Hierarchical photovoltaic string anomaly detection method and visual analysis system based on dimension reduction mode

Abstract

The invention discloses a hierarchical photovoltaic string anomaly detection method based on a dimension reduction mode, and compared with an existing detection method, the method employs time sequence electrical quantity data and weather data as sample data, combines the dimension reduction mode and a hierarchical model training architecture, and achieves the detection of the anomaly of a photovoltaic string. A dimension reduction mode graph of time sequence electrical quantity data with weather labels is used as an anomaly detection model to be input, the model input cost is reduced, the model prediction accuracy is improved, and a dynamic prediction scene of a real photovoltaic power plant can be effectively dealt with; the invention further provides a visual analysis system based on the photovoltaic string anomaly detection method, the string anomaly recognition result of the anomaly detection model is supported to be displayed and analyzed, and the anomaly detection model is optimized through analysis of the anomaly result and simple labeling. A user is helped to quickly position an abnormal string through interactive operation and formulate a reasonable solution, and effective troubleshooting of a photovoltaic string fault is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a hierarchical photovoltaic string anomaly detection method and a visual analysis system based on a dimensionality reduction model. Background Art

[0002] With the gradual depletion of fossil fuels and the intensification of environmental pollution problems, the use of renewable energy has become particularly important. As one of the core sources of renewable energy, solar energy has become increasingly important in the global energy structure due to its clean, sustainable and economic advantages. Photovoltaic power generation technology is a key means of efficiently converting solar energy into electrical energy, and its importance is self-evident. However, with the gradual commissioning of a large number of photovoltaic power stations, photovoltaic strings, a core device, frequently experience abnormalities. Abnormalities in photovoltaic systems not only reduce power generation efficiency, but may also bring safety hazards such as fire. At the same time, with the rapid increase in installed capacity, the traditional method of relying on manual maintenance of equipment one by one has increasingly exposed the problem of manpower shortage. Therefore, the development of intelligent photovoltaic string abnormality detection technology has become a key task at present.

[0003] In recent years, in order to improve the performance of photovoltaic power stations, researchers have proposed many abnormality detection methods for photovoltaic strings. For example, Reference 1 (Detection and analysis of hot-spot formation in solar cells) uses infrared thermal imaging technology to draw the surface temperature distribution of solar cells under reverse bias, and proposes a fault diagnosis method based on infrared image analysis. Reference 2 (Online Fault Detection in PV Systems) studies the operating status of photovoltaic strings under different light intensity conditions, simulates the correct photovoltaic string surface temperature and irradiation intensity, and compares the difference between the theoretical normal value and the actual measured value to confirm whether an abnormality occurs.

[0004] However, the actual discrimination and accuracy of abnormalities in photovoltaic power station systems greatly rely on thresholds set based on expert knowledge, and the detection of abnormalities has a certain degree of randomness. For example, the invention application with publication number CN114123971A discloses a photovoltaic string abnormality detection method and system based on VaDE, which specifically obtains the current mean value, voltage mean value and installed capacity of each photovoltaic string in the photovoltaic power station during the detection period for a preset time, and determines the first data indicators of each photovoltaic string during the detection period for a preset time based on these data and inputs them into the pre-established photovoltaic string abnormality detection model, obtains the Gaussian distribution weights corresponding to each photovoltaic string, and determines the abnormal photovoltaic strings during the photovoltaic power station detection period based on the Gaussian distribution weights. Although this method improves the accuracy of abnormality detection, it relies on high-quality labeled data, which has a high acquisition cost in practical applications. In addition, when the actual photovoltaic strings are running, inherent factors such as equipment defects, installation inclination angles, and transient interferences such as dark cloud cover often cause false alarms in the model, increasing processing costs.

[0005] In addition, the combination of artificial intelligence and photovoltaic string anomaly detection has improved the accuracy of the model, but due to the black box structure of the model itself, it is not conducive to users understanding the relationship between input and output. Data visual analysis systems are widely used in the field of deep learning. They encode the attributes of data with intuitive visual charts and integrate human perception into data analysis through human-computer interaction, which enhances accuracy and interpretability and helps users understand model behavior to a certain extent. Therefore, in the actual operation of photovoltaic power stations, how to combine domain knowledge to detect photovoltaic string anomalies is a problem that needs to be solved urgently. Summary of the invention

[0006] The purpose of the present invention is to provide a hierarchical photovoltaic string anomaly detection method and a visual analysis system based on a dimensionality reduction model, which combines the dimensionality reduction model and the hierarchical model training architecture, introduces weather clustering information, and constructs a detection model that effectively identifies photovoltaic string anomalies, reduces the input cost of the model and the model false alarm rate, and combines interactive visual analysis operations. Users can intuitively understand how the anomaly detection model identifies abnormal photovoltaic strings and the possible causes of photovoltaic string anomalies, helping users to quickly locate anomalies and formulate corresponding processing plans.

[0007] To achieve the above-mentioned purpose of the invention, an embodiment provides a hierarchical photovoltaic string anomaly detection method and a visual analysis system based on a dimensionality reduction model, including a hierarchical photovoltaic string anomaly detection method based on a dimensionality reduction model and a visual analysis system based on the hierarchical photovoltaic string anomaly detection method.

[0008] In one embodiment, the hierarchical photovoltaic string anomaly detection method based on dimensionality reduction mode comprises the following steps:

[0009] Collect the time-series electrical quantity data and weather data of photovoltaic strings and perform data preprocessing;

[0010] Based on the preprocessed time series electrical quantity data, a dimension reduction pattern diagram of the time series electrical quantity data is calculated;

[0011] Based on the preprocessed weather data, weather labels are formed by clustering, and the weather labels are introduced into the dimension reduction pattern diagram of the time series electrical quantity data to obtain the dimension reduction pattern diagram of the time series electrical quantity data with weather labels;

[0012] Based on the dimensionality reduction pattern diagram of time-series electrical quantity data with weather labels, a hierarchical approach is used to train anomaly detection models. The trained anomaly detection model is used to identify abnormal PV strings, and the anomaly detection results of all PV strings are obtained.

[0013] In one embodiment, the data preprocessing includes: error value processing, missing value filling and noise filtering;

[0014] The error value processing means that when the sequential electrical quantity data exceeds the rated value or a negative value appears, it is considered that there is an error value and the error value is eliminated;

[0015] The missing value filling method fills a large number of continuous empty values ​​appearing in the time series electrical quantity data and weather data of the photovoltaic strings by an interpolation method;

[0016] The noise filtering smoothes the filtering of the time-series electrical quantity data and weather data of the photovoltaic strings by a sliding average method, thereby eliminating the interference of weather factors.

[0017] In one embodiment, the calculation to obtain the dimension reduction pattern diagram of the time series electrical quantity data includes: performing a dimension transformation operation on the preprocessed time series electrical quantity data and obtaining the dimension reduction pattern diagram of the time series electrical quantity data based on UMAP calculation.

[0018] In one embodiment, the photovoltaic string-based weather data is clustered to form weather labels, including: clustering and color-coding the preprocessed weather data using a KMeans clustering method to form weather labels, wherein blue represents low irradiation conditions, green represents medium irradiation conditions, and red represents high irradiation conditions.

[0019] In one embodiment, the hierarchical approach to training anomaly detection models includes: using the dimensionality reduction pattern diagram of time-series electrical quantity data with weather labels as the input of the anomaly detection model, grouping the dimensionality reduction pattern diagrams of all photovoltaic strings according to the inverters they belong to, and then grouping the dimensionality reduction pattern diagrams of the photovoltaic strings under each group of inverters based on the weather labels, and training an anomaly detection model separately for the dimensionality reduction pattern diagram of each type of weather label.

[0020] In one embodiment, the anomaly detection model combines three unsupervised models: a single-class support vector machine, an elliptical envelope, and a local anomaly factor, and the weighted average of the three anomaly models is used as the output of the anomaly detection model.

