CN-122001295-A - Photovoltaic output abnormality detection method and system based on multi-feature self-adaptive hypersphere

CN122001295ACN 122001295 ACN122001295 ACN 122001295ACN-122001295-A

Abstract

The application relates to a photovoltaic output abnormality detection method and system based on a multi-feature self-adaptive hypersphere, belongs to the technical field of photovoltaic system abnormality detection, and aims to solve the problems that an existing detection method is single in feature dimension, not suitable for a timing working condition and lacks a dynamic judgment mechanism. The method comprises the steps of constructing a multi-dimensional feature vector, dividing the whole day into 24 independent time periods, respectively constructing a super-sphere model for the independent time periods meeting construction conditions, implementing three-level judgment, calculating a basic abnormality score, correcting the score by combining with photovoltaic operation experience, dynamically adjusting a threshold value, and filtering extremely low power to finish judgment. The method realizes the fusion of multidimensional features, adapts to the output dynamic change rules of different time periods, and realizes the optimal balance of false alarm rate and false alarm rate.

Inventors

SUN WENCHUAN
HAN LIN
CHEN YU
JIANG RUI
YU SHANZHI
WANG JUN
CHEN HONGDA
FANG HAITENG

Assignees

国网山东省电力公司临沂供电公司

Dates

Publication Date: 20260508
Application Date: 20251229

Claims (10)

