core feature; fluctuation frequency can be determined by analyzing the spectral distribution after performing a Fourier transform on the core feature; fluctuation duration can be obtained by calculating the average duration of continuous fluctuation events in the core feature sequence; and fluctuation stability can be characterized by evaluating the coefficient of variation of the core feature over different time periods. These indicators provide a multi-faceted quantitative description of photovoltaic fluctuation characteristics.

[0074] Next, step S109 is executed. The extracted features need to be verified to ensure that they can accurately reflect the fluctuations in photovoltaic performance and have practical application value. The validity of the features is verified. Through verification methods, the validity of the extracted features can be preliminarily judged.

[0075] Finally, step S110 is executed, outputting a verified set of fluctuation characteristics and a report on the correlation between the characteristics and photovoltaic performance. This report may include the definition of each characteristic, the calculation method, and a qualitative or quantitative description of its relationship with photovoltaic performance (such as power generation efficiency and grid stability). For example, the report may list which characteristics are highly correlated with the downward trend of photovoltaic power generation, or which characteristics have a causal relationship with grid voltage fluctuations. This report provides a basis for decision-making in subsequent fault diagnosis, operation and maintenance optimization, and grid dispatch.

[0076] Based on the above examples, this method demonstrates a significant technical contribution to the mining of performance fluctuation characteristics in distributed photovoltaic systems.

[0077] First, addressing the issue of insufficient data reliability in existing methods, this method performs comprehensive preprocessing on the original dataset, including outlier removal and missing value imputation, ensuring data quality for subsequent analysis. For example, in the example above, the system can effectively handle abnormal data caused by sensor malfunctions and missing data caused by transmission interruptions. Compared to traditional methods that only use simple filtering or mean imputation, this significantly improves data reliability and lays a solid foundation for accurate feature mining.

[0078] Secondly, this method fully considers the driving effect of environmental factors on photovoltaic power output. By calculating the correlation between environmental parameters and photovoltaic output power and identifying key influencing factors, feature construction no longer relies solely on the statistical characteristics of output data. In the example, irradiance and ambient temperature are identified as key influencing factors and incorporated into the construction of the initial time-series fluctuation feature set. This allows the extracted features to more accurately reflect the physical mechanism of photovoltaic fluctuations, avoiding the feature bias caused by traditional methods ignoring environmental coupling relationships.

[0079] Furthermore, in terms of model construction, this method introduces Variational Mode Decomposition (VMD) to handle the fluctuation standard deviation sequence. VMD can decompose complex fluctuation signals into three types of intrinsic mode components: high-frequency, mid-frequency, and low-frequency, effectively avoiding the mode aliasing and boundary effect problems existing in traditional wavelet decomposition and other algorithms. In the example, VMD can clearly separate multi-scale fluctuation patterns caused by different physical reasons, such as distinguishing between instantaneous fluctuations caused by cloud cover and slow fluctuations caused by changes in solar radiation intensity, thereby achieving accurate separation of multi-scale features of photovoltaic fluctuations and improving the refinement of features.

[0080] Furthermore, this method effectively solves the redundancy problem of high-dimensional feature sets by using principal component analysis to reduce the dimensionality of the expanded feature set and extract core features. Based on this, four types of quantitative indicators are calculated: fluctuation intensity, fluctuation frequency, fluctuation persistence, and fluctuation stability. These indicators provide a comprehensive and intuitive quantitative description of photovoltaic fluctuation characteristics. Compared with traditional methods that rely on only a single statistical indicator, this method can deeply analyze fluctuation characteristics from multiple dimensions.

[0081] Finally, this method establishes a feature validity verification mechanism, using both feature discriminative power and prediction contribution as dual indicators to ensure the practical value of the extracted features. In the example, by comparing feature differences under different operating conditions and evaluating the prediction error reduction rate, it is ensured that the output fluctuation feature set can effectively support fault diagnosis and operation and maintenance decisions. This overcomes the deficiency of most existing methods in lacking a feature-photovoltaic performance correlation verification mechanism, making the extracted features more engineering-practical and able to directly guide operation and maintenance personnel to make accurate decisions, thereby improving the grid's acceptance of photovoltaics.

[0082] As a feasible specific implementation of this application, in order to improve data reliability, multi-dimensional monitoring data collection and raw data preprocessing were performed. The structure of the distributed photovoltaic power station monitoring system is as follows: Figure 2 As shown, multiple sources of raw data were collected, and outliers in the raw dataset were removed based on the 3σ criterion and trend consistency; missing values ​​were filled in by an improved nearest neighbor weighted completion algorithm.

[0083] Outlier removal from the original dataset based on the 3σ criterion and trend consistency refers to identifying and removing data points that significantly deviate from the overall pattern. Outliers may be caused by measurement errors, equipment malfunctions, or transient interference; if not removed, they will severely interfere with the data analysis results. The 3σ criterion is a statistical method used to identify data points in normally distributed data that deviate from the mean by more than three standard deviations; these points are considered outliers. Combining trend consistency further considers the time-series characteristics of the data, ensuring that the removed outliers are not only numerically abnormal but also inconsistent with the overall trend of the data, avoiding the accidental deletion of normal fluctuations. For example... One approach is to first calculate the mean and standard deviation of the data sequence, marking data points that exceed the mean plus or minus three standard deviations as potential outliers. Then, by analyzing the trends of these potential outliers and their adjacent points (e.g., comparing local slopes or differences), if the trend of a point is significantly inconsistent with that of its surrounding data points, it is ultimately confirmed as an outlier and removed. Another approach is to first perform a moving average or exponential smoothing on the data to obtain a smooth trend line, then calculate the residual between each data point and this trend line. Finally, apply the 3σ criterion to the residuals and combine this with the local trend changes of the residual sequence to identify outliers.

[0084] Improving the nearest neighbor weighted completion algorithm to fill in missing values ​​refers to using the similarity between data points to estimate and fill in missing data points in a dataset. Missing values ​​are a common problem in datasets and can lead to incomplete or biased data analysis. The nearest neighbor weighted completion algorithm fills in the missing value by finding the K most similar neighbors of the missing data point and performing a weighted average based on the known values ​​of these neighbors. The improvement lies in optimizing the similarity measure or weighting method to improve the accuracy of the completion, especially when processing time series data, and better capturing the dynamic characteristics of the data. For example, when finding nearest neighbors, it can adopt... The similarity of time series can be measured using dynamic time warping (DTW) distance or distance based on correlation coefficients, rather than simple Euclidean distance. After determining the K nearest neighbors, they can be imputed by weighted averaging based on the values ​​of these neighbors at the missing time. The weights can be the reciprocal of the similarity measure or a time decay factor can be introduced so that neighboring data points closer to the missing time have higher weights. Another approach is to combine machine learning models, such as random forests or gradient boosting trees, with neighboring data points as input features of the model. The model can then predict the missing values, thereby achieving more intelligent weighted imputation.

[0085] Specifically, the method used in the embodiments of this application is as follows:

[0086] The collected multi-source raw data specifically includes:

[0087] The output parameters of photovoltaic modules, such as DC voltage U, DC current I, and output power P, can be collected by the sensors built into the string inverter.