[0021] In one embodiment, the identification of abnormal photovoltaic strings is performed based on the output results of the abnormal detection model calculation, and the output results include: abnormal value R a and degradation rate R d ;

[0022] The outlier value is obtained by selecting the anomaly detection model architecture and normalizing the results using the MinMax method to quantify the degree of abnormality of each photovoltaic string. The outlier value R a It is expressed as:

[0023]

[0024] Among them, Out represents the output result of the anomaly detection model of the i-th photovoltaic string, the intermediate variable m=OCSVM|EE|LOF represents that the selected anomaly detection model is a single-class support vector machine, a local anomaly factor and an elliptical envelope, n represents the number of photovoltaic strings; Min represents the minimum value of the mean value of the output results of the anomaly detection model of n photovoltaic strings; Max represents the maximum value of the mean value of the output results of the anomaly detection model of n photovoltaic strings;

[0025] The degradation rate is obtained by quantifying the relative value of each point in the dimension reduction pattern to the normal data cluster C. N The method uses the same training method as the anomaly detection model to identify abnormal photovoltaic strings based on the method of fitting the distribution boundary of the normal dimensionality reduction pattern, and selects the Gaussian mixture model as the model structure. The Gaussian mixture model approximates the normal data cluster C through the weighted sum of multiple Gaussian distributions. N The probability distribution p(x) is used to calculate the degradation rate R d The MinMax method is used to normalize the calculation results of the Gaussian mixture model to the range of [0, 1]; and the degradation rate R d The groups are grouped by time t, and the minimum likelihood of all normal PV strings in each group is used as the boundary. The ratio of the likelihood of all normal PV strings in the group to the minimum likelihood is calculated, and the weighted sum of the ratios of the PV strings at each time point is used to calculate the approximate degradation rate R. d , degradation rate R d It is expressed as:

[0026]

[0027] Among them, x represents a multidimensional data point in the dimensionality reduction pattern graph, x∈C N , CN Represents normal data clusters in the dimensionality reduction pattern graph; Represents the likelihood of the parameter θ, which is used to estimate the value that can make x equal to θ when the multidimensional data point x in the dimensionality reduction pattern is known. is numerically identical to the probability distribution p(x); θ represents the parameter set of the Gaussian mixture model, including the mixture weight, mean, and covariance matrix; K represents the number of Gaussian distributions, and π k represents the mixture weight of the k-th Gaussian distribution, represents the kth Gaussian distribution, μ k is the mean of the kth Gaussian distribution, ∑ k is the covariance matrix of the kth Gaussian distribution; MinMax indicates that when calculating the degradation rate R d The MinMax method is used to normalize the calculation results of the Gaussian mixture model to the range of [0, 1], and max is represented by the maximum value in the range of [0, 1]; R d (t) represents the degradation rate R d Group by time t, the time range is [1, n]; represents the likelihood of grouping by time t, and the minimum likelihood of all normal PV strings in each group is expressed as

[0028] In order to clearly demonstrate the hierarchical photovoltaic string anomaly detection method based on the dimensionality reduction mode, a visualization analysis system of the hierarchical photovoltaic string anomaly detection method is provided. The system uses the photovoltaic power generation prediction method in the visualization analysis. The visualization analysis system includes: a global view module, a hierarchical view module, a sorting view module, a mode view module and an analysis view module;

[0029] The global view module provides a frequency histogram and a rectangular tree diagram for displaying the recognition results of the anomaly detection model on the input data, wherein the frequency histogram displays the distribution of abnormal values ​​of all photovoltaic strings in the photovoltaic power station calculated by the anomaly detection model, and provides an overall overview of the abnormal photovoltaic strings identified by the anomaly detection model; the rectangular tree diagram displays the abnormal values ​​of all photovoltaic strings in the photovoltaic power station based on the hierarchical structure of the photovoltaic power station box transformer-inverter-photovoltaic string;

[0030] The hierarchical view module is used to display the overall distribution of the dimension reduction pattern diagrams of all photovoltaic strings under the inverter in the form of scatter plots and contour plots, and to obtain the distribution boundary of the dimension reduction pattern diagram of normal photovoltaic strings based on the distribution boundary fitting algorithm of the dimension reduction pattern diagram of normal strings, so as to help users understand the reasons why the abnormal detection model identifies the photovoltaic strings as abnormal and further optimize the abnormal detection model;

[0031] The sorting view module is used to display the original current data of each photovoltaic string under the selected inverter in the global view and the output results of the abnormal detection model calculation, and provides filtering and sorting functions to help users quickly locate the photovoltaic strings of interest;

[0032] The mode view module is used to display the original time series electrical quantity data and the dimension reduction mode diagram of the photovoltaic strings selected in the sorting view by providing line graphs and scatter graphs, so as to help users explore the similarities and differences between normal and abnormal photovoltaic strings in the dimension reduction mode diagrams, and provide reasonable explanations for the recognition results of the anomaly detection model;

[0033] The analysis view module is used to provide relevant abnormal characteristic analysis indicators of the specified photovoltaic string and a simple data annotation function to assist the user in analyzing the causes of the photovoltaic string abnormality.

[0034] In one embodiment, the normal photovoltaic string dimension reduction pattern diagram distribution boundary is obtained based on the normal photovoltaic string dimension reduction pattern diagram distribution boundary fitting, including:

[0035] Based on the normal photovoltaic strings identified by the anomaly detection model, the Gaussian mixture model is used to generate a contour map to depict the boundaries of the normal photovoltaic string distribution. The contour line boundary value k of the contour map is calculated by the k-sigma method, indicating that the contour line is calculated based on the μ-kσ of the photovoltaic string, and the value range of k is [1, 5];

[0036] The further optimized anomaly detection model is optimized by adjusting the contour line boundary value k of the contour map and the number of sub-distributions in the Gaussian mixture model.

[0037] In one embodiment, the provision of the relevant abnormal characteristic analysis index of the specified photovoltaic string includes: providing the normal rate R n , relative power generation rate R rpg and irradiation correlation coefficient C Irr Indicators and visualize them through radar charts to facilitate estimation of the values ​​of each indicator;

[0038] The normal rate R n Indicates the proportion of normal PV strings in the dimension reduction mode diagram;

[0039] The relative power generation rate R rpg is the PV string S under the same inverter and on the same day relative to the optimal PV string S max The ratio of the power generation P is expressed as:

[0040]

[0041] Where N represents the number of days, P s，d represents the power generation of the PV string S under the same inverter and on the dth day, Pmax，d represents the optimal photovoltaic string S max Power generation under the same inverter and on day d;

[0042] The irradiation correlation coefficient C Irr is the Pearson correlation coefficient between the time series current data I and the irradiation data R of the photovoltaic string, which ranges from [0, 1] and is expressed as:

[0043]

[0044] For a normal photovoltaic string S, the current data I is usually proportional to the irradiation data R, and the irradiation correlation coefficient C Irr The closer it is to 1, the more normal the PV string is. N represents time, I represents s，t It represents the time series current data of a normal photovoltaic string S at time t. represents the average value of the time series current data of a normal photovoltaic string S, R t represents the irradiation data of the normal photovoltaic string S at time t, represents the mean value of the irradiation data of the normal photovoltaic string S, and max represents the maximum value in the range of [0, 1] as the Pearson correlation coefficient;

[0045] The visualization of the indicator by radar chart includes: setting the center of the radar chart to represent the indicator value as 0, the outermost vertex to represent the value as 1, the smaller the area of ​​the radar chart, the more abnormal the photovoltaic string is, and the radar chart represents the relevant abnormal characteristics of the specified photovoltaic string under low irradiation, medium irradiation and high irradiation conditions through unified color coding.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The hierarchical photovoltaic string anomaly detection method based on dimensionality reduction pattern provided by the present invention is based on time-series electrical quantity data and weather data as sample data, combined with dimensionality reduction pattern and hierarchical model training architecture, and uses dimensionality reduction pattern diagram of time-series electrical quantity data with weather labels as anomaly detection model input compared to existing detection methods, thereby reducing model input cost and improving model prediction accuracy, and can effectively cope with the prediction scenario of real photovoltaic power plant dynamics; the present invention also provides a visualization analysis system based on the photovoltaic string anomaly detection method, which supports the display and analysis of string anomaly recognition results of the anomaly detection model, optimizes the anomaly detection model by analyzing the anomaly results and simply annotating the anomaly detection model, and helps users quickly locate abnormal strings through interactive operations and formulate reasonable solutions, thereby realizing effective troubleshooting of photovoltaic string failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0049] Figure 1 is a flow chart of the photovoltaic string abnormality detection method of the present invention;

