Method, device and equipment for generating abnormal test data of photovoltaic monitoring system and medium

Abstract

The embodiment of the invention relates to the field of photovoltaic technology, and discloses a photovoltaic monitoring system abnormal test data generation method, device, equipment and medium, and the method comprises the steps: constructing a GAN model, constructing a fusion framework of a knowledge graph and a probability graph model, and constructing an ontology model of the photovoltaic field, establishing a probability graph model according to a constraint relationship in the ontology model, iteratively training the GAN model until convergence, outputting a preset number of evaluation test samples of a preset photovoltaic fault type, inputting the preset number of evaluation test samples into the probability graph model, and outputting a multi-dimensional influence probability corresponding to each evaluation test sample, and obtaining a business influence degree corresponding to the evaluation test sample according to the multi-dimensional influence probability, and screening the to-be-evaluated test sample according to the business influence degree to obtain test data. The test data generated by the embodiment of the invention can be widely applied to test and verification of functions such as data cleaning, fault-tolerant processing, fault detection and the like of the photovoltaic monitoring system.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic technology, specifically to a method and system for generating abnormal test data in a photovoltaic monitoring system. Background Technology

[0002] The data monitoring system of a photovoltaic power station needs to process massive amounts of data every day, including environmental parameters, equipment operating parameters, and geographical location information. When equipment malfunctions during the actual operation of the photovoltaic power station, if the data monitoring system fails to identify and process the abnormal data in a timely manner, it will lead to incorrect judgments, affecting the operation and maintenance efficiency of the power station and causing economic losses.

[0003] Existing methods for testing the data processing capabilities of photovoltaic (PV) monitoring systems primarily rely on limited anomaly scenarios generated in real-world environments, or on manually constructed or randomly generated anomaly data based on simple rules. This data is then input into the PV monitoring system for processing, and the system's data processing capabilities are evaluated based on the results. While test data generated from limited anomaly scenarios in real-world environments offers high realism and reliability, it suffers from limitations and incompleteness. Manually constructed or randomly generated test data can provide diverse anomaly scenarios, but lacks realism, and the amount of data generated is insufficient to comprehensively evaluate the PV monitoring system's data processing capabilities. Therefore, there is an urgent need for a method that can generate comprehensive and realistic test data to meet the requirements for evaluating the data processing capabilities of PV monitoring systems. Summary of the Invention

[0004] In view of the above problems, embodiments of the present invention provide a method, apparatus, equipment and medium for generating abnormal test data of photovoltaic monitoring system, which is used to solve the problem that the amount of test data generated for photovoltaic monitoring system testing in the prior art is small and not comprehensive enough.

[0005] According to one aspect of the present invention, a method for generating abnormal test data of a photovoltaic monitoring system is provided, the method comprising: Constructing a GAN model includes building a generator network and a discriminator network. The generator network includes a physical constraint module, a multimodal encoder-decoder structure, and a temporal generation module. The discriminator network includes a physical rule filtering layer, a multimodal discrimination branch, and a temporal analysis module. A framework for integrating knowledge graphs and probabilistic graphical models is constructed, including constructing an ontology model in the photovoltaic field to form a knowledge graph in the photovoltaic field, and establishing the probabilistic graphical model based on the constraint relationships in the knowledge graph; The generator network and the discriminator network are iteratively trained until convergence based on the photovoltaic fault type, the characteristic parameters corresponding to the photovoltaic fault type, and the key features corresponding to the historical operation data of the photovoltaic monitoring system; the generator network that has completed training outputs a preset number of evaluation test samples of predetermined photovoltaic fault types. The preset number of evaluation test samples are input into the probabilistic graph model. The multi-dimensional impact probability corresponding to each evaluation test sample is output according to the mapping of the constraint relationship of the knowledge graph. The business impact degree corresponding to the evaluation test sample is obtained according to the multi-dimensional impact probability. The evaluation test samples are selected based on the business impact to obtain test data for testing the anomaly detection capability of the photovoltaic monitoring system.

[0006] In one optional approach, the step of inputting a preset number of evaluation test samples into the probabilistic graphical model, outputting the multi-dimensional impact probability corresponding to each evaluation test sample based on the mapping of the constraint relationships of the knowledge graph, and obtaining the business impact degree corresponding to the evaluation test sample based on the multi-dimensional impact probability, further includes: The evaluation test sample is input into the probabilistic graphical model, and the influence path corresponding to the evaluation test sample is output. Based on the constraints of the ontology model, the influence path is mapped to domain entity relationships, and the influence path graph corresponding to the evaluation test sample is output based on the domain entity relationships. Based on the impact path diagram, the multi-dimensional impact probability, and the business impact degree, output a test report for the evaluation test sample. The test report includes: fault type, impact path diagram, multi-dimensional impact probability, and business impact degree.