[0088] Environmental parameters, such as light intensity G, ambient temperature T, and wind speed v, can be collected by a weather station installed at the center of the component array.

[0089] Operating status parameters, such as inverter efficiency η and grid connection point AC voltage Upcc, can be collected through the distribution network monitoring terminal. The data sampling frequency is set to 1 minute, and the time span is continuous for 72 hours, covering typical weather scenarios such as sunny, cloudy, and overcast. Figure 3 This is a graph showing the variation patterns of surface irradiance and extraterrestrial irradiance corresponding to the four measured weather types.

[0090] Setting the original dataset Where m is the number of data sample time points, and n is the feature dimension. This represents the raw monitoring value of the j-th type of parameter at time i. Raw data preprocessing is divided into noise detection and anomaly removal, and data completion sub-steps. Differentiated processing strategies are designed for different types of data problems, including:

[0091] Step S1011, Noise Detection and Anomaly Removal:

[0092] The 3σ criterion, combined with trend consistency judgment, is used to avoid misjudging normal fluctuations. First, the mean μ and standard deviation σ of the j-th parameter are calculated as follows; if x satisfies |x-μ|>3σ, the parameter change trend of the three moments before and after that moment is further judged. If it is opposite to the overall trend, such as the power being stable at around 100kW at adjacent moments and then suddenly dropping to 0 at that moment, it is judged as an outlier and marked; if it is a normal fluctuation, such as a sudden drop in power caused by rapid cloud cover, the data is retained.

[0093]

[0094]

[0095] Step S1012, Data Completion:

[0096] For marked outliers and missing data, an improved nearest-neighbor weighted completion algorithm is adopted. Unlike traditional mean-based completion, this algorithm assigns weights based on the temporal relevance of the data; the closer the time interval, the greater the weight, thus improving completion accuracy. Five normal data points before and after the given time are selected, and the weighted average is calculated as the completed value x'. The weight w decreases linearly with increasing time distance |k|, as shown in the following formula. For data missing for more than three consecutive time points, corrections are made based on the changing trends of environmental parameters during the same period to avoid the completed data becoming out of sync with actual working conditions.

[0097]

[0098] Where wk is the weight of the data at time i+k. When k=±1, w=0.9, which is the maximum weight; when k=±5, wk=0.5, which is the minimum weight, ensuring that recent data plays a dominant role in the completion result.

[0099] The output is the preprocessed dataset after denoising and completion. Typically, data integrity is required to be ≥99.5% and noise removal rate is required to be ≥98%.

[0100] The above technical solutions effectively address the common problem of outliers and missing values ​​in the original data when mining the performance fluctuation characteristics of distributed photovoltaic (PV) systems. Outlier removal based on the 3σ criterion combined with trend consistency avoids the drawbacks of traditional methods that might mistakenly delete normal fluctuation data, ensuring data purity and enabling subsequent analysis to be based on more realistic and reliable data. Simultaneously, by improving the nearest neighbor weighted completion algorithm to fill in missing values, not only is data integrity restored, but also, because the algorithm considers the similarity between data points and the time series characteristics, the completed data better maintains the inherent structure and trend of the original data, avoiding biases that might be introduced by simple interpolation. This significantly improves the quality of the preprocessed dataset, providing high-quality input for subsequent steps such as calculating the correlation between environmental parameters and PV output power, and constructing the initial time-series fluctuation feature set of fluctuation indices. This ensures that the finally mined fluctuation features are more accurate and reliable, improving the precision and effectiveness of distributed PV performance fluctuation characteristic analysis.

[0101] As a feasible specific implementation of the present application, the present application further proposes to standardize the preprocessed dataset, and to construct an initial time series fluctuation feature set by inputting the standardized photovoltaic output power sequence and key environmental parameter sequence.

[0102] Standardizing the preprocessed dataset refers to preprocessing the data to eliminate differences in the dimensions and numerical ranges of different features, making them comparable. Its purpose is to transform the data to a uniform scale, avoiding the situation where certain features have excessively high or low weights in subsequent analysis due to differences in dimensions. Implementation methods can include, but are not limited to: Z-score standardization, which transforms the original data into a distribution with a mean of 0 and a standard deviation of 1; or Min-Max standardization, which linearly maps the original data to a specified interval, such as [0,1] or [-1,1]. Using the standardized photovoltaic output power sequence and key environmental parameter sequence as input means using the standardized photovoltaic output power sequence and key environmental parameter sequence as input data for constructing the initial time-series fluctuation feature set. Its purpose is to ensure the consistency of the input data in terms of dimensions and numerical ranges, providing a fair and comparable basis for subsequent feature calculations. Constructing the initial time-series fluctuation feature set aims to extract a series of indicators reflecting the fluctuation characteristics of photovoltaic performance based on the input sequence through a sliding event window, forming an initial feature set. Its function is to extract key information from raw time-series data that can quantify and describe fluctuation behavior, laying the foundation for subsequent fluctuation feature analysis and mining. The construction method can include, but is not limited to: calculating statistics such as mean, standard deviation, and variance within each sliding window to capture the central tendency and dispersion of the data; or calculating time-domain features such as trend slope and autocorrelation coefficient within the window to reflect the pattern of data change over time.

[0103] In this embodiment of the application, the preprocessed dataset is based on the aforementioned output. To ensure that all parameters are free of outliers and missing values, and considering the significant dimensional differences in photovoltaic monitoring data (e.g., illuminance in W / m² and voltage in V), direct analysis could lead to large numerical parameters dominating the correlation results. Therefore, a Z-score standardization method was used to map all parameters to a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of dimensions. The calculation is shown in the following formula. The standardized data can then be directly used for correlation analysis and subsequent feature calculations.

[0104]

[0105] Where μ'j and σ'j are the mean and standard deviation of the j-th type of parameter after preprocessing, respectively, which are calculated by traversing the j-th column of data in X'; To standardize the dataset, ensure that all parameters are compared on the same order of magnitude.

[0106] This application's approach standardizes the preprocessed dataset, unifying the photovoltaic output power sequence and key environmental parameter sequences to the same scale, eliminating differences in dimensions and numerical ranges between different physical quantities. Based on this, an initial time-series fluctuation feature set is constructed using the standardized photovoltaic output power sequence and key environmental parameter sequences as input. This approach ensures that all features are within a comparable range when calculating volatility indices within a sliding event window, avoiding feature weight bias caused by inconsistent original data scales. For example, if the light intensity value is much greater than the ambient temperature, unstandardized processing may lead to overemphasis on light intensity-related fluctuation features in subsequent analyses. Through standardization, the fluctuation characteristics of each key environmental parameter and photovoltaic output power can be fairly evaluated and extracted, enabling the constructed initial time-series fluctuation feature set to more accurately and objectively reflect the true fluctuation patterns of distributed photovoltaic performance. This provides high-quality input for subsequent feature dimensionality reduction, core feature extraction, and quantification index calculation, thereby improving the accuracy and reliability of the entire fluctuation feature mining method.