[0050] Figure 2 Schematic diagram of a photovoltaic string abnormality detection method according to the present invention;

[0051] Figure 3 It is a schematic diagram of the distribution boundary fitting algorithm of the system normal group string dimension reduction pattern diagram;

[0052] Figure 4 is a schematic diagram of a visual analysis system of the present invention;

[0053] Figure 5 is a schematic diagram of the global view module;

[0054] Figure 6 is a schematic diagram of a hierarchical view module;

[0055] Figure 7 is a schematic diagram of a sorting view module;

[0056] Figure 8 is a schematic diagram of a mode view module;

[0057] Fig. 9 A schematic diagram of the analysis view module. DETAILED DESCRIPTION

[0058] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0059] In order to effectively identify abnormal photovoltaic strings, the embodiment provides a hierarchical photovoltaic string anomaly detection method based on a dimensionality reduction model. The anomaly detection method combines the dimensionality reduction model with the hierarchical model training architecture, introduces weather clustering information, and constructs a detection model that effectively identifies photovoltaic string anomalies, thereby reducing the input cost of the model and the model false alarm rate. Figure 1 As shown, the photovoltaic string abnormality detection method provided in the embodiment includes the following steps:

[0060] S1. Collect the time-series electrical quantity data and weather data of the photovoltaic strings and perform data preprocessing.

[0061] In the embodiment, the time-series electrical quantity data and weather data of the photovoltaic strings are obtained from the photovoltaic power station as the original data. However, there are a large number of abnormal data caused by equipment and communication failures in the original data set, so it is necessary to design an appropriate data preprocessing process, including:

[0062] (1) Error value processing: Due to defects in sampling equipment, the obtained time series electrical quantity data and weather data are erroneous and violate the actual conditions. For example, there are still current values ​​under no light conditions at night, the current values ​​exceed the rated current value or appear negative, etc. Such errors need to be eliminated;

[0063] (2) Missing value filling: Due to equipment communication failure, the acquired time series electrical quantity data and weather data are lost during the transmission process, and a large number of continuous null values ​​appear in the original data set, which seriously affects the training and verification of the subsequent anomaly detection model. It needs to be processed by linear or nonlinear interpolation methods;

[0064] (3) Noise filtering: Factors such as the stability of equipment communication and cloud cover will interfere with the collected raw data, which will be reflected in the time series electrical quantity data as noise or glitches. The sliding average method is used to filter and smooth the time series electrical quantity data to eliminate interference.

[0065] Based on the above data preprocessing process, the preprocessed time series electrical quantity data and weather data are obtained.

[0066] S2. Based on the preprocessed time series electrical quantity data, a dimension reduction pattern diagram of the time series electrical quantity data is calculated.

[0067] In the embodiment, Figure 2 As shown in A, the dimension reduction pattern diagram of the time series electrical quantity data is obtained by using the dimension transformation operation on the preprocessed time series electrical quantity data and calculating based on UMAP, which is specifically:

[0068] Firstly, the preprocessed time series electrical quantity data is downsampled. The time granularity of the preprocessed time series electrical quantity data is 1 minute. In order to reduce the calculation cost without reducing the accuracy, its time granularity is converted to 1 hour by downsampling, and the aggregation method adopts the averaging method. Assuming that the total length of the entire time series is n, and the number of features used is 1 (mainly focusing on the real-time current of the string), the shape of the input vector after downsampling changes from (n, 1) to (n / 60, 1); the second is the dimension transformation. In the previous step of downsampling, the number of points corresponding to one day becomes 24, so every 24 points are converted into one row, and the shape of the input vector changes from (n / 60, 1) to (n / 1440, 24), so that each row represents the electrical quantity data of the photovoltaic string for one day; finally, UMAP is used to convert the data after the dimension change into a pattern diagram. By converting the 24-dimensional vector into a 2-dimensional vector, the shape of the input vector changes from (n / 1440, 24) to (n / 1440, 2), and projecting it onto a 2-dimensional plane diagram, a dimensionality reduction pattern diagram of the time series electrical quantity data is obtained. Each point in the dimensionality reduction pattern diagram represents the power generation status of the photovoltaic string on a certain day.

[0069] S3. Clustering is performed based on the preprocessed weather data to form weather labels, and the weather labels are introduced into the dimensionality reduction pattern diagram of the time series electrical quantity data to obtain the dimensionality reduction pattern diagram of the time series electrical quantity data with weather labels.

[0070] In the embodiment, Figure 2 As shown in B, the preprocessed weather data is clustered to form weather labels, and the weather labels are mapped back to the dimension reduction pattern diagram of the time series electrical quantity data of S2 to complete the introduction of weather labels, specifically:

[0071] First, the preprocessed weather data is clustered. Assuming that the total length of the entire time series is n and the number of features used is 1 (mainly focusing on the power station irradiation data), the shape of the input vector is (n, 1). The collection frequency of irradiation data is also 1 minute. Therefore, the downsampling method is used to convert the time granularity into the hourly level, and the vector shape becomes (n / 60, 1), reducing the calculation amount of the anomaly detection model and the data fluctuation caused by interference such as communication failures; then, the irradiation data is divided by day, and the shape of a single vector becomes (24, 1) as the object of the Kmeans algorithm for clustering. The irradiation data sequences with similar weather conditions are grouped into the same category, and finally three clusters representing low irradiation, medium irradiation and high irradiation are obtained. The weather labels are encoded in blue (#7883BB), green (#86BB78) and red (#DE7676) respectively to obtain weather labels; the weather labels are mapped back to the dimensionality reduction pattern diagram of the time series electrical quantity data to obtain the dimensionality reduction pattern diagram of the time series electrical quantity data with weather labels as the input of the anomaly detection model.

[0072] S4. Based on the dimension reduction pattern diagram of the time-series electrical quantity data with weather labels, the anomaly detection model is trained in a hierarchical manner to identify abnormal PV strings and obtain the anomaly detection results of all PV strings.

[0073] In the embodiment, Figure 2 As shown in C, when modeling the dimensionality reduction pattern diagram of all photovoltaic strings under a single inverter, different weather conditions such as irradiation have a great influence on the distribution of string data points in the dimensionality reduction pattern diagram. There is a large change in the spatial distance between high irradiation and low irradiation points. At the same time, the outlier characteristics of abnormal photovoltaic string data points in space are easy to confuse the recognition ability of the model. Therefore, the weather clustering label is introduced to group the aggregated dimensionality reduction pattern diagram of the inverter level. An anomaly detection model is trained separately for each type of weather condition to eliminate the data distribution deviation caused by the irradiation factor and further optimize the performance of the model. Specifically:

[0074] Firstly, the dimension reduction pattern diagram of time series electrical quantity data is used as the input of the anomaly detection model, and the hierarchical idea is introduced to group the photovoltaic strings into Inv according to their inverters. 1 ,Inv 2 , ..., Inv n , where n represents the total number of inverters in the input data set, Inv i Represents the dimension reduction mode diagram of the time series electrical quantity data of all photovoltaic strings under inverter i; the second step is to convert each Inv i All the dimensionality reduction pattern graphs in the dimensionality reduction pattern graph are combined together, and each point in the dimensionality reduction pattern graph is grouped according to the corresponding weather label. i It is further divided into Inv i，low ,Inv i，med ,Inv i，high Three groups, corresponding to low, medium and high irradiation conditions; finally, for each Inv i,label (i=1,2,...,n;label=low|med|high) trains an anomaly detection model separately.

