CN-121980354-A - Photovoltaic fault diagnosis method, system, equipment and medium

CN121980354ACN 121980354 ACN121980354 ACN 121980354ACN-121980354-A

Abstract

The invention relates to the technical field of photovoltaic operation safety guarantee, in particular to a photovoltaic fault diagnosis method, a system, equipment and a medium, wherein the method comprises the steps of injecting a plurality of fault modes into a photovoltaic simulation system, collecting simulation operation data, and constructing a basic RCET-TGATEDGNN model based on the simulation operation data; the method comprises the steps of inputting actual operation data of a real photovoltaic system into a basic RCET-TGATEDGNN model as a verification sample to obtain a verification diagnosis result, classifying the verification sample based on a manual confirmation result of the verification diagnosis result to obtain a classification sample data set, obtaining a target RCET-TGATEDGNN model based on a classification sample data set optimization basic RCET-TGATEDGNN model, and obtaining photovoltaic fault information of the photovoltaic system to be diagnosed based on the operation data and the target RCET-TGATEDGNN model. The invention can realize the classification and positioning of faults of the photovoltaic array equipment.

Inventors

SHEN CHUNMING
ZHAO WEN
HAN FANGWEI
DENG ZHE
TANG XIRU
LIU KEHUI

Assignees

北京市科学技术研究院

Dates

Publication Date: 20260505
Application Date: 20260205

Claims (10)