[0107] As a feasible specific implementation of this application's embodiment, in the distributed photovoltaic performance fluctuation feature mining method, accurately and comprehensively collecting the output parameters, environmental parameters, and operating status parameters of photovoltaic modules is the foundation for constructing the original dataset. However, if the collected parameter types are not specific enough or fail to cover key factors that significantly affect photovoltaic performance fluctuations, the subsequent feature mining process may not be able to fully reveal the essential laws of the fluctuations, thereby affecting the accuracy and reliability of the quantitative analysis of fluctuation intensity, frequency, persistence, and stability. This may result in the constructed fluctuation feature set not fully reflecting the actual operating status of the photovoltaic system, thus limiting the in-depth understanding and effective management of photovoltaic performance fluctuations.

[0108] In this regard, this application further proposes that the output parameters of the photovoltaic module include DC voltage, DC current, and output power; the environmental parameters include light intensity, ambient temperature, and wind speed; and the operating status parameters include inverter efficiency and grid connection point AC voltage.

[0109] Among the output parameters of a photovoltaic (PV) module, DC voltage refers to the voltage generated by the PV module during power generation. It is one of the important indicators for measuring the power output capability of the PV module, and its fluctuations directly reflect the impact of changes in the internal or external environment on power generation performance. DC voltage can be measured by connecting a voltage sensor in parallel at the output terminal of the PV module, or by obtaining it through the sampling circuit inside the inverter. DC current refers to the current generated by the PV module during power generation. It is a key parameter that, together with DC voltage, determines the output power of the PV module, and its fluctuations are usually closely related to factors such as light intensity and module temperature. DC current can be measured by connecting a current sensor in series in the output circuit of the PV module, or by obtaining it through the sampling circuit inside the inverter. Output power refers to the actual electrical power generated by the PV module, usually the product of DC voltage and DC current. It is the most direct indicator reflecting the power generation performance of the PV module, and its fluctuations are the core focus of this method. Output power can be measured directly by a power meter, or it can be calculated from the collected DC voltage and DC current.

[0110] Among environmental parameters, irradiance refers to the solar radiation power received per unit area, and it is the most direct and important environmental factor affecting the power generation of photovoltaic (PV) modules. Irradiance can be measured using a solar intensity meter installed at the PV power plant site, or estimated using satellite remote sensing data. Ambient temperature refers to the temperature of the air surrounding the PV modules. It not only affects the temperature of the PV cells, thus affecting their power generation efficiency and output power, but may also affect the operating status of other equipment. Ambient temperature can be measured using temperature sensors installed at the PV power plant site, or obtained from weather station data. Wind speed refers to the speed of airflow. It mainly affects the heat dissipation of the PV modules through convection heat transfer, thereby indirectly affecting the cell temperature and power generation efficiency. Wind speed can be measured using anemometers installed at the PV power plant site, or obtained from weather station data.

[0111] Inverter efficiency, a key operating parameter, refers to the efficiency with which the inverter converts direct current (DC) generated by photovoltaic (PV) modules into alternating current (AC). It is a critical indicator of inverter performance, and its fluctuations directly impact the overall output power of the PV system. Inverter efficiency can be obtained by monitoring the inverter's input DC power and output AC power and calculating their ratio. Grid-connected AC voltage refers to the AC voltage at the connection point between the PV system and the grid. Its stability and fluctuations reflect the grid's operating status and the PV system's impact on the grid. Grid-connected AC voltage can be measured by installing a voltage transformer and voltmeter at the grid-connected point, or by utilizing the inverter's internal monitoring functions.

[0112] In the method for mining the performance fluctuation characteristics of distributed photovoltaic (PV) systems, to ensure sufficient evidence and accuracy in analyzing PV system performance fluctuations, this application explicitly specifies the types of PV module output parameters, environmental parameters, and operating status parameters to be collected. This lays a solid data foundation for subsequent data preprocessing, correlation calculation, fluctuation feature set construction, and quantitative index calculation. Specifically, by collecting three types of PV module output parameters—DC voltage, DC current, and output power—the power generation performance and fluctuations of the PV modules can be comprehensively reflected. Among these, output power, as a core indicator, is the focus of this method. Simultaneously, by collecting three types of environmental parameters—illuminance, ambient temperature, and wind speed—the direct or indirect impacts of external environmental changes on PV power generation performance can be accurately captured. These environmental factors are the main driving forces behind PV output power fluctuations. Furthermore, by collecting two operating status parameters—inverter efficiency and grid-connected AC voltage—the performance of the internal conversion links of the PV system and its interaction with the grid can be reflected. Fluctuations in these parameters also significantly affect the overall performance of the PV system. The explicit and specific designation of these parameters makes the construction of the original dataset more targeted and complete. In subsequent steps, such as calculating the correlation between environmental parameters and photovoltaic output power, these specific and key environmental parameters can more accurately identify the factors that have the greatest impact on photovoltaic output power fluctuations. When constructing the initial time-series fluctuation feature set of the volatility index, using these specific photovoltaic output power sequences and key environmental parameter sequences as input ensures that the constructed feature set can comprehensively and meticulously characterize the multi-dimensional characteristics of photovoltaic performance fluctuations. This specific limitation on parameter types avoids analytical biases caused by improper parameter selection or missing information, thus enabling the entire fluctuation feature mining method to more effectively reveal the deep mechanisms of distributed photovoltaic performance fluctuations and provide a more reliable basis for subsequent fluctuation prediction, control, and management.

[0113] In practical distributed photovoltaic (PV) power plants, a series of sensors and data acquisition devices can be deployed to obtain the aforementioned parameters. For example, the acquisition of PV module output parameters can be achieved by installing DC voltage and DC current sensors at the output end of each PV string. These sensors transmit real-time data to the data acquisition unit, which then calculates the output power. Environmental parameters can be acquired by setting up a small weather station within the power plant, including a high-precision solar intensity meter to measure solar radiation intensity, a thermistor temperature sensor to measure ambient temperature, and a cup anemometer to measure wind speed. This environmental data is also transmitted to the data acquisition unit in real time. For operating status parameters, inverter efficiency can be obtained by reading the inverter's internal operating logs or through its communication interface. Typically, the inverter provides real-time input DC power and output AC power data, allowing efficiency calculation. Grid connection point AC voltage can be monitored in real time by installing voltage transformers and smart meters at the grid connection cabinet. All acquired data is aggregated to a central data server via wired or wireless communication to form a raw dataset for subsequent fluctuation feature mining algorithms.

[0114] As a feasible specific implementation of the embodiments of this application, some of the above-described embodiments of this application propose to calculate the correlation between environmental parameters and photovoltaic output power in order to determine key influencing factors. However, in actual operation, how to accurately and quantitatively assess the degree of correlation between different environmental parameters and photovoltaic output power, and objectively screen out the key factors that truly have a significant impact, may face problems such as inappropriate method selection or vague evaluation standards, thereby affecting the accuracy and effectiveness of subsequent fluctuation feature mining.

[0115] In this regard, this application further proposes the following steps for determining key influencing factors: constructing a correlation matrix based on environmental parameters using the Pearson correlation coefficient; calculating the absolute value of the correlation degree of each environmental parameter; and determining the key influencing factors based on the absolute value of the correlation degree of each environmental parameter.