[0075] In the embodiment, a composite model is selected for anomaly detection. The structure of the composite model can make full use of the advantages of each model in different data distributions. For example, One Class SVM (single-class support vector machine) is a probability distribution model that does not rely on data, and is suitable for data sets with unknown or complex distributions. The introduction of kernel techniques enables it to capture the nonlinear characteristics of data and improve the accuracy of detection; EllipticEnvelope (elliptical envelope) uses the covariance matrix to fully consider the correlation between dimensions. The calculation is simple and efficient, and it is suitable for large-scale data sets that meet the multivariate normal distribution; Local OutlierFactor (local anomaly factor) evaluates the degree of anomaly by comparing the density of the local area of ​​the data point. It does not rely on the global data distribution and can handle a variety of density distribution patterns. Based on the idea of ​​ensemble learning, the three unsupervised models of One Class SVM, EllipticEnvelope and Local Outlier Factor are selected to be combined together, and the weighted average of the three models is used as the output of the anomaly detection model;

[0076] The anomaly detection model outputs all outliers in the dimensionality reduction pattern graph grouped in a hierarchical manner. In order to quantify the degree of anomaly of each photovoltaic string for anomaly detection, the outlier value R of the string is a Defined as:

[0077]

[0078] Among them, Out represents the output result of the anomaly detection model of the i-th PV string, the intermediate variable m=OCSVM|EE|LOF indicates that the selected anomaly detection model is a single-class support vector machine, a local anomaly factor and an elliptical envelope, n represents the number of PV strings; Min represents the minimum value of the mean value of the output results of the anomaly detection model of n PV strings; Max represents the maximum value of the mean value of the output results of the anomaly detection model of n PV strings.

[0079] The abnormal value of each photovoltaic string is obtained by weighted summing the number of corresponding abnormal points in the dimensionality reduction pattern diagram. The larger the abnormal value, the higher the abnormal degree of the photovoltaic string. Set the abnormal threshold arr_thr of the photovoltaic string abnormality detection model, and use the common 3σ criterion statistical method to automatically calculate the value of arr_thr. If the abnormal value R of the photovoltaic string is a If it is greater than the abnormal threshold arr_thr, the abnormal detection model will identify it as an abnormal string, and vice versa.

[0080] In addition to identifying whether the PV string is abnormal, the model also needs to evaluate the degree of efficiency loss relative to the normal PV string, because faults in real PV power plant scenarios do not necessarily lead to a decline in PV string performance, and operation and maintenance personnel mainly focus on those PV strings whose power generation performance has declined due to faults. If the degree of degradation of the PV string can be quantified, operation and maintenance personnel can use this indicator to prioritize fault repairs. Therefore, it is set to approximate the degradation rate of the PV string by measuring the degree of deviation of each point in the reduced-dimensional pattern diagram relative to the normal data cluster.

[0081] In the embodiment, in order to calculate the deviation degree of data points in the dimension reduction data map of the photovoltaic string, it is proposed to use the degradation rate for evaluation. The degradation rate is obtained by quantifying the deviation degree of each point in the dimension reduction pattern map relative to the normal data cluster C. N The degree of deviation can be used to approximate the assessment, including: Figure 3 The method based on fitting the distribution boundary of the normal dimensionality reduction pattern graph shown in FIG. adopts the same training method as the anomaly detection model to identify abnormal photovoltaic strings. It is grouped according to the inverter and weather labels, and the normal photovoltaic strings filtered by the threshold arr_thr are used as input. The Gaussian mixture model is selected as the model structure. The Gaussian mixture model (GMM) approximates the normal data cluster C through the weighted sum of multiple Gaussian distributions. N The probability distribution p(x) of the fitted GMM model can calculate the likelihood of any data point in the dimensionality reduction pattern graph Higher likelihood Indicates that the point belongs to the normal data cluster C N The probability of is greater, and vice versa;

[0082] In calculating the degradation rate R d The MinMax method is used to normalize the calculation results of the Gaussian mixture model to the range of [0, 1]; and the degradation rate R d The groups are grouped by time t, and the minimum likelihood of all normal PV strings in each group is used as the boundary. The ratio of the likelihood of all normal PV strings in the group to the minimum likelihood is calculated, and the weighted sum of the ratios of the PV strings at each time point is used to calculate the approximate degradation rate R. d , degradation rate R d It is expressed as:

[0083]

[0084] Among them, x represents a multidimensional data point in the dimensionality reduction pattern graph, x∈C N , C N Represents normal data clusters in the dimensionality reduction pattern graph; Represents the likelihood of the parameter θ, which is used to estimate the value that can make x equal to θ when the multidimensional data point x in the dimensionality reduction pattern is known. is numerically identical to the probability distribution p(x); θ represents the parameter set of the Gaussian mixture model, including the mixture weight, mean, and covariance matrix; K represents the number of Gaussian distributions, and π k represents the mixture weight of the k-th Gaussian distribution, represents the kth Gaussian distribution, μ k is the mean of the kth Gaussian distribution, ∑ k is the covariance matrix of the kth Gaussian distribution; MinMax indicates that when calculating the degradation rate R d The MinMax method is used to normalize the calculation results of the Gaussian mixture model to the range of [0, 1], and max is represented by the maximum value in the range of [0, 1]; R d (t) represents the degradation rate R d Group by time t, the time range is [1, n]; represents the likelihood of grouping by time t, and the minimum likelihood of all normal PV strings in each group is expressed as

[0085] The main challenge in the current operation and maintenance of photovoltaic power stations is that inefficient photovoltaic strings are difficult to identify. Existing studies have shown that machine learning and deep learning technologies have been applied to the anomaly detection of photovoltaic strings, helping to identify faulty strings and facilitating repairs by operation and maintenance personnel. However, most detection methods rely on high-quality annotated data, which is costly to obtain in practical applications. In addition, inherent factors such as equipment defects, installation inclination angles, and transient interferences such as cloud cover often lead to false alarms in the model, increasing processing costs. At the same time, the black box characteristics of the model make the output results lack interpretability, and it is difficult for users to understand the impact of different input features, and thus it is impossible to draw effective analysis conclusions. A hierarchical photovoltaic string anomaly detection method based on a dimensionality reduction model. The present invention also provides a visualization analysis system that combines the photovoltaic string anomaly detection method with visualization to effectively identify inefficient photovoltaic strings, guide users to quickly explore photovoltaic string patterns, improve the interpretability of the model, and accurately understand the model results. Based on the visualization analysis system combined with human perception and expert experience, valuable annotation data is quickly generated to formulate reasonable solutions and achieve effective troubleshooting of photovoltaic string faults. Figure 4 As shown, the visual analysis system provided in the embodiment includes: a global view module, a hierarchical view module, a sorting view module, a pattern view module and an analysis view module.

[0086] In the embodiment, Figure 5As shown in the figure, the global view module provides frequency histograms and rectangular tree diagrams to display the recognition results of the PV string anomaly detection model for the input data set, and provides the actual distribution of abnormal PV strings from the perspective of outliers and PV power station structure. After the user selects an inverter of interest for analysis, the relevant visual data content will be displayed accordingly in the hierarchical view and sorting view.

[0087] The frequency distribution histogram shows the distribution of abnormal values ​​of all photovoltaic strings in the photovoltaic power station calculated by the anomaly detection model. The horizontal axis represents the size of the abnormal value, and the vertical axis represents the number of photovoltaic strings within the corresponding abnormal value range. The red vertical line marks the abnormal threshold of the photovoltaic string automatically calculated by the 3σ criterion statistical method, providing an overall overview of the model's abnormal identification. The rectangular tree diagram hierarchically displays the abnormal values ​​of all photovoltaic strings in the photovoltaic power station.

[0088] In the initial state, the rectangular units in the rectangular tree diagram represent the box transformers in the photovoltaic power station. The box transformer number (such as BT001) is marked in the upper left corner of each rectangular unit. The color coding of the rectangle represents the abnormal value. The darker the color, the larger the abnormal value. When you click on a box transformer, the rectangular tree diagram will display the abnormal value of the inverter at the next level, and so on. Users can gain in-depth understanding of the abnormal distribution at different levels.

[0089] In addition, the global view module also provides a slider function for adjusting the abnormal threshold. Users can drag the slider left and right or click the button to accurately adjust the threshold. After adjustment, the data in the associated rectangular tree diagram and other views will be updated synchronously, helping users to quickly locate the abnormal string group of interest; at the same time, by clicking on the rectangular unit, users can further analyze the operation history of the string group in other associated views. After clicking, the rectangular unit will be highlighted, and the color coding will remain consistent, which improves the interactive experience and the intuitiveness of information visualization.