1. A method of diagnosing a photovoltaic failure, comprising: Constructing a photovoltaic simulation system, injecting a plurality of fault modes into the photovoltaic simulation system, collecting simulation operation data, and constructing a foundation RCET-TGATEDGNN model for photovoltaic fault diagnosis based on the simulation operation data; acquiring actual operation data of a real photovoltaic system, inputting the actual operation data as a verification sample into the basic RCET-TGATEDGNN model to obtain a verification diagnosis result, manually confirming the verification diagnosis result, and classifying the verification sample based on the manual confirmation result to acquire a classification sample data set; Optimizing the basic RCET-TGATEDGNN model based on the classified sample dataset to obtain a target RCET-TGATEDGNN model; and acquiring operation data of the photovoltaic system to be diagnosed, and acquiring photovoltaic fault information of the photovoltaic system to be diagnosed based on the operation data and the target RCET-TGATEDGNN model.
2. The method of claim 1, wherein the simulated operational data comprises fault characteristics and fault types; The fault characteristics comprise basic electrical parameters, curve morphological characteristics, slope and geometric characteristics, comprehensive performance indexes and statistical characteristics; the basic electrical parameters include open circuit voltage Short circuit current Maximum power point voltage Maximum power point current Maximum output power ; The curve morphological characteristics comprise the peak value number of the P-U curve and the peak value number of the I-U curve; The slope and geometric characteristics comprise that the I-U curve is within @, the slope and the geometric characteristics comprise , ) Slope at point and P-U curve at% , ) Slope at point, I-U curve , ) Point and [ (ii) a ] 0) Slope of line between points, I-U curve , ) The point is the sum of (0), ) Slope of the line between points; the comprehensive performance index comprises an I-U curve filling factor; the statistical characteristics comprise an area under an I-U curve, an area under a P-U curve, a current variance, a voltage variance, a power skewness and a power average; The fault types include the most common faults in photovoltaic systems.
3. The method of claim 2, wherein constructing a base RCET-TGATEDGNN model for photovoltaic fault diagnosis based on the simulated operational data comprises: Applying interference to fault features of the photovoltaic simulation system to generate an original data set and an interference data set, calculating interference sensitivity factors of each fault feature based on the original data set and the interference data set, determining anti-noise Type of each fault feature based on the interference sensitivity factors, determining anti-noise Type coding Type i of each sensitivity factor based on the anti-noise Type; wherein, the calculation formula of the interference sensitive factor is as follows, Wherein, the Representing fault characteristics Is a factor of interference sensitivity; representing fault characteristics Standard deviation change rate after treatment; representing fault characteristics Is a characteristic level decay rate of (1); Wherein, the Representing fault characteristics The minimum standard deviation rate of change of (2); representing standard deviation coefficients; representing fault characteristics Is a standard deviation of interference data of (2); representing fault characteristics Standard deviation of raw data of (2); representing fault characteristics Is used to determine the correlation coefficient of the original data, Representing fault characteristics Is a correlation coefficient of interference data; Generating feature nodes of each fault feature based on standardized feature values and anti-noise type codes of the fault feature, and generating edges of the feature nodes based on physical strength and the anti-noise type of the feature nodes, wherein the edges of the feature nodes comprise three edge types; Generating a three-channel network structure diagram based on the characteristic nodes and the edge types, and initializing the weight of each edge; Performing Type-Scaling Type driving Scaling based on the anti-noise Type coding Type i , after Scaling the input features of each node in different channels, sending the network structure diagram of each channel into a convolution layer of the corresponding channel in parallel, dynamically and adaptively regulating and controlling weights of three channels on feature nodes through a Type gating layer, and performing adaptive fusion of multi-channel features to obtain fusion features; And processing the fusion features sequentially by a pooling layer, a classification layer and a loss layer, and finally outputting a fault classification result to complete the construction of a basic RCET-TGATEDGNN model.
4. A method of diagnosing a photovoltaic fault according to claim 3, wherein determining the noise immunity type for each fault feature based on the disturbance sensitive factor comprises: Setting the anti-noise type of the fault characteristic of which the interference sensitivity factor is smaller than a first preset threshold value as a strong anti-noise type; Setting the anti-noise type of the fault characteristic of which the interference sensitivity factor is greater than or equal to a first preset threshold value and less than a second preset threshold value as a moderate sensitive anti-noise type; Setting the anti-noise type of the fault characteristics with the interference sensitivity factor larger than or equal to a second preset threshold value as a high-sensitivity anti-noise type; The strong anti-noise type is encoded as [1,0], the medium sensitive anti-noise type is encoded as [0,1], and the high sensitive anti-noise type is encoded as [0,0].
5. The method of claim 3, wherein the edges of the feature nodes comprise physical identity/geometric edges, statistically robust edges, noise-prone auxiliary edges; The physical identity/geometric edge represents that a physical conservation relation or a collection relation exists between two fault characteristic nodes; the statistical robust edge represents that the pearson correlation coefficient of two fault characteristic nodes is more than 80% and the absolute value of variation of the pearson correlation coefficient under the interference condition is less than 15%; the easy-to-noise auxiliary edge represents that the correlation between the characteristics of two fault characteristic nodes is not obvious or the correlation between the characteristics is obvious but the attenuation under the interference is obvious.
6. The method of claim 1, wherein classifying the verification sample based on the manual validation result to obtain a classified sample dataset, comprising: Acquiring target operation data of an actually operated photovoltaic system, and preprocessing the target operation data to acquire a target data set; Inputting feature nodes of each fault feature to the basic RCET-TGATEDGNN model based on the target dataset, wherein the feature nodes of the fault feature comprise standardized feature values of a current sample and anti-noise type codes corresponding to interference sensitive factors of the node under the version of the current basic RCET-TGATEDGNN model; manually checking whether the output result of the basic RCET-TGATEDGNN model is higher than a confidence threshold value and is consistent with the actual fault condition, wherein the manual checking refers to that operation and maintenance personnel confirms the model output based on-site inspection, instrument retest or operation and maintenance record/alarm information; if the output result is higher than the confidence threshold and is consistent with the actual fault condition, classifying the verification sample as a class A verification sample; if the output result is lower than the confidence threshold or conflicts with the actual fault condition, and no new fault feature or new fault feature combination exists, classifying the verification sample as a B-class verification sample; and if the output result does not belong to the existing fault type, classifying the verification sample into a C-type verification sample.
7. The method of claim 6, wherein optimizing the base RCET-TGATEDGNN model based on the classified sample dataset to obtain a target RCET-TGATEDGNN model comprises: Training the basic RCET-TGATEDGNN model based on the class A verification sample and the class B verification sample when the accumulation of the class A verification sample and the class B verification sample reaches a preset scale, and outputting a target RCET-TGATEDGNN model when the accuracy of the model output result reaches a preset accuracy threshold; when the environment or the environment interference change is monitored to exceed a preset condition or the number of C-class verification samples exceeds an optimization threshold, redefining the fault characteristics and the fault types, retraining the basic RCET-TGATEDGNN model based on the updated fault characteristics and the updated fault types, and further training to obtain the target RCET-TGATEDGNN model based on the updated basic RCET-TGATEDGNN model.
8. A photovoltaic fault diagnosis system, comprising: the basic model construction module is used for constructing a photovoltaic simulation system, injecting various fault modes into the photovoltaic simulation system, collecting simulation operation data, and constructing a basic RCET-TGATEDGNN model for photovoltaic fault diagnosis based on the simulation operation data; The classification sample acquisition module is used for acquiring actual operation data of the real photovoltaic system, inputting the actual operation data into the basic RCET-TGATEDGNN model as a verification sample to obtain a verification diagnosis result, manually confirming the verification diagnosis result, and classifying the verification sample based on the manual confirmation result to acquire a classification sample data set; the target model construction module is used for optimizing the basic RCET-TGATEDGNN model based on the classification sample data set to obtain a target RCET-TGATEDGNN model; The photovoltaic fault diagnosis module is used for collecting operation data of the photovoltaic system to be diagnosed and obtaining photovoltaic fault information of the photovoltaic system to be diagnosed based on the operation data and the target RCET-TGATEDGNN model.
9. An electronic device comprising a processor, a memory, and a computer program stored on the memory and executable on the processor, the computer program implementing the photovoltaic fault diagnosis method according to any one of claims 1-7 when executed by the processor.
10. A computer readable storage medium, characterized in that the computer readable storage medium has stored thereon a computer program which, when executed by a processor, implements the photovoltaic fault diagnosis method according to any of claims 1-7.