[0116] When constructing an association matrix based on environmental parameters using the Pearson correlation coefficient, the Pearson correlation coefficient is a widely used statistical indicator for measuring the degree of linear correlation between two variables, with a value ranging from -1 to 1. The larger the absolute value of the coefficient, the stronger the linear correlation between the two variables. By calculating the Pearson correlation coefficient between each environmental parameter (e.g., light intensity, ambient temperature, wind speed, etc.) and photovoltaic output power, these coefficients can be organized into an association matrix for subsequent unified analysis and processing. For example, the Pearson correlation coefficient between each environmental parameter and photovoltaic output power can be calculated separately, and these coefficients can be arranged into a row vector or column vector to form an association matrix. Alternatively, the correlation coefficients between all pairs of environmental parameters and between environmental parameters and photovoltaic output power can be calculated to form a symmetric matrix containing information on the linear correlation between all variables.

[0117] When calculating the absolute value of the correlation between various environmental parameters, since the Pearson correlation coefficient can be positive (positive correlation) or negative (negative correlation), the absolute value of the calculated Pearson correlation coefficient needs to be taken to uniformly measure the strength of the correlation without considering its direction. This ensures that the correlation strength with photovoltaic output power can be fairly assessed regardless of whether the impact is positive or negative. Besides directly taking the absolute value of the Pearson correlation coefficient, other statistical methods, such as Spearman's rank correlation coefficient or mutual information, can be used to calculate the correlation and then take its absolute value to address potential nonlinear relationships or more complex dependencies.

[0118] When determining key influencing factors based on the absolute values ​​of the correlation degrees of various environmental parameters, key environmental parameters that significantly affect photovoltaic output power can be identified by comparing and filtering the absolute values ​​of the correlation degrees of all environmental parameters. This can be achieved by setting a preset threshold for the absolute values ​​of the correlation degrees, and identifying all environmental parameters whose absolute values ​​of the correlation degrees exceed this threshold as key influencing factors. Alternatively, all environmental parameters can be sorted in descending order of their absolute values ​​of correlation degrees, and the top N (where N is a preset value) environmental parameters with the largest absolute values ​​of correlation degrees can be selected as key influencing factors.

[0119] Specifically, the core driving factor for photovoltaic power output fluctuations is environmental parameters. The linear correlation between three environmental parameters—illuminance G, ambient temperature T, and wind speed v—and photovoltaic output power P is calculated. The correlation matrix R=[rpq] is constructed using the Pearson correlation coefficient, as shown in equation (5). The Pearson correlation coefficient ranges from [-1, 1]. The closer the absolute value is to 1, the stronger the correlation; a value close to 0 indicates a weak correlation.

[0120]

[0121] Where p represents the environmental parameter dimension, with 1-3 corresponding to G, T, and v respectively; q represents the output power dimension, which is fixed as the 4th column. , These are the mean values ​​of the corresponding standardized parameters. Figure 4 This is a histogram showing the correlation coefficient distribution between photovoltaic (PV) power output and different factors. Based on engineering experience, environmental parameters with an absolute correlation value ≥ 0.7 are defined as key influencing factors. Typically, solar irradiance (G) has the highest correlation and is the core driving factor; ambient temperature (T) is second; and wind speed (v) has the lowest correlation and is not a key factor. Subsequent feature construction focuses on the core driving factors, and the output results are the environment-power output correlation matrix R = [Rpq] and a list of key influencing factors.

[0122] This application's approach introduces the Pearson correlation coefficient as a tool to quantify correlation strength and further calculates its absolute value, providing an objective and quantitative method for identifying key environmental influencing factors. This method enables the accurate assessment of the linear correlation strength between different environmental parameters and photovoltaic output power after collecting output parameters, environmental parameters, and operating status parameters of photovoltaic modules, constructing an original dataset, and preprocessing the original dataset to obtain a preprocessed dataset. By screening these quantitative correlations, it ensures that when subsequently using photovoltaic output power sequences and key environmental parameter sequences as input to construct an initial time-series fluctuation feature set of volatility indicators based on a sliding event window, the selected key environmental parameters are factors that truly have a significant impact on photovoltaic output power fluctuations. This not only improves the accuracy and reliability of identifying key influencing factors but also lays a solid foundation for subsequent fluctuation feature mining and analysis, avoiding the accumulation of errors caused by subjective judgments or fuzzy standards.

[0123] The above technical solution provides a clear and quantitative method to identify environmental parameters that significantly impact the performance fluctuations of distributed photovoltaic (PV) systems. This effectively addresses the potential ambiguity and subjectivity in determining key influencing factors, ensuring that the selected key environmental parameters are highly relevant. Therefore, in the subsequent fluctuation feature mining process, the fluctuation characteristics of PV output power can be captured more accurately, thereby improving the accuracy and reliability of the entire distributed PV performance fluctuation feature mining method and providing a more solid data foundation for PV system operation optimization and fault diagnosis.

[0124] As a feasible specific implementation of this application, in the distributed photovoltaic performance fluctuation feature mining method, constructing an initial time-series fluctuation feature set of fluctuation indices based on a sliding event window, using the photovoltaic output power sequence and key environmental parameter sequence as input, is a fundamental step in understanding the dynamic behavior of the photovoltaic system. However, if the specific composition of the fluctuation indices is not fully defined and quantified, the constructed feature set may fail to comprehensively and meticulously reflect the multi-dimensional characteristics of photovoltaic performance fluctuations, thereby affecting the accuracy and effectiveness of subsequent feature dimensionality reduction, core feature extraction, and quantitative index calculation.

[0125] In response, this application further proposes an initial time-series volatility feature set for constructing volatility indices, including six basic features: window mean, volatility standard deviation, peak factor, trend slope, environment-power coupling coefficient, and volatility index.

[0126] The window mean refers to the average value of photovoltaic output power or key environmental parameters within a set sliding event window. Its function is to characterize the overall level or central trend of data within a specific time period, reflecting the average performance of the photovoltaic system under different operating conditions. The standard deviation of fluctuation refers to the dispersion of photovoltaic output power or key environmental parameters within a set sliding event window. It quantifies the fluctuation amplitude of data points relative to their mean and is a commonly used indicator to measure the severity of fluctuations. The peak factor is the ratio of the peak value to the effective value (such as the root mean square value) of the data sequence within a set sliding event window. This feature is used to characterize whether there are prominent peaks or instantaneous high values ​​in the data sequence, reflecting the possibility of extreme fluctuations in photovoltaic output power within a short period. The trend slope refers to the changing trend of photovoltaic output power or key environmental parameters within a set sliding event window. It reflects the rising, falling, or stable state of the data within the window period, as well as the rate of change. The environment-power coupling coefficient refers to the degree of mutual influence between key environmental parameters and photovoltaic output power within a set sliding event window. This coefficient is used to quantify the direct or indirect driving effect of environmental factors on photovoltaic power fluctuations, revealing the correlation between the two. For example, it can be characterized by calculating the Pearson correlation coefficient between the environmental parameter series and the photovoltaic output power series within the window, or by using methods such as mutual information and Granger causality to assess the degree of coupling. The volatility index is a comprehensive measure used to broadly describe the overall volatility characteristics of photovoltaic output power within a defined sliding event window. It may be a composite index based on multiple fundamental statistics, aiming to provide a single, easily understood quantification of the degree of volatility. For example, it could be a normalized index based on the difference between the maximum and minimum values ​​within the window, or a weighted average combining standard deviation and mean absolute deviation.