[0090] In the embodiment, Figure 6 As shown, the hierarchical view module is used to display the overall distribution of the dimension reduction pattern diagrams of all PV strings under the inverter in the form of scatter plots and contour plots, and to obtain the distribution boundaries of the normal PV string dimension reduction pattern diagram based on the normal string dimension reduction pattern diagram distribution boundary fitting algorithm, so as to help users understand the reasons why the anomaly detection model identifies the PV strings as abnormal and further optimize the anomaly detection model.

[0091] After selecting the inverter number to be analyzed, the visual analysis system will obtain the dimensionality reduction pattern diagrams of all PV strings under the selected inverter and merge them into a scatter plot for display. In order to ensure the comparability between different inverters, the x-axis and y-axis ranges of the scatter plot are uniformly set to the global maximum and minimum values ​​of the dimensionality reduction pattern diagrams of all PV strings in the PV power station, and 5% of the range is added or subtracted on this basis. In this way, when switching between different inverters, abnormal patterns can be easily discovered for further analysis.

[0092] Based on the selected inverter, each point in the scatter plot is color-coded according to the weather conditions corresponding to the PV strings. The visual analysis system divides the weather labels into three categories, using blue (#7883BB) for low irradiation, green (#86BB78) for medium irradiation, and red (#DE7676) for high irradiation. This weather coding method helps users distinguish the distribution of strings under different weather conditions and understand the impact of weather on the dimensionality reduction pattern diagram. In addition, the scatter plot supports zooming and panning functions, and users can use these interactive functions to conduct in-depth analysis of overall and local features. When the mouse hovers over a scatter point, detailed attribute information of the point will be displayed, including the PV string number, time, and outliers.

[0093] To further enhance the expressiveness of the dimensionality reduction pattern diagram of the PV strings, the system also adds a contour diagram. The contour diagram is generated by the Gaussian mixture model trained by the normal PV strings identified by the anomaly detection model, and is used to depict the boundaries of the normal PV string distribution. The contour line boundary value k of the contour diagram is calculated by the k-sigma method. The default value of k is 3, indicating that the contour line is calculated based on the μ-kσ of the PV strings.

[0094] In the Hierarchical View module, users can adjust the contour parameters through the two sliders above the scatter plot. The first slider "boundary" controls the size of k in the contour boundary value of the contour plot, and the value range is [1, 5]; the second slider "n_components" is used to adjust the number of sub-distributions in the GMM model, and the value range is [1, m], where m is the total number of strings under the inverter. By adjusting these two sliders, users can further optimize the fitting effect of the model to make it more accurately reflect the normal distribution of PV strings.

[0095] In the embodiment, Figure 7 As shown, the sorting view module is used to display the original current data of each PV string under the selected inverter in the global view and the output results calculated by the anomaly detection model, and provides filtering and sorting functions to help users quickly locate the PV strings of interest.

[0096] The sorting view module is displayed in a tabular form, with the PV string number (Pid), current status, abnormal value (Ra) and degradation rate (Rd) from left to right. The Pid column indicates the number of the string under the inverter. For example, if the inverter is BT001-I001 and the Pid is PV1, the PV string number format is BT001-I001-PV1. The second column, current status, shows the time-series current data of the PV string, which is displayed through a line thumbnail. The x-axis is time, and the time range covers December 9, 2022 to March 26, 2023. The y-axis is the current value, ranging from the minimum to maximum value of the current of all PV strings in the PV power station, and a 5% range is added to ensure visual contrast. R a and R d The columns show the abnormal values ​​and degradation rates of the PV strings, both in the range of [0, 1], and are presented in the form of horizontal bar graphs. The length of the bar indicates the numerical value of the abnormal value and degradation rate. Users can sort by Pid, ​​R a or R d Sort the data in ascending or descending order by clicking on the columns, and drag the current timing diagram of the selected PV string to the mode view for further analysis. When the amount of data exceeds the page length, users can switch to view the results of all PV strings through the paging function at the bottom.

[0097] In the embodiment, Figure 8 As shown, the pattern view module is used to display the original time-series electrical quantity data and dimensionality reduction pattern diagram of the PV strings selected in the sorting view by providing line graphs and scatter graphs, helping users to explore the similarities and differences between normal and abnormal PV strings in the dimensionality reduction pattern diagrams, and providing a reasonable explanation for the recognition results of the anomaly detection model.

[0098] The mode view module consists of a line graph and a scatter graph. The line graph displays the time series current data of the PV strings selected from the sort view module by dragging. The settings of the x-axis and y-axis are the same as those of the thumbnail line graph in the sort view module. Figure 1 The line chart of the mode view module contains two time-series current curves. The PV string coded with color #c77272 is the analysis string, and the PV string coded with color #4682b4 is the reference string. By comparing the current deviation of the two time-series current curves in the same time period, users can analyze the R a or R d The line chart also provides horizontal zooming and dragging functions, so users can zoom in on local areas for in-depth analysis.

[0099] like Figure 8As shown in the figure, the scatter plot in the mode view module consists of the analysis string dimensionality reduction pattern diagram on the left and the reference string dimensionality reduction pattern diagram on the right. The upper left corner of the analysis string dimensionality reduction pattern diagram and the reference string dimensionality reduction pattern diagram displays the PV string number, and the number format is BT001-I001-PV1. The color coding is consistent with the line chart in the mode view module. The dimensionality reduction pattern diagram uses the same representation method as the hierarchical view module, showing the scatter plot based on weather label clustering and color coding and the corresponding contour map of the normal PV string point distribution. However, in the model view module, only the scatter points of the selected PV string are displayed, not all the scatter points under the entire inverter. In order to highlight the abnormal mode, the abnormal scatter points identified by the model are represented by stars, and the size is also larger than the normal PV string points. The upper right corner of each dimensionality reduction pattern diagram also provides an associated mapping function. The first button can restore the layout to the default state, and the second button is a circle selection function. The user can circle the specified area in the dimensionality reduction pattern diagram, and the scatter points corresponding to the time on the left and right sides will be highlighted synchronously. When you hover the mouse over any scatter point, detailed information about the scatter point, such as time, outliers, etc., will be displayed, and the corresponding time period in the line chart of the pattern view module will also be highlighted. This design helps users analyze and mine the potential correlation between abnormal patterns and original current data, and further explain the recognition results of the model.

[0100] In the embodiment, Fig. 9 As shown, the analysis view module is used to provide relevant abnormal characteristic analysis indicators of the specified PV string and a simple data annotation function to assist users in analyzing the causes of PV string abnormalities.

[0101] The left side of the analysis view module shows the statistical characteristic indicators of the analysis string and the reference string, which helps users to explain the causes of abnormal patterns in the PV strings in detail. In order to reasonably explain these abnormal patterns, three important indicators for result analysis are designed and introduced: normal rate R n , relative power generation rate R rpg and irradiation correlation coefficient C Irr , where the normal rate R n Represents the proportion of normal photovoltaic strings in the dimensionality reduction mode diagram; relative power generation rate R rpg is the PV string S under the same inverter and on the same day relative to the optimal PV string S max The ratio of the power generation P is expressed as:

[0102]

[0103] Where N represents the number of days, P s，d represents the power generation of the PV string S under the same inverter and on the dth day, P max，d represents the optimal photovoltaic string S max Power generation under the same inverter and on day d;

[0104] Irradiation correlation coefficient C Irr is the Pearson correlation coefficient between the time series current data I and the irradiation data R of the photovoltaic string, which ranges from [0, 1] and is expressed as:

[0105]

[0106] For a normal photovoltaic string S, the current data I is usually proportional to the irradiation data R, and the irradiation correlation coefficient C Irr The closer it is to 1, the more normal the PV string is. N represents time, I represents s，t It represents the time series current data of a normal photovoltaic string S at time t. represents the average value of the time series current data of a normal photovoltaic string S, R t represents the irradiation data of the normal photovoltaic string S at time t, represents the mean value of the irradiation data of the normal PV string S, and max represents the maximum value in the range of [0, 1] as the Pearson correlation coefficient.