1. The photovoltaic output abnormality detection method based on the multi-feature self-adaptive hypersphere is characterized by comprising the following steps of: step S1, constructing a multi-dimensional feature vector, namely extracting multi-dimensional features of the original photovoltaic output time sequence based on the original photovoltaic output time sequence, and constructing the multi-dimensional feature vector; Step S2, a step of constructing a time-interval self-adaptive hypersphere model, wherein the whole day is divided into 24 independent time intervals, the hypersphere model is constructed for the independent time intervals meeting the construction conditions, and if training data points of the current time interval are used Constructing a hypersphere model for the corresponding independent time period according to the construction conditions; And S3, an abnormality judgment step, namely returning an abnormal value score for an independent period corresponding to the unsuccessful construction of the hypersphere model, judging that the corresponding independent period is abnormal, and judging whether the abnormality exists or not by adopting a three-level judgment strategy for the hypersphere model of the independent period which is successfully constructed.
2. The method for detecting abnormal photovoltaic output based on the multi-feature adaptive hypersphere according to claim 1, wherein the multi-dimensional features comprise time sequence statistics features, periodic coding features and physical model features; the time sequence statistical features comprise an original power feature, a power absolute difference feature, a sliding average value feature and a sliding standard deviation feature; The periodic coding features comprise periodic time feature sine codes and periodic time feature cosine codes; the physical model features include power ratio features, expected power features.
3. The method for detecting abnormal photovoltaic output based on the multi-feature adaptive hypersphere according to claim 2, wherein the original photovoltaic output time sequence is expressed as: wherein T is the total number of data points for all training days and predictive days, The original power characteristics of the original photovoltaic output data at the t-th time point are obtained; Absolute differential power characteristics The mathematical expression is: sliding mean feature Is defined as: Wherein the window size Is the original power characteristic value of the kth time point, 1 ; Sliding standard deviation feature Is defined as: ; Wherein, the In order to actually end the upper limit of the range, In order for the sample to be of a practically effective amount, Is the actual lower starting limit; Converting the original timestamp into periodic time characteristics, acquiring periodic time characteristic sine codes and periodic time characteristic cosine codes by adopting sine and cosine code processing, wherein the mathematical expression is as follows: Wherein, the For a periodic temporal feature sinusoidal coding, For the periodic time feature cosine coding, For the point in time An integer hour value in the range of 0-23, Setting 15-minute sampling intervals as minute values, wherein the minute values are 0, 15, 30 and 45; Expected power Based on annual seasonal factors And ideal irradiance model The mathematical expression is: Wherein, the Is the nominal maximum power; annual seasonal factor The mathematical expression of (2) is: Wherein, the Day of the year; Ideal irradiance model The mathematical expression of (2) is: Wherein, the In the form of a value for a minute, In order to make the sunrise time available, For sunset time, the mathematical expressions are respectively: Power ratio The mathematical expression of (2) is: Wherein, the In order to prevent small constants for zero-debugging, As a feature of the original power, it is, Is a desired power characteristic; Selecting effective features and constructing an original feature vector of a t-th time point The mathematical expression is: Wherein, the The characteristic of the sliding average value of 6 hours and the characteristic of the sliding standard deviation of 6 hours are respectively adopted.
4. The photovoltaic output abnormality detection method based on the multi-feature self-adaptive hypersphere according to claim 1 is characterized in that the whole day is divided into 24 independent time periods, and a hypersphere model is built for the independent time periods h meeting the building conditions; the original characteristic vector corresponding to the time point of the independent period h in all training days Sequentially arranging to construct training data matrix The mathematical expression is: Wherein, the For the original feature vector corresponding to the i-th time point after the time points t are sequentially arranged, i=1, 2, & gt, n is the total number of the time points of the period in all training days; Wherein, the independent period h is fixed at a certain time, Is that A dimension training data matrix; If the training data point n of the current period is less than 20, not building a hypersphere model; If training data points of the current period And constructing a super-sphere model, and preprocessing a training data matrix.
5. The method for detecting abnormal photovoltaic output based on multi-feature adaptive hypersphere as claimed in claim 4, wherein the preprocessing includes the step of determining the original feature vectors corresponding to all time points for the independent period h Is a characteristic value of each dimension of (a) Standardized, the mathematical expression is: Wherein, the As a result of the normalized feature values, As a dimension of the features, Respectively is All training samples No The median and the quartile range of the dimensional feature.
6. The method for detecting abnormal photovoltaic output based on the multi-feature adaptive hypersphere according to claim 5, wherein the hypersphere model parameters comprise a hypersphere center point and a hypersphere radius; each dimension characteristic value Normalized feature vector Is composed of The dimensional training data matrix is expressed as ; Calculating the median of the normalized training data matrix according to each characteristic dimension as the central coordinate vector of the hypersphere The mathematical expression of the center coordinate vector is: The radius of the hypersphere is defined as 80% of the fractional number of the distance from all training data points to the center of the hypersphere in the current period, wherein the 80% fractional number refers to the value at the position of 80% after all values are arranged from small to large; The mathematical expression of the hypersphere radius is as follows: Wherein defaults to , For training coordinate vectors of data points To the central coordinate vector If the mahalanobis distance to the central coordinate vector cannot be calculated, selecting the Euclidean distance to the central coordinate vector; The mathematical expression for the mahalanobis distance of the training data point coordinate vector to the center coordinate vector is: Wherein, the Is the inverse of the covariance matrix; Covariance matrix The mathematical expression of (2) is: Wherein, the For each dimension of the eigenvalue Normalized feature vector Is composed of The matrix of training data is maintained, , Is an 8 x 8 dimensional identity matrix; The mathematical expression of the euclidean distance of the training data point coordinate vector to the center coordinate vector is: When (when) At this time, the super-sphere model construction was successful.
7. The photovoltaic output abnormality detection method based on the multi-feature self-adaptive hypersphere according to claim 1 is characterized in that for an independent period corresponding to unsuccessful construction of a hypersphere model, returning an outlier score of 0.5, and judging that no abnormality exists in the corresponding independent period; The three-level decision strategy comprises: A first-stage judgment strategy is to calculate a basic anomaly score based on the center distance of the hypersphere, and primarily screen anomalies to obtain the basic anomaly score; A second-stage judging strategy is to correct the basic abnormal score according to the operation experience of the photovoltaic system to obtain a final abnormal score; and a third stage of judging strategy, namely dynamically adjusting a threshold value according to the independent period, filtering extremely low power and finally judging whether the abnormality exists.
8. The method for detecting abnormal photovoltaic output based on multi-feature adaptive hypersphere as recited in claim 7, wherein the first level decision strategy uses corresponding independent time periods The standardized mode of each dimension feature is obtained And calculate its distance to the center of the hypersphere ; Setting a base anomaly score The mathematical expression of (2) is: Wherein, the Is the central coordinate vector of the hypersphere, Is of super sphere radius if Indicating that the point falls outside the boundary of the hypersphere, and preliminarily judging as abnormal; The second stage of judgment strategy is to introduce a weight function to describe the influence of the operation experience of the photovoltaic system on the photovoltaic detection and based on the current time h and the actual power value Correcting the basic anomaly score and correcting the anomaly score The mathematical expression is: Wherein, the The specific value of the weight function is as follows: ; And a third-stage judgment strategy, wherein the threshold value is dynamically adjusted according to the independent time period, and the mathematical expression is as follows: Wherein, the For the dynamic threshold of the adaptive independent period, As a basis for the threshold value, For the independent period adjustment factor, the independent period adjustment factor takes the following value expression: ; based on correction of anomaly scores Dynamic threshold Filtering the extremely low power to obtain abnormality determination The mathematical expression is: 。
9. The photovoltaic output abnormality detection system based on the multi-feature self-adaptive hypersphere is characterized by comprising a multi-dimensional feature vector module, a self-adaptive hypersphere model module and an abnormality judgment module; The multi-dimensional characteristic vector constructing module extracts multi-dimensional characteristics of the module based on an original photovoltaic output time sequence and constructs multi-dimensional characteristic vectors; the step of constructing the adaptive hypersphere model module and the time-interval adaptive hypersphere model divides the whole day into 24 independent time intervals, constructs the hypersphere model for the independent time intervals meeting the construction conditions, and if training data points of the current time interval are used Constructing a hypersphere model for the corresponding independent time period according to the construction conditions; The abnormality judging module returns an abnormal value score for an independent period corresponding to the independent period which is not successfully constructed, judges that the corresponding independent period is abnormal, and judges whether the abnormality exists or not by adopting a three-level judging strategy for the independent period which is successfully constructed.
10. The multi-feature adaptive hypersphere-based photovoltaic output abnormality detection system of claim 9, wherein the multi-dimensional features include timing statistics, periodic coding, physical model features; the time sequence statistical features comprise an original power feature, a power absolute difference feature, a sliding average value feature and a sliding standard deviation feature; The periodic coding features comprise periodic time feature sine codes and periodic time feature cosine codes; the physical model features include power ratio features, expected power features.