[0127] The proposed solution concretizes the initial time-series fluctuation feature set into six basic features: window mean, fluctuation standard deviation, peak factor, trend slope, environment-power coupling coefficient, and volatility index. This allows for a comprehensive capture of the fluctuation characteristics of photovoltaic output power from multiple dimensions. These features complement each other, forming a multi-level, multi-angle fluctuation feature description system. This enables subsequent variational mode decomposition to more accurately separate fluctuation components of different frequencies, principal component analysis to more effectively extract core features, and ultimately achieves precise calculation of quantitative indicators such as fluctuation intensity, fluctuation frequency, fluctuation persistence, and fluctuation stability. Consequently, it significantly improves the depth and breadth of mining the fluctuation characteristics of distributed photovoltaic performance.

[0128] As a possible specific implementation of this application, when constructing the initial time-series volatility feature set of the volatility index, a fixed-length sliding event window can be set, such as 15 minutes or 30 minutes. As the window slides across the photovoltaic output power sequence and the key environmental parameter sequence, the aforementioned six basic features are calculated for each window. Specifically, the window mean can be obtained by calculating the arithmetic mean of all photovoltaic output power data points within the window; the volatility standard deviation can be obtained by calculating the standard deviation of the photovoltaic output power data points within the window relative to their mean; the peak factor can be calculated by dividing the maximum value of the photovoltaic output power within the window by its root mean square value; the trend slope can be determined by performing a least-squares linear regression on the photovoltaic output power data within the window and taking the slope of the regression line; the environment-power coupling coefficient can be obtained by calculating the Pearson correlation coefficient between the photovoltaic output power sequence and the key environmental parameter sequence (e.g., the irradiance sequence) within the window; and the volatility index can be defined as the difference between the maximum and minimum values ​​of the photovoltaic output power within the window divided by the window mean. In this way, a vector containing these six features can be generated for each sliding window, thereby constructing the initial time-series volatility feature set.

[0129] Specifically, the input data consists of the photovoltaic output power sequence and the determined key environmental parameter sequence from the standardized dataset Z;

[0130] Photovoltaic fluctuations exhibit significant temporal correlations. A sliding time window method is employed to construct features, effectively capturing the fluctuation characteristics across different time scales. Considering engineering requirements, and to balance local fluctuations with computational efficiency, the window size is set to 15 minutes; to ensure feature continuity, the step size is set to 5 minutes. The specific feature construction is as follows:

[0131] Based on a sliding time window with a window size of 15 minutes and a step size of 5 minutes, features are constructed to comprehensively characterize the fluctuation characteristics of photovoltaic power output, taking into account dimensions such as average level, dispersion, extreme potential, changing trend, environmental correlation, and fluctuation frequency. This includes the following steps:

[0132] Step S1041: Obtain the window mean fi1

[0133] The average level of photovoltaic output within the current window is the core parameter for judging fluctuations, as shown in the following formula. Changes in the window mean can directly reflect the overall trend of photovoltaic output; for example, the window mean remains high during midday and gradually decreases in the evening.

[0134]

[0135] Step S1042: Obtain the standard deviation of fluctuation fi2

[0136] The standard deviation reflects the dispersion of photovoltaic power output within the window and is a direct indicator of the intensity of fluctuations, as shown in the following formula. The larger the standard deviation, the more drastic the power output fluctuations within the window. For example, the standard deviation will increase significantly when clouds are moving rapidly; under stable conditions in clear weather, the standard deviation will be smaller.

[0137]

[0138] Step S1043: Obtain the peak factor fi3

[0139] The ratio of the maximum power within the window to the average power within the window reflects the extreme fluctuation potential of photovoltaic output, as shown in the following formula. When the peak factor is greater than 1.5, it indicates that there is a significant power spike within the window, which may be caused by instantaneous strong light, such as direct sunlight through gaps in clouds, and attention should be paid to the impact on the inverter.

[0140]

[0141] Step S1044: Obtain the trend slope fi4

[0142] The slope of the linear regression of the power series within the window quantitatively reflects the trend of fluctuation, as shown in the following formula. A positive slope with a large absolute value indicates a rapid upward trend in power output; a negative slope indicates a downward trend; and a slope close to 0 indicates stable power output.

[0143]

[0144] Step S1045: Obtain the environment-power coupling coefficient fi5

[0145] The covariance between illuminance and photovoltaic power within the window reflects the driving effect of environmental factors on photovoltaic fluctuations, as shown in the following formula. A positive coupling coefficient indicates that illuminance and power change in the same direction, which conforms to the physical laws of photovoltaic power generation; if the coupling coefficient is negative or close to 0, it may indicate that there is a fault in the photovoltaic module, such as dust accumulation leading to a decrease in light absorption efficiency.

[0146]

[0147] Where μ is the mean light intensity within the current window, which is calculated by traversing the Z sequence within the window to ensure the accuracy of covariance calculation.

[0148] Step S1046: Obtain the volatility indicator fi,6

[0149] The sum of the absolute values ​​of power changes at adjacent moments within a window reflects the frequency of photovoltaic power output fluctuations, as shown in the following formula. The larger this index is, the more frequent the power changes within the window. For example, during cloudy weather, continuous cloud cover causes frequent power fluctuations, resulting in a significantly higher volatility index.

[0150]

[0151] The output is the initial time series fluctuation feature set F=[fik], where k=1-6 correspond to six basic features: window mean, fluctuation standard deviation, peak factor, trend slope, environment-power coupling coefficient, and volatility index. Each feature has a clear physical meaning.

[0152] The aforementioned technical solution concretizes the initial time-series fluctuation feature set into six basic features: window mean, fluctuation standard deviation, peak factor, trend slope, environment-power coupling coefficient, and volatility index. This overcomes the limitation that relying on a single or few indicators is insufficient to comprehensively characterize the complexity of photovoltaic performance fluctuations. These multi-dimensional features can more precisely capture the fluctuation patterns of photovoltaic output power under different time scales and influencing factors. For example, the window mean reflects the average level, the fluctuation standard deviation reflects the degree of dispersion, the peak factor reveals extreme events, the trend slope captures the direction of change, the environment-power coupling coefficient quantifies external drivers, and the volatility index provides a comprehensive assessment. This comprehensive feature description provides richer and more accurate input for subsequent variational mode decomposition, enabling the decomposed intrinsic mode components to more realistically reflect the fluctuation components at different frequencies. Simultaneously, it provides a more solid foundation for feature dimensionality reduction and core feature extraction based on principal component analysis, ensuring that the extracted core features can more effectively represent the essence of photovoltaic performance fluctuations, thereby significantly improving the accuracy, comprehensiveness, and reliability of distributed photovoltaic performance fluctuation feature mining.