[0107] The design of the relevant abnormal characteristic analysis indicators based on the above-mentioned specified PV strings ensures that the value range of all characteristic indicators is [0, 1] and has the same directionality, that is, the closer the indicator is to 1, the more normal the PV string is, and vice versa, which helps to enhance the expression effect of these indicators when mapped into a visual form.

[0108] In order to more vividly display these characteristic indicators, the system uses radar charts for visualization. Since three characteristic indicators are involved, the background of the radar chart is a triangle, and it is divided into four equal parts by alternating gray and white colors, which is convenient for users to quickly estimate the values ​​of each indicator visually. The center of the radar chart represents the indicator value of 0, and the outermost vertex represents the value of 1. The smaller the area of ​​the radar chart, the more abnormal the photovoltaic string. The radar chart is arranged in a 2×3 format. The upper and lower rows correspond to the analysis string and the reference string in the mode view, respectively. The color coding remains consistent. Each row contains three columns, representing the characteristics of the photovoltaic string under low irradiation, medium irradiation, and high irradiation conditions. When the user hovers the mouse over the boundary line of the radar chart, the detailed values ​​of each feature will be displayed.

[0109] The right side of the analysis view module also provides a data annotation interface, where users can correct erroneous results identified by the anomaly detection model to optimize subsequent model iterations. The annotation form contains four parts: string ID, anomaly label, inferred cause, and remarks. The string ID is used to display the selected PV string number; the anomaly label indicates whether the selected PV string has an anomaly, and the initial value is the identification result of the anomaly detection model, which users can modify based on the actual analysis results; the inferred cause is a multiple-choice option, indicating the inefficient cause inferred through time series current data analysis. Common causes of PV string inefficiency include the following four:

[0110] 1.0 current: When the PV string is in sufficient weather, the current value is 0 for a long period of time (such as 1 day or 1 week);

[0111] 2. Dust accumulation or shading: Due to dust accumulation or shading (such as trees, shadows, etc.), the power generation of photovoltaic strings is low for a long time and does not change with the weather;

[0112] 3. Internal failure: Failure of some components of the PV string (such as short circuit, fragmentation, hot spot, etc.), resulting in a decrease in power generation, but proportional to weather changes.

[0113] 4. Single port connected to two strings: Due to engineering problems, one inverter port is connected to two PV strings, resulting in a 2:1 ratio between the PV string current and the normal PV string current.

[0114] The Remarks column is used to add comments for the PV string analysis for subsequent expert model training. After filling in the form, click the "Annotate" button to complete the string annotation task, and finally generate a PV string dataset containing the annotation results through the visual analysis system export button.

[0115] In order to better illustrate the technical effect of the present invention, anomaly detection is performed using real operating data collected from a centralized photovoltaic power station. The total installed capacity of the photovoltaic power station is about 25MWp, and the overall architecture adopts a solution of string inverter, on-site boost, and centralized grid connection. The power station uses more than 50,000 high-efficiency single crystal components of the same specifications, and is built in a hierarchical structure of components-strings-inverters-box transformers, and the data set is collected through the supporting photovoltaic power station NCS background monitoring system. The data set includes two parts: time-series electrical quantity data and weather data. The time-series electrical quantity data includes 8 box transformers, 75 inverters and 1350 strings. The DC side current and voltage, AC side three-phase current and voltage, active power, reactive power, IGBT temperature, daily power generation and total power generation of photovoltaic strings during operation from December 9, 2022 to March 26, 2023 with a time granularity of 1 minute are collected. The numbering format of photovoltaic strings is BT[box transformer ID]-I[inverter ID]-PV[string ID]; the weather data is the global irradiation, temperature and wind speed of the entire photovoltaic power station in the same time period. Related environmental factors that affect the power generation performance of photovoltaic strings are obtained by sampling at a frequency of 1 minute. Based on the developed photovoltaic string anomaly detection method and visualization analysis system, the following tasks are performed:

[0116] (1) Explore the inefficient strings in power plants;

[0117] After the visualization analysis system loads the collected PV power station data set, it first analyzes the global view module to obtain the recognition results of the anomaly detection model for the inefficient state of the entire PV power station string. Figure 4 As shown in A1 in the figure, the distribution of abnormal values ​​of photovoltaic strings displayed by the frequency histogram of the global view module shows that the abnormal values ​​of most photovoltaic strings are around 0.1, and the abnormality discrimination threshold automatically calculated by the anomaly detection model based on the 3σ criterion statistical method is 0.20. Observe the position of the red threshold line in the histogram to confirm that the discrimination threshold reasonably distinguishes whether the photovoltaic strings are abnormal or not. Next, Figure 4 As shown in A2 in the figure, the distribution of abnormal PV strings in physical locations is explored through the rectangular tree diagram of the global view module. The mapping relationship between the color and the abnormal degree of the PV strings in the reference scale is used. The darker the color of the rectangular unit, the more abnormal the corresponding box transformer is. By comparison, it is found that the box transformers BT002, BT004 and BT005 are in serious abnormal conditions. Clicking to enter the next level, it is found that the inverters BT002-I009 and BT002-I012 under the BT002 box transformer, the inverter BT004-I004 under the BT004 box transformer, and the inverter BT005-I017 under the BT005 box transformer are in serious abnormal conditions.

[0118] Furthermore, in order to understand the current situation of the PV strings under the inverter with more serious abnormal conditions, such as Figure 4 As shown in A2 in Figure 1, after selecting a specified inverter, such as the darkest BT005-I017, the real-time current status of all PV strings from PV1 to PV18 under the inverter is displayed in the sorting view module ( Figure 4 According to the outlier value R a Arrange in descending order and find R a There are two PV strings whose current is greater than the abnormal discrimination threshold of 0.2, namely PV18 and PV13, and the corresponding abnormal values ​​are 1 and 0.26. Checking the real-time current timing diagram, it is found that the current of PV18 is 0, while the overall amplitude of PV13 current is small, which obviously indicates that there is an abnormality, indicating that the anomaly detection model can identify the inefficient PV strings in the data set.

[0119] (2) Mining the characteristics of the cluster dimension reduction pattern graph;

[0120] In the Global View module, select inverter BT005-I017 as the analysis object ( Figure 4 A2 in Figure 4 As shown in B in the figure, the hierarchical view module shows the combination of all PV string dimensionality reduction pattern diagrams in the inverter. The points in the combination pattern diagram are divided into high ( Figure 4 As shown in B1), Figure 4 As shown in B2), low irradiation ( Figure 4 There are three types of PV strings (shown in B3 in the figure), and the corresponding normal cluster distribution ranges. From the perspective of the distribution boundaries of each type, the anomaly detection model fitting the inverter divides and combines the PV strings under the three types of weather conditions in two-dimensional space. From the bottom to the top in the vertical direction, they are arranged as low irradiation, medium irradiation, and high irradiation. There is a partial intersection phenomenon in the distribution boundaries of each part, among which the intersection of high irradiation and medium irradiation accounts for a large proportion. By analyzing the arrangement of the dimensionality reduction points of various types of PV strings, it can be found that from bottom to top in the vertical direction of the dimensionality reduction space, the better the power generation state corresponding to the scattered points in the dimensionality reduction pattern diagram of the PV strings, that is, the height of the dimensionality reduction space is proportional to the power generation performance of the PV strings on that day, and the same observation results can be obtained by replacing different inverter objects.

[0121] In addition, if Figure 4 As shown in B, the cross-overlap in the PV string BT005-I017 indicates that the anomaly detection model has errors in the judgment boundary of weather condition clustering, among which the overlap between high irradiation and medium irradiation is relatively higher. The possible reason is that the fluctuation range of low irradiation is smaller than that of medium irradiation, and the difference is more obvious.