Description

Photovoltaic fault diagnosis method, system, equipment and medium Technical Field The invention relates to the technical field of photovoltaic operation safety guarantee, in particular to a photovoltaic fault diagnosis method, a system, equipment and a medium. Background At the moment of the vigorous development of the photovoltaic industry, the stable operation of a photovoltaic system is important for guaranteeing energy supply and improving energy utilization efficiency. Photovoltaic fault diagnosis is always a research hotspot in the field as a key link for ensuring reliable operation of a photovoltaic system. At present, a photovoltaic fault diagnosis method based on deep learning emerges a plurality of typical schemes. The Chinese patent application with publication number of CN120046051A proposes a fault diagnosis method of a photovoltaic inverter based on GA-LSTM-GPR, which optimizes the attention weight of an LSTM network by means of a genetic algorithm, and builds a fault diagnosis model by combining Gaussian process regression, and mainly relies on electrical parameters such as voltage, current, power and the like to carry out fault detection work. The Chinese patent application with publication number of CN119442021A gives a photovoltaic array fault diagnosis method based on soft voting integration of a composite model, integrates a plurality of sub-models through a soft voting integration algorithm, determines super parameters by utilizing a global optimizing algorithm, and mainly performs fault diagnosis based on P-V curve data. The Chinese patent application with publication number of CN118763993A proposes a photovoltaic string fault diagnosis method based on reference state selection, wherein the reference state string is determined by calculating probability density curve indexes of the generated energy and conversion efficiency of the photovoltaic string, and then the relative deviation between other strings and the reference string is calculated, and the fault diagnosis is performed by combining statistical characteristics and distribution characteristics. The Chinese patent application with publication number of CN120471101A gives a photovoltaic array fault diagnosis method based on LSTM-CNN-ArcLoss fusion network, builds a photovoltaic array intelligent fault diagnosis model through LSTM-CNN-ArcLoss fusion network algorithm, and mainly relies on time sequence data to carry out fault classification. However, the above solution exposes a number of drawbacks in practical applications. First, environmental interference sensitivity issues are prominent. Most of the prior art mainly depends on electrical parameters (such as IV curves, voltage and current, etc.) or image data, and the actual running environment of the photovoltaic system is complex and changeable, and environmental factors such as illumination intensity, temperature, shadow shielding, etc. can have significant influence on the data, so that the diagnosis precision is greatly reduced in the complex environment. Second, feature robustness is not sufficient. Traditional features, such as current, voltage, etc., have poor stability when being disturbed by the outside, and in order to obtain reliable feature information, complicated preprocessing steps or additional sensor data are often needed, which not only increases the complexity and cost of the system, but also may introduce new errors. Furthermore, physical constraints are not well-coupled with data driving. The existing models, such as LSTM, CNN, soft voting integration and the like, are mainly based on data driving for modeling, lack of explicit modeling on physical rules, so that the model has poor interpretability, and the requirements of operation and maintenance personnel on deep understanding of fault causes are difficult to meet. In addition, the weak dynamic adaptation capability is also a big and short board in the prior art. The existing method, such as reference state selection, generally depends on a static threshold value or a complex optimization algorithm, is difficult to actively adapt to the dynamic change of environmental interference, and the diagnosis performance is seriously affected when the environmental condition is changed rapidly. Finally, interference immunity and interpretability are difficult to be compatible. In the process of improving the self anti-interference capability in the prior art, the interpretation of a model is often sacrificed, and the requirement of the operation and maintenance of a photovoltaic power station on high-reliability diagnosis is required to be met, so that the fault diagnosis method has strong anti-interference capability and can provide clear and easily-understood diagnosis results, and the prior art obviously cannot meet the requirement. Disclosure of Invention The invention aims to solve at least one technical problem and provide a photovoltaic fault diagnosis method, a system, equipment and a medium, which aim t