[0153] As a feasible specific implementation of this application, the fluctuation standard deviation sequence is decomposed into three types of intrinsic mode components (high-frequency, mid-frequency, and low-frequency) and residual trend components through variational mode decomposition. This application further proposes that the decomposition of the fluctuation standard deviation sequence through variational mode decomposition includes: constructing a variational constraint model; setting the fluctuation periods of high-frequency, mid-frequency, and low-frequency; setting the number of decomposition modes, penalty factor, and iteration stopping threshold, wherein the number of decomposition modes is 3; solving the variational model using the alternating direction multiplier algorithm, and iteratively updating the intrinsic mode components and center frequency until the iteration stopping threshold is met; and outputting the three types of intrinsic mode components (high-frequency, mid-frequency, and low-frequency) and residual trend components.

[0154] This application's approach first constructs a variational constraint model, transforming the complex signal decomposition problem into an optimizable mathematical framework. Then, by setting the periods of high-frequency, mid-frequency, and low-frequency fluctuations, as well as the number of decomposition modes, penalty factor, and iteration stopping threshold—specifically, explicitly setting the number of decomposition modes to 3—this method can effectively decompose the photovoltaic fluctuation standard deviation sequence into three physically meaningful main frequency components: high-frequency, mid-frequency, and low-frequency fluctuations. This pre-setting and parameterization ensures the validity and interpretability of the decomposition results. Furthermore, the variational model is solved using the alternating direction multiplier algorithm, and the intrinsic mode components and center frequencies are iteratively updated until the iteration stopping threshold is met, ensuring the convergence and accuracy of the decomposition process. The introduction of the ADMM algorithm enables the efficient finding of optimal mode components while satisfying the constraints, thereby accurately separating the fluctuation characteristics of different frequencies. Finally, the high-frequency, mid-frequency, and low-frequency intrinsic mode components, along with residual trend components, are output, providing refined and physically meaningful input for subsequent construction of expanded feature sets. This refined decomposition process makes it possible to analyze the multi-scale characteristics of photovoltaic performance fluctuations, thereby solving the problems of feature confusion and inaccurate analysis that may be caused by simple decomposition.

[0155] Specifically, the fluctuation trend separation process based on variational mode decomposition uses the fluctuation standard deviation sequence f from the initial time-series fluctuation feature set as input data.

[0156] Photovoltaic fluctuations are a superposition of fluctuations at different scales, such as instantaneous noise, cloud cover, and diurnal variations. Traditional EMD algorithms suffer from mode aliasing and cannot effectively separate these fluctuations. Variational Mode Decomposition (EMD) algorithms, by pre-setting the number of modes and constraints, can achieve accurate separation of multi-scale fluctuations. The steps include:

[0157] Step S1051: Construction of variational mode decomposition model:

[0158] The core of variational mode decomposition (VMD) is to construct a variational constraint model. By iteratively searching for the optimal solution, the wave sequence is decomposed into multiple intrinsic mode components (IMFs) with sparse characteristics. Each IMF corresponds to a center frequency, ensuring that waves at different scales are accurately separated. The variational constraint model of VMD is shown in the following equations:

[0159] The objective function is to minimize the sum of the bandwidths of all IMFs; the constraint condition is to ensure that the sum of all IMFs equals the original volatility sequence.

[0160]

[0161]

[0162] Where {uk} is the decomposed IMF component set, k=1,2,3; {wk} is the center frequency of each IMF; δ(t) is the Dirac function; * is the convolution operation; j is the imaginary unit; ∂ represents the partial derivative with respect to time t; ||·|| represents the L2 norm.

[0163] Step S1052: Decomposition Parameter Setting and Solution

[0164] Based on the characteristics of photovoltaic fluctuations, the high-frequency fluctuation period was set to 1-5 minutes, the mid-frequency to 15-60 minutes, and the low-frequency to several hours to one day. Parameters were optimized through extensive experiments: the number of decomposition modes K=3, corresponding to the high, mid, and low-scale fluctuations respectively; the penalty factor α=2000 to balance bandwidth and reconstruction error; and the iteration stopping threshold ε=10 to ensure decomposition accuracy. The Alternating Direction Multiplier (ADMM) algorithm was used to solve the variational model, iteratively updating the IMF components and center frequency until the stopping condition was met, ultimately outputting three IMF components and one residual trend component Res.

[0165] The output consists of three intrinsic mode components at different scales and one residual trend component, denoted as... Among them, IMF1 corresponds to high-frequency random fluctuations, IMF2 corresponds to medium-frequency periodic fluctuations, IMF3 corresponds to low-frequency trend fluctuations, and Res corresponds to long-term trends.

[0166] By employing the aforementioned technical solution, variational mode decomposition of the fluctuation standard deviation sequence can ensure the accuracy and effectiveness of the decomposition process through a clear variational constraint model, precise parameter settings, and efficient solution algorithms. This refined decomposition allows the fluctuation characteristics of photovoltaic output power to be clearly divided into three categories of intrinsic mode components: high-frequency, mid-frequency, and low-frequency, thereby avoiding confusion between fluctuation characteristics of different frequencies. This not only improves the accuracy of feature extraction but also provides a more reliable foundation for subsequent calculations of quantitative indicators such as fluctuation intensity, fluctuation frequency, fluctuation persistence, and fluctuation stability based on these components, thus enhancing the overall accuracy and depth of distributed photovoltaic performance fluctuation feature mining.

[0167] As a feasible specific implementation of this application, a scheme for feature dimensionality reduction and core feature extraction of the expanded feature set based on principal component analysis is further proposed. The scheme includes: calculating the covariance matrix of the expanded feature set; performing eigenvalue decomposition on the covariance matrix to obtain 9 eigenvalues ​​and corresponding unit eigenvectors; calculating the cumulative contribution rate of each eigenvalue; setting a cumulative contribution rate threshold and selecting the number of principal components s, and selecting the first s principal components that exceed the cumulative contribution rate threshold as core features; multiplying the expanded feature set with the matrix formed by the first s eigenvectors to obtain the core feature set.

[0168] Specifically, based on principal component analysis (PCA) for feature dimensionality reduction and core feature extraction, the input data consists of the initial time-series fluctuation feature set F=[fik] output in step 103 and the three IMF components after VMD decomposition. The initial feature set and IMF components, after merging, have high dimensionality, and there may be redundancy such as strong correlation between fluctuation standard deviations and IMF components, increasing the computational cost of subsequent models. PCA is used to map high-dimensional features to a low-dimensional space through linear transformation, retaining the principal components with high information content, thus achieving feature dimensionality reduction. The steps include:

[0169] Step S1051: Feature Expansion

[0170] The three IMF components obtained from VMD decomposition are used as new features and merged column-wise with the initial time series fluctuation feature set F to form an expanded feature set. This ensures that all features participate in the dimensionality reduction analysis.

[0171] Step S1052 Principal Component Analysis Dimensionality Reduction:

[0172] First, calculate the expanded feature set. The covariance matrix C reflects the degree of linear correlation between features, as shown in the following formula; then, the covariance matrix C is decomposed into eigenvalues ​​to obtain 9 eigenvalues ​​and corresponding unit eigenvectors. The larger the eigenvalue, the more information the corresponding principal component contains.

[0173]

[0174] in for The mean vector of F is obtained by calculating the mean of each column of F; This is the decentralized feature matrix, ensuring that the covariance calculation is not affected by the feature mean.