[0122] like Figure 7 As shown in the figure, further explore and analyze each photovoltaic string under the photovoltaic string, and sort the photovoltaic strings according to the abnormal value R in the sorting view module. a To sort in descending order, select R a The lowest two strings PV16 (R a =0) and PV7(R a =0) as the analysis string (left) and reference string (right). Through observation, it is found that the distribution of points in the dimensionality reduction pattern of normal strings under three types of weather conditions also follows the rule that the height of the dimensionality reduction space is proportional to the current of the photovoltaic string. The abnormal star-shaped data points in the anomaly detection model deviate from the data cluster to which they belong in terms of spatial distance, which indirectly confirms the reliability of the anomaly detection model. The analysis string is modified to R a The highest PV18, the PV string current is almost 0, R a =1 means that the abnormal detection model believes that the current performance of the photovoltaic string is abnormal every day. At the same time, all the scattered points in the dimensionality reduction pattern are almost concentrated in the same local area, which is consistent with the actual situation. a =0.26), such as Figure 4 As shown in D, the timing current diagram of PV13 and PV7 in the comparison mode view module ( Figure 4 D1 in Figure 1 shows that the current data of PV13 at some time points are all lower than that of the normal PV string PV7, indicating that the overall power generation performance is poor. Figure 4 D2 and Figure 4 As shown in D3 in the figure, observing the dimensionality reduction pattern corresponding to PV13, it is found that the red part (high irradiation) and the green part (medium irradiation) of the photovoltaic string have an overall downward shift compared with the reference string on the right, such as high irradiation scatter points appear in the medium irradiation range, and medium irradiation scatter points appear in the low irradiation range, and all points with severe downward shifts are identified as abnormal points by the anomaly detection model. According to the rule that the spatial height is proportional to the photovoltaic string current, it means that the power generation performance of the photovoltaic string is lower than the normal state during the high and medium irradiation time periods, while it remains normal under low irradiation conditions. Mapping back to the time series current diagram for comparison and verification, it is found that the actual results are consistent with the above inference.

[0123] (3) Comparative analysis of string abnormalities;

[0124] In the sorting view module, sort by outlier R a Arrange in descending order, select the photovoltaic strings with high abnormal values, combine with the time series electrical quantity data, combine with the radar chart for detailed analysis and simple annotation ( Figure 4 The following are the specific situations of various abnormalities during the analysis:

[0125] (3.1) Long-term zero current. For example, the photovoltaic string BT005-I017-PV1 8. This type of anomaly is manifested in the time series current as a current of 0 or close to 0 within a certain period of time (several days or more) when the environmental factors such as radiation are sufficient. The dimensionality reduction pattern diagram in the pattern view module is manifested as data points clustered in a smaller range on both sides of the dimensionality reduction space. The corresponding radar chart in the analysis view shows a single point under the three types of weather conditions, that is, the value of each indicator is 0. When marking this type of data, check the "Long-term zero current" option to infer the cause.

[0126] (3.2) Dust accumulation or obstruction. For example, BT002-I008-PV2. By analyzing the radar chart on the left side of the analysis view module, we found that the radiation correlation coefficient of this type of inefficient photovoltaic strings has decreased. For example, the radiation correlation coefficient of BT002-I008-PV2 is 0.94, 0.75 and 0.73 under low, medium and high weather conditions, respectively, indicating that the real-time current value of the photovoltaic string is less affected by weather radiation. This may be due to the obstruction of grass or buildings, the accumulation of dust and feces, etc., which reduces the amount of weather radiation exposed to the photovoltaic string. The stronger the weather radiation, the greater the impact. However, there is still an approximately proportional relationship overall. Corresponding to the dimensionality reduction pattern diagram, the scattered points of medium and high radiation appear to move downward in large numbers and are identified as abnormalities by the anomaly detection model. In addition, the relative power generation rates of the photovoltaic strings were 0.47, 0.33 and 0.2 respectively, indicating that the current value of the photovoltaic strings has decreased compared with the normal photovoltaic strings, and the stronger the weather radiation, the greater the degree of decrease in the current value of the photovoltaic strings. Compared with the results of the time series current diagram in the mode view module, there is indeed a phenomenon that the overall current amplitude decreases and the deviation is greater in the high radiation part. When marking this type of data, check the "dust accumulation or obstruction" option to infer the cause.

[0127] (3.3) Internal defects. For example, BT001-I008-PV1. Compared with the second type of anomaly, the causes of this type of anomaly include diode breakdown, surface cracking or hot spots, which will cause some components of the photovoltaic string to fail to work properly. While reducing the current amplitude of the photovoltaic string, it also seriously affects the correlation between current and irradiation. The irradiation correlation coefficients corresponding to the string BT001-I008-PV1 are 0.95, 0.73 and 0.41, respectively, which also decrease. The stronger the irradiation, the greater the decrease. This is because internal defects cause some components of the photovoltaic string to fail to work properly. Even if they are exposed to light, they cannot generate current, thereby reducing the impact of irradiation on the current value of the photovoltaic string, and it is more prominent under high irradiation conditions. Similarly, the relative power generation rate is reduced to 0.58, 0.45 and 0.33, and the corresponding medium and high irradiation points in the dimensionality reduction pattern diagram also have a large downward shift and are marked as abnormal. Different from the second type of time series current value, which tends to a lower value regardless of the irradiation level, the current of the internal defective string is proportionally reduced, because the internal defect causes some components of the photovoltaic string to fail to work, which is equivalent to generating electricity with a reduced number of components. When marking this type of data, check the "Internal Defect" option to infer the cause.

[0128] (3.4) Single port connected to two strings. For example, BT004-I003-PV18. This type of anomaly is quite special and was discovered during the on-site investigation. The reason for this is that the PV strings were restricted by the actual environment during installation, resulting in two strings being connected to a single port in the inverter. The timing current of this type of PV string is 2:1 compared to the normal PV string, and all the scattered points in the entire dimensionality reduction pattern graph move upward. In addition, the relative power generation rate of this type of PV string in the mode view module is close to 1, while that of the normal PV string is close to 0.5. When marking this type of data, check the "Single port connected to two strings" option to infer the cause.

[0129] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A hierarchical photovoltaic string anomaly detection method based on dimensionality reduction mode, characterized in that: The following steps are involved: Collect the time-series electrical quantity data and weather data of photovoltaic strings and perform data preprocessing; Based on the preprocessed time series electrical quantity data, a dimension reduction pattern diagram of the time series electrical quantity data is calculated; Based on the preprocessed weather data, weather labels are formed by clustering, and the weather labels are introduced into the dimension reduction pattern diagram of the time series electrical quantity data to obtain the dimension reduction pattern diagram of the time series electrical quantity data with weather labels; Based on the dimensionality reduction pattern diagram of time-series electrical quantity data with weather labels, a hierarchical approach is used to train anomaly detection models. The trained anomaly detection model is used to identify abnormal PV strings, and the anomaly detection results of all PV strings are obtained.

2. The hierarchical photovoltaic string abnormality detection method according to claim 1, characterized in that: The data preprocessing includes: error value processing, missing value filling and noise filtering; The error value processing means that when the sequential electrical quantity data exceeds the rated value or a negative value appears, it is considered that there is an error value and the error value is eliminated; The missing value filling method fills a large number of continuous empty values ​​appearing in the time series electrical quantity data and weather data of the photovoltaic strings by an interpolation method; The noise filtering smoothes the filtering of the time-series electrical quantity data and weather data of the photovoltaic strings by a sliding average method, thereby eliminating the interference of weather factors.

3. The hierarchical photovoltaic string abnormality detection method according to claim 1, characterized in that: The calculation to obtain the dimension reduction pattern diagram of the time series electrical quantity data includes: performing a dimension transformation operation on the preprocessed time series electrical quantity data and obtaining the dimension reduction pattern diagram of the time series electrical quantity data based on UMAP calculation.

4. The hierarchical photovoltaic string abnormality detection method according to claim 1, characterized in that: The clustering of weather data based on photovoltaic strings to form weather labels includes: clustering and color-coding the preprocessed weather data using the KKMeans clustering method to form weather labels, wherein blue represents low irradiation conditions, green represents medium irradiation conditions, and red represents high irradiation conditions.