Description

Photovoltaic output abnormality detection method and system based on multi-feature self-adaptive hypersphere Technical Field The invention belongs to the technical field of abnormality detection of photovoltaic systems, and particularly relates to a photovoltaic output abnormality detection method and system based on a multi-feature self-adaptive hypersphere. Background At present, photovoltaic power generation is used as an important component of a sustainable energy system, the application scale of the photovoltaic power generation system is continuously expanded, but the photovoltaic array is easily influenced by various factors such as shadow shielding, inverter faults and component aging in actual operation, so that abnormal output is caused, the power generation efficiency is influenced or safety problems are caused. The photovoltaic system anomaly detection method is divided into a model-based frame and a data-based frame, the model-based photovoltaic anomaly detection method relies on an accurate system mathematical or physical model, and modeling of a large-scale and complex photovoltaic system is time-consuming and difficult. The data-based photovoltaic anomaly detection method comprises a deep learning method which is high in training complexity, extremely sensitive to super-parameter configuration, not capable of effectively integrating physical mechanism knowledge of a photovoltaic system, not capable of associating dynamic characteristics of photovoltaic output under typical operation conditions, not capable of establishing effective association with operation experience of the photovoltaic system, and not capable of lacking an adaptive adjustment mechanism for detection sensitivity in a specific period. The invention patent with the publication number of CN120222962A discloses a detection method of a photovoltaic running state, which is characterized by comprising the steps of obtaining photovoltaic data of a photovoltaic system, detecting the running state of the photovoltaic system based on the distance between the photovoltaic data and the center of at least one preset hypersphere model to obtain a detection result, wherein the preset hypersphere model is used for representing a plurality of historical photovoltaic data of the photovoltaic system in the running state corresponding to the preset hypersphere model, and the center is obtained based on the plurality of historical photovoltaic data corresponding to the preset hypersphere model. The prior art has the following defects that a multidimensional feature fusion system is not constructed, the time-division modeling is not carried out on the time sequence implementation of the photovoltaic output under the typical working condition, and a dynamic scoring and threshold adjustment mechanism guided by domain knowledge is lacked. This is a disadvantage of the prior art. In view of the above, it is desirable to provide a method and a system for detecting abnormal photovoltaic output based on multi-feature adaptive superspheres, so as to solve the above-mentioned drawbacks in the prior art. Disclosure of Invention Aiming at the technical problems that a multi-dimensional characteristic fusion system is not constructed in the prior art, the time-period modeling is not carried out on the time sequence of the photovoltaic output under the typical working condition, and a dynamic scoring and threshold adjustment mechanism guided by field knowledge is lacked, the invention provides a photovoltaic output abnormality detection method and system based on a multi-characteristic self-adaptive hypersphere, and aims to solve the technical problems. In a first aspect, the invention provides a photovoltaic output abnormality detection method based on a multi-feature adaptive hypersphere, comprising the following steps: step S1, constructing a multi-dimensional feature vector, namely extracting multi-dimensional features of the original photovoltaic output time sequence based on the original photovoltaic output time sequence, and constructing the multi-dimensional feature vector; the multidimensional features comprise time sequence statistical features, periodic coding features and physical model features; the time sequence statistical features comprise an original power feature, a power absolute difference feature, a sliding average value feature and a sliding standard deviation feature; The periodic coding features comprise periodic time feature sine codes and periodic time feature cosine codes; the physical model features include power ratio features, expected power features. The raw photovoltaic output time series is expressed as: wherein T is the total number of data points for all training days and predictive days, The original power characteristics of the original photovoltaic output data at the t-th time point are obtained; Absolute differential power characteristics The mathematical expression is: sliding mean feature Is defined as: Wherein the wi