[0175] Step S1053 Core Feature Selection:

[0176] Calculate the cumulative contribution rate ηCon of each eigenvalue. The cumulative contribution rate reflects the proportion of total information contained in the first n principal components, as shown in the following formula. Based on engineering requirements, select the first s principal components with a cumulative contribution rate ≥ 85% as core features. Multiply the expanded feature set F with the matrix formed by the first s eigenvectors to obtain the core feature set, thus achieving feature dimensionality reduction.

[0177]

[0178] The output is the core fluctuation feature set F after dimensionality reduction, with a dimension of m×s (s≤9), a cumulative contribution rate of ≥85%, and retains the core information of the original features.

[0179] The above technical solution effectively solves the problems of excessively high dimensionality, feature redundancy, and potential noise interference in the expanded feature set by performing feature dimensionality reduction and core feature extraction based on principal component analysis. This allows subsequent calculations of the four quantitative indicators—fluctuation intensity, fluctuation frequency, fluctuation persistence, and fluctuation stability—to be based on a more concise, efficient, and information-rich core feature set. Therefore, it not only significantly reduces computational complexity and storage requirements, improving data processing efficiency, but also enhances the representativeness and discriminative power of the features by removing redundancy and noise. This results in more accurate and reliable quantitative analysis of the fluctuation characteristics of distributed photovoltaic performance, and helps to more clearly reveal the inherent laws governing photovoltaic performance fluctuations.

[0180] As a feasible specific implementation of this application, fluctuation characteristics are quantified and dynamic evaluation indicators are calculated based on the obtained core fluctuation feature set F. The core feature set is high-dimensional data and needs to be further quantified into intuitive engineering indicators to facilitate understanding and use by operation and maintenance and scheduling personnel. Combining the needs of photovoltaic performance evaluation, four types of quantitative indicators are designed, covering the intensity, frequency, persistence, and stability of fluctuations. Specific calculations include:

[0181] Wave intensity S1:

[0182] The global variance of the core feature set, reflecting the overall fluctuation amplitude, is as follows:

[0183]

[0184] Where f is the element in the i-th row and t-th column of F. Let be the mean of the t-th core feature.

[0185] Fluctuation frequency S2:

[0186] The dominant frequency of the high-frequency IMF component reflects the frequency of fluctuations and is calculated using a Fast Fourier Transform (FFT), as shown in the following formula:

[0187]

[0188] Where FFT(·) represents Fast Fourier Transform, |·| represents amplitude operation, and argmax represents taking the frequency f corresponding to the maximum amplitude.

[0189] Fluctuation persistence S3:

[0190] The autocorrelation coefficient of the core feature sequence reflects the time-duration characteristics of fluctuations, as shown in the following formula:

[0191]

[0192] Fluctuation stability S4:

[0193] The coefficient of variation of fluctuation intensity based on the sliding window reflects the dynamic stability of the fluctuation, as shown in the following formula:

[0194]

[0195] Where S4 is the fluctuation intensity sequence within the sliding window, σ(·) is the standard deviation, and μ(·) is the mean.

[0196] The output is a set of quantified photovoltaic performance fluctuation characteristic indicators. These correspond to four core indicators: fluctuation intensity, fluctuation frequency, fluctuation duration, and fluctuation stability. Each indicator has a specific value that can be directly used for engineering decision-making.

[0197] As a possible specific implementation of this application, a further step is proposed to verify the effectiveness of a feature, including: using a dual-indicator verification of feature discrimination and prediction contribution. When both feature discrimination and prediction contribution are not less than a threshold, the feature is deemed effective. Feature discrimination refers to the ability of the extracted feature to exhibit differences under different operating conditions or states. For example, a feature with high discrimination will show significant differences in its numerical distribution when the photovoltaic system is operating normally and experiencing a fault. Methods for achieving feature discrimination may include, but are not limited to: comparing the mean or variance of features under different operating conditions through statistical hypothesis testing (such as t-tests or analysis of variance); or quantifying the degree of difference by calculating the distance metric between feature distributions under different operating conditions (such as KL divergence or JS divergence). Prediction contribution refers to the effectiveness of the extracted feature in improving the performance of the photovoltaic output power prediction model. A feature with high prediction contribution, when introduced into the prediction model, can significantly reduce the model's prediction error. Methods for assessing predictive contribution can include, but are not limited to: comparing the reduction in prediction error (e.g., root mean square error (RMSE) or mean absolute error (MAE) of the prediction model with and without the feature; or using model-based methods (e.g., feature importance scoring using random forests or gradient boosting trees) to evaluate the influence of a feature on the model output. Dual-metric validation refers to using both feature discrimination and predictive contribution to comprehensively evaluate the effectiveness of a feature. This validation method aims to ensure that the selected feature is not only statistically discriminative but also has a positive impact on the prediction of key performance indicators in practical applications. A threshold is a pre-set critical value used to determine whether feature discrimination and predictive contribution have reached acceptable levels. A feature is considered effective only when the calculated result of feature discrimination or predictive contribution is not less than its respective threshold. This threshold can be set based on practical application needs, historical data analysis, or expert experience to balance the rigor and practicality of the feature. Determining feature effectiveness means ultimately determining whether a feature is considered effective based on the comparison results of feature discrimination and predictive contribution with their respective thresholds. Typically, a feature is considered valid and included in the final fluctuation feature set only when both indicators meet the condition of being no less than a threshold.

[0198] In the method for mining the performance fluctuation features of distributed photovoltaics (PV), to ensure the practical application value and reliability of the extracted fluctuation feature set, this method introduces a feature validity verification step after calculating four quantitative indicators: fluctuation intensity, fluctuation frequency, fluctuation persistence, and fluctuation stability. This verification step employs a dual-indicator verification mechanism of feature discrimination and prediction contribution to rigorously screen each feature in the core feature set. Specifically, feature discrimination is used to evaluate the feature's ability to distinguish between different operating conditions (e.g., normal operation and fault state), ensuring that the selected features can effectively distinguish different states of the PV system, providing a basis for fault diagnosis and anomaly detection. Simultaneously, prediction contribution is used to quantify the feature's improvement effect on the performance of the PV output power prediction model, ensuring that the selected features can significantly reduce prediction errors in actual prediction tasks, thereby improving prediction accuracy and reliability. By setting reasonable thresholds and requiring that both feature discrimination and prediction contribution are not less than their respective thresholds, this method can screen out effective features that possess both good discrimination ability and practical contribution to the prediction task. This dual verification mechanism avoids the bias that a single indicator might bring. For example, a feature may be statistically discriminative but not very helpful for actual prediction, or it may contribute to prediction but fail to effectively distinguish different operating conditions. Therefore, this verification mechanism ensures that the final output set of fluctuation features is comprehensive, reliable, and has practical application value, thereby significantly improving the practicality and accuracy of the entire distributed photovoltaic performance fluctuation feature mining method.