[0007] In one alternative approach, the framework for constructing the fusion of knowledge graph and probabilistic graphical model includes constructing an ontology model in the photovoltaic field to form a knowledge graph in the photovoltaic field, and establishing the probabilistic graphical model based on the constraint relationships in the knowledge graph, including: Construct an ontology model Q=(E, R, A) in the photovoltaic system, where E is the entity set, R is the relation set, and A is the attribute set. The constraint relations of the knowledge graph include the entity set, relation set, and attribute set. Building the probabilistic graphical model based on the constraints in the knowledge graph includes: Based on the relation set and attribute set in the ontology model, a Bayesian structure of a parameter-index causal network is established, wherein the nodes of the Bayesian structure represent system parameters X or business indicators I, and the edges represent conditional probability distributions P(I|X) derived from the relation set R. The Bayesian structure and the ontology model are jointly mapped through shared variable nodes to form a factor graph model G=(V F V B V F V represents the relation-attribute factor node in the knowledge graph. B For the probability distribution nodes in the Bayesian network, the factor graphical model is the probability graphical model; Based on the historical operating data of the photovoltaic system, the factor function parameters corresponding to the factor nodes in the factor graph model are calculated using the maximum likelihood estimation method, thus obtaining the trained probabilistic graphical model.

[0008] In one alternative approach, the step of establishing a Bayesian structure for a parameter-index causal network based on the relation set and attribute set in the ontology model, wherein the nodes of the Bayesian structure represent system parameters X or business indicators I, and the edges represent conditional probability distributions P(I|X) derived from the relation set R, including: Map the attributes in the attribute set A to the system parameter nodes in the Bayesian structure; Map the predefined business metrics to business metric nodes in the Bayesian structure; Based on the relation set R, determine the existence and direction of directed edges between the system parameter nodes and between the system parameter nodes and the business indicator nodes, and obtain the Bayesian structure.

[0009] In an alternative approach, the Bayesian structure and the ontology model are jointly mapped through shared variable nodes to form a factor graph model G=(V F V B V F V represents the relation-attribute factor node in the knowledge graph. B For the probability distribution nodes in a Bayesian network, the factor graphical model is the probabilistic graphical model, including: Each conditional probability distribution in the Bayesian structure is mapped to a factor node in the factor graph model, and the factor node is connected to all variable nodes involved in the conditional probability distribution. The relations in the relation set R are mapped to factor nodes in the factor graph model, and the factor nodes are connected to the variable nodes corresponding to the entities participating in the relations; wherein, the set of variable nodes V B The factor graph model is constructed by consisting of the system parameter nodes and the business indicator nodes.

[0010] In one optional approach, a preset number of evaluation test samples are input into the probabilistic graphical model. Based on the mapping of constraints in the knowledge graph, a multi-dimensional impact probability corresponding to each evaluation test sample is output. The business impact degree corresponding to the evaluation test sample is obtained based on the multi-dimensional impact probability, including: An improved belief propagation algorithm is executed in the factor graph model to obtain the multi-dimensional probability and influence path of the evaluation test sample on the business indicator node; The improvements to the belief propagation algorithm include: An attenuation factor is introduced into the message passing rules of the belief propagation algorithm for improvement. The improved message passing rules include: An attenuation factor α is introduced into the message passing rules from the system parameter node to the functional node to improve the message passing rules: ; An attenuation factor is introduced into the message passing rules from the functional node to the system parameter node to improve the message passing rules: ; The attenuation factor is defined as: , d is the shortest path hop count between the two nodes in the factor graph G, and β is the adjustable attenuation coefficient.

[0011] In an alternative approach, the improvement to the belief propagation algorithm further includes introducing an anomaly feature weight function w(e) into the belief propagation algorithm: Based on the photovoltaic fault type e corresponding to the evaluation test sample, obtain the corresponding abnormal feature weight w(e); When calculating the message passed from the functional node to the parameter node, the message is weighted according to the weight w(e) of the fault type e: .