5. The hierarchical photovoltaic string abnormality detection method according to claim 1, characterized in that: The method of training the anomaly detection model in a hierarchical manner includes: using the dimensionality reduction pattern diagram of the time-series electrical quantity data with weather labels as the input of the anomaly detection model, grouping the dimensionality reduction pattern diagrams of all photovoltaic strings according to the inverters they belong to, and then grouping the dimensionality reduction pattern diagrams of the photovoltaic strings under each group of inverters based on the weather labels, and training an anomaly detection model separately for the dimensionality reduction pattern diagram of each type of weather label.

6. The hierarchical photovoltaic string abnormality detection method according to claim 5, characterized in that: The anomaly detection model is a combination of three unsupervised models: a single-class support vector machine, an elliptical envelope, and a local anomaly factor. The weighted average of the three anomaly models is used as the output of the anomaly detection model.

7. The hierarchical photovoltaic string abnormality detection method according to claim 1, characterized in that: The identification of abnormal photovoltaic strings is performed based on the output results calculated by the abnormal detection model, and the output results include: abnormal value R a and degradation rate R d ; The outlier value is obtained by selecting the anomaly detection model architecture and normalizing the results using the MinMax method to quantify the degree of abnormality of each photovoltaic string. The outlier value R a It is expressed as: Among them, Out represents the output result of the anomaly detection model of the i-th photovoltaic string, the intermediate variable m=OCSVM|EE|LOF represents that the selected anomaly detection model is a single-class support vector machine, a local anomaly factor and an elliptical envelope, n represents the number of photovoltaic strings; Min represents the minimum value of the mean value of the output results of the anomaly detection model of n photovoltaic strings; Max represents the maximum value of the mean value of the output results of the anomaly detection model of n photovoltaic strings; The degradation rate is obtained by quantifying the relative value of each point in the dimension reduction pattern to the normal data cluster C. N The method uses the same training method as the anomaly detection model to identify abnormal photovoltaic strings based on the method of fitting the distribution boundary of the normal dimensionality reduction pattern, and selects the Gaussian mixture model as the model structure. The Gaussian mixture model approximates the normal data cluster C through the weighted sum of multiple Gaussian distributions. N The probability distribution p(x) is used to calculate the degradation rate R d The MinMax method is used to normalize the calculation results of the Gaussian mixture model to the range of [0, 1]; and the degradation rate R d The groups are grouped by time t, and the minimum likelihood of all normal PV strings in each group is used as the boundary. The ratio of the likelihood of all normal PV strings in the group to the minimum likelihood is calculated, and the weighted sum of the ratios of the PV strings at each time point is used to calculate the approximate degradation rate R. d , degradation rate R d It is expressed as: Among them, x represents a multidimensional data point in the dimensionality reduction pattern graph, x∈C N , C N Represents normal data clusters in the dimensionality reduction pattern graph; Represents the likelihood of the parameter θ, which is used to estimate the value that can make x equal to θ when the multidimensional data point x in the dimensionality reduction pattern is known. is numerically identical to the probability distribution p(x); θ represents the parameter set of the Gaussian mixture model, including the mixture weight, mean, and covariance matrix; K represents the number of Gaussian distributions, and π k represents the mixture weight of the k-th Gaussian distribution, represents the kth Gaussian distribution, μ k is the mean of the kth Gaussian distribution, ∑ k is the covariance matrix of the kth Gaussian distribution; MinMax indicates that when calculating the degradation rate R d The MinMax method is used to normalize the calculation results of the Gaussian mixture model to the range of [0, 1], and max is represented by the maximum value in the range of [0, 1]; R d (t) represents the degradation rate R d Group by time t, the time range is [1, n]; represents the likelihood of grouping by time t, and the minimum likelihood of all normal PV strings in each group is expressed as 8. A visual analysis system for a hierarchical photovoltaic string anomaly detection method based on a dimensionality reduction model, characterized in that: The system uses the hierarchical photovoltaic string anomaly detection method based on dimensionality reduction mode according to any one of claims 1 to 7 in visual analysis, and the visual analysis system includes: a global view module, a hierarchical view module, a sorting view module, a mode view module and an analysis view module; The global view module provides a frequency histogram and a rectangular tree diagram for displaying the recognition results of the anomaly detection model on the input data, wherein the frequency histogram displays the distribution of abnormal values ​​of all photovoltaic strings in the photovoltaic power station calculated by the anomaly detection model, and provides an overall overview of the abnormal photovoltaic strings identified by the anomaly detection model; the rectangular tree diagram displays the abnormal values ​​of all photovoltaic strings in the photovoltaic power station based on the hierarchical structure of the photovoltaic power station box transformer-inverter-photovoltaic string; The hierarchical view module is used to display the overall distribution of the dimension reduction pattern diagrams of all photovoltaic strings under the inverter in the form of scatter plots and contour plots, and to obtain the distribution boundary of the dimension reduction pattern diagram of normal photovoltaic strings based on the distribution boundary fitting algorithm of the dimension reduction pattern diagram of normal strings, so as to help users understand the reasons why the abnormal detection model identifies the photovoltaic strings as abnormal and further optimize the abnormal detection model; The sorting view module is used to display the original current data of each photovoltaic string under the selected inverter in the global view and the output results of the abnormal detection model calculation, and provides filtering and sorting functions to help users quickly locate the photovoltaic strings of interest; The mode view module is used to display the original time series electrical quantity data and the dimension reduction mode diagram of the photovoltaic strings selected in the sorting view by providing line graphs and scatter graphs, so as to help users explore the similarities and differences between normal and abnormal photovoltaic strings in the dimension reduction mode diagrams, and provide reasonable explanations for the recognition results of the anomaly detection model; The analysis view module is used to provide relevant abnormal characteristic analysis indicators of the specified photovoltaic string and a simple data annotation function to assist the user in analyzing the causes of the photovoltaic string abnormality.

9. The visual analysis system according to claim 8, characterized in that: The method of obtaining the distribution boundary of the normal photovoltaic string dimensionality reduction pattern diagram based on the normal string dimensionality reduction pattern diagram distribution boundary fitting includes: Based on the normal photovoltaic strings identified by the anomaly detection model, the Gaussian mixture model is used to generate a contour map to depict the boundaries of the normal photovoltaic string distribution. The contour line boundary value k of the contour map is calculated by the k-sigma method, indicating that the contour line is calculated based on the μ-kσ of the photovoltaic string, and the value range of k is [1, 5]; The further optimized anomaly detection model is optimized by adjusting the contour line boundary value k of the contour map and the number of sub-distributions in the Gaussian mixture model.

10. The visual analysis system according to claim 8, characterized in that: The provision of relevant abnormal characteristic analysis indicators for a specified photovoltaic string includes: providing a normal rate R n , relative power generation rate R rpg and irradiation correlation coefficient C Irr Indicators and visualize them through radar charts to facilitate estimation of the values ​​of each indicator; The normal rate R n Indicates the proportion of normal PV strings in the dimension reduction mode diagram; The relative power generation rate R rpg is the PV string S under the same inverter and on a certain day relative to the optimal PV string S max The ratio of the power generation P is expressed as: Where N represents the number of days, P s，d represents the power generation of the PV string S under the same inverter and on the dth day, P max，d represents the optimal photovoltaic string S max Power generation under the same inverter and on day d; The irradiation correlation coefficient C Irr is the Pearson correlation coefficient between the time series current data I and the irradiation data R of the photovoltaic string, which ranges from [0, 1] and is expressed as: For a normal photovoltaic string S, the current data I is usually proportional to the irradiation data R, and the irradiation correlation coefficient C Irr The closer it is to 1, the more normal the PV string is. N represents time, I represents s，t It represents the time series current data of a normal photovoltaic string S at time t. represents the average value of the time series current data of a normal photovoltaic string S, R t represents the irradiation data of the normal photovoltaic string S at time t, represents the mean value of the irradiation data of the normal photovoltaic string S, and max represents the maximum value in the range of [0, 1] as the Pearson correlation coefficient; The visualization of the indicator by radar chart includes: setting the center of the radar chart to represent the indicator value as 0, the outermost vertex to represent the value as 1, the smaller the area of ​​the radar chart, the more abnormal the photovoltaic string is, and the radar chart represents the relevant abnormal characteristics of the specified photovoltaic string under low irradiation, medium irradiation and high irradiation conditions through unified color coding.

Images