[0199] This application provides a specific method for verifying and outputting feature validity. The input data includes a quantitative feature index set S, historical fault records of photovoltaic power plants, and power prediction error data. The verified distributed photovoltaic performance fluctuation feature set and feature-performance correlation report are output. Verification is performed using a dual-indicator approach: "feature discrimination" and "prediction contribution." Feature discrimination is achieved through the difference between features under fault and normal operating conditions, as shown in the following formula;

[0200]

[0201] The predicted contribution is achieved through the error reduction rate of feature-based power prediction, as shown in the following formula:

[0202]

[0203] in and σ represents the mean value of feature S under fault and normal operating conditions, respectively; σ is the corresponding standard deviation; MAE is the mean absolute error of power prediction without core features, and MAE is the prediction error after adding core features. When D≥1.5 and E≥10%, the feature effectiveness meets the standard.

[0204] By combining the above output verification of the fluctuation feature set, and generating a correlation report between the features and photovoltaic performance (power generation efficiency, failure risk), the physical meaning and engineering application value of each feature are clarified.

[0205] In some other embodiments, this application further proposes that feature distinguishability is achieved by the difference between features under fault and normal operating conditions, and prediction contribution is achieved by the error reduction rate of feature participation in power prediction.

[0206] This application's solution provides a clear and quantifiable standard for validating the effectiveness of distributed photovoltaic performance fluctuation features by defining feature discrimination as the difference between features under fault and normal operating conditions, and prediction contribution as the error reduction rate of the feature's participation in power prediction. After feature dimensionality reduction and core feature extraction of the expanded feature set, effectiveness verification is necessary to ensure that the extracted core features truly have practical application value. Measuring feature discrimination by difference directly assesses the feature's ability to identify system anomalies or faults; a good fluctuation feature should exhibit significant differences when the system state changes. Simultaneously, measuring prediction contribution by error reduction rate evaluates the feature's practicality from another dimension—whether the feature can effectively improve the accuracy of photovoltaic power prediction. The combined use of these two indicators ensures that the feature verification process focuses not only on the feature's ability to represent the system state but also on its gain effect on the actual prediction task. Only when both indicators reach preset thresholds can the feature be ultimately determined to be effective, thus avoiding the selection of features that, while statistically significant, have little practical application value, ensuring that the final output fluctuation feature set has higher reliability and practicality. This dual verification mechanism enables the entire feature mining method to filter out truly meaningful fluctuation features, providing high-quality input for subsequent photovoltaic performance analysis and prediction.

[0207] The above technical solutions clarify the specific quantitative methods for feature discrimination and prediction contribution, making the feature effectiveness verification process more objective and operable. Defining feature discrimination as the difference between features under fault and normal operating conditions ensures that the selected features have the ability to identify system anomalies or faults, thereby improving the accuracy of photovoltaic system operation status monitoring and diagnosis. Simultaneously, defining prediction contribution as the error reduction rate of features participating in power prediction directly measures the actual gain of features on photovoltaic power prediction accuracy, avoiding the introduction of redundant features that are detrimental to prediction performance. This dual quantitative verification mechanism ensures that the final selected fluctuation feature set not only effectively characterizes photovoltaic performance fluctuations but also significantly improves the reliability and accuracy of subsequent analysis and prediction tasks, providing a high-quality data foundation for the refined management and optimized operation of distributed photovoltaic systems.

[0208] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A distributed photovoltaic performance fluctuation feature mining method, characterized in that, The method comprises the following steps: Collecting output parameters, environmental parameters and operating state parameters of a photovoltaic module to construct an original data set; Preprocessing the original data set to obtain a preprocessed data set; Calculating the correlation degree of environmental parameters and photovoltaic output power to determine key influencing factors; Taking the photovoltaic output power sequence and the key environmental parameter sequence as inputs, and based on a sliding event window, an initial time sequence fluctuation feature set of the fluctuation index is constructed; Taking the fluctuation standard deviation sequence in the initial feature set as input, the fluctuation standard deviation sequence is decomposed into three types of intrinsic mode components and a residual trend component by variational mode decomposition; The three types of intrinsic mode components obtained are combined with the initial feature set to form an expanded feature set; Based on principal component analysis, the expanded feature set is subjected to feature dimension reduction and core feature extraction; Based on the core feature set, four types of quantitative indexes of fluctuation intensity, fluctuation frequency, fluctuation persistence and fluctuation stability are calculated; Verify the effectiveness of the features; Output the fluctuation feature set that meets the verification and the report on the correlation between the features and the performance of the photovoltaic module.

2. The method of claim 1, wherein, The preprocessing of the original data set comprises: Based on the 3σ criterion combined with trend consistency, the outliers in the original data set are removed; The missing values are filled by an improved adjacent weighted completion algorithm.

3. The method of claim 1, wherein, The preprocessed data set is standardized, and the initial time sequence fluctuation feature set is constructed by inputting the standardized photovoltaic output power sequence and the key environmental parameter sequence.

4. The method of claim 1, wherein, The photovoltaic module output parameters include direct current voltage, direct current, and output power; the environmental parameters include light intensity, environmental temperature, and wind speed; and the operating state parameters include inverter efficiency and grid-connected point alternating current voltage.

5. The method of claim 1, wherein, The determination of the key influencing factors comprises: Based on the environmental parameters, a correlation matrix is constructed using the Pearson correlation coefficient; The absolute value of the correlation degree of each environmental parameter is calculated; According to the absolute value of the correlation degree of each environmental parameter, the key influencing factors are determined.

6. The method of claim 1, wherein, The initial time sequence fluctuation feature set of the fluctuation index comprises six types of basic features, including window mean, fluctuation standard deviation, peak factor, trend slope, environmental-power coupling coefficient and fluctuation index.

7. The method of claim 1, wherein, The decomposition of the fluctuation standard deviation sequence by variational mode decomposition comprises: A variational constraint model is constructed; The high-frequency, medium-frequency and low-frequency fluctuation periods are set; The decomposition mode number, the penalty factor and the iteration stop threshold are set, and the decomposition mode number is 3; The variational model is solved by using the alternating direction multiplier algorithm, and the intrinsic mode component and the center frequency are updated by iteration until the iteration stop threshold is met; The high-frequency, medium-frequency and low-frequency intrinsic mode components and the residual trend component are output.

8. The method of claim 1, wherein, The feature dimension reduction and core feature extraction of the expanded feature set based on principal component analysis comprise: The covariance matrix of the expanded feature set is calculated; The covariance matrix is subjected to eigenvalue decomposition to obtain 9 eigenvalues and corresponding unit eigenvectors; The cumulative contribution rate of each eigenvalue is calculated; The cumulative contribution rate threshold and the selected number of principal components s are set, and the first s principal components exceeding the cumulative contribution rate threshold are selected as the core features; The core feature set is obtained by multiplying the expanded feature set and the matrix formed by the first s eigenvectors.

9. The method of claim 1, wherein, The verification of the effectiveness of the features comprises: The feature discrimination degree and the prediction contribution degree are verified by using two indexes, and when the feature discrimination degree and the prediction contribution degree are not less than threshold values, the feature is determined to be effective.

10. The method of claim 9, wherein, The feature discrimination degree is realized by the difference degree of the feature under the fault and normal working conditions, and the prediction contribution degree is realized by the error reduction rate of the feature participating in power prediction.

Images