[0012] In an alternative approach, the improvement to the belief propagation algorithm further includes: executing the improved belief propagation algorithm using a hierarchical iterative strategy to obtain the influence probability and influence path, including: Input the device parameters in the evaluation test sample, perform belief propagation between nodes within a single photovoltaic device, and output device layer messages; The device layer message is input into the system layer, belief propagation is performed between cross-device nodes, and the system layer message is output. The system-level message is input, globally transmitted to the business indicator node, and the impact path and the impact probability are output.

[0013] According to a second aspect of the present invention, a photovoltaic monitoring system abnormal test data generation device is provided, comprising: The Generative Adversarial Network (GAN) module is used to construct GAN models, including building a generator network and a discriminator network. The generator network includes a physical constraint module, a multimodal encoder-decoder structure, and a temporal generation module. The discriminator network includes a physical rule filtering layer, a multimodal discrimination branch, and a temporal analysis module. The evaluation framework construction module is used to construct a fusion framework of knowledge graph and probabilistic graphical model, including constructing an ontology model in the photovoltaic field to form a knowledge graph in the photovoltaic field, and establishing the probabilistic graphical model based on the constraint relationships in the knowledge graph; The test data generation module is used to iteratively train the generator network and the discriminator network until convergence based on the photovoltaic fault type, the characteristic parameters corresponding to the photovoltaic fault type, and the key features corresponding to the historical operation data of the photovoltaic monitoring system; the generator network that has completed training outputs a preset number of evaluation test samples of a predetermined photovoltaic fault type. The business impact assessment module is used to input a preset number of assessment test samples into the probabilistic graph model, output the multi-dimensional impact probability corresponding to each assessment test sample according to the mapping of the constraint relationship of the knowledge graph, and obtain the business impact degree corresponding to the assessment test sample according to the multi-dimensional impact probability. The test data acquisition module is used to filter the evaluation test samples according to the business impact to obtain test data for testing the anomaly detection capability of the photovoltaic monitoring system.

[0014] According to a third aspect of the present invention, a computer device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction that causes the processor to perform the operation of the photovoltaic monitoring system abnormal test data generation method as described in any of the first aspects above.

[0015] According to another aspect of the present invention, a storage medium is provided, the storage medium comprising: storing at least one executable instruction, wherein when the executable instruction is executed on a computer device, the computer device performs the photovoltaic monitoring system abnormal test data generation method as described in any of the first aspects above.

[0016] This invention constructs a Generative Adversarial Network (GAN) model that integrates physical constraints, a multimodal encoder-decoder structure, and a temporal generation module. Combined with a deep fusion framework of knowledge graphs and probabilistic graphical models, it can rapidly and automatically generate test data covering more than several photovoltaic (PV) fault types. This test data is input into a PV monitoring system to test the system's data processing capabilities. This solves the problems of incomplete coverage, low authenticity, and poor consistency caused by traditional test data relying on limited real-world anomaly scenarios or manual construction. By introducing physical constraint mechanisms, temporal correlation preservation strategies, and multimodal data fusion to improve the standard GAN model, it ensures that the generated evaluation test samples conform to the operating rules of the PV system and possess temporal continuity and cross-modal consistency. Through joint reasoning using knowledge graphs and probabilistic graphical models, it achieves quantitative assessment and path tracing of the business impact of the evaluation test samples, significantly improving the relevance, practicality, and evaluation efficiency of the test data. This invention can be widely applied to the testing and verification of data cleaning, fault tolerance processing, and fault detection functions in PV monitoring systems, possessing significant engineering application value and promising prospects for widespread adoption.

[0017] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0018] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating the method for generating abnormal test data of a photovoltaic monitoring system provided in Embodiment 1 of the present invention is shown. Figure 2 A flowchart illustrating the method for generating abnormal test data of a photovoltaic monitoring system provided in Embodiment 1 of the present invention is shown. Figure 3 A schematic diagram of the structure of the photovoltaic monitoring system abnormal test data generation device provided in Embodiment 2 of the present invention is shown; Figure 4 A schematic diagram of the structure of the computer device provided in Embodiment 3 of the present invention is shown. Detailed Implementation

[0019] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Example 1

[0020] Please see Figure 1 and Figure 2 This embodiment 1 provides a method for generating abnormal test data of a photovoltaic monitoring system, the method comprising: S1: Construct a GAN model, including building a generator network and a discriminator network. The generator network includes a physical constraint module, a multimodal encoder-decoder structure, and a temporal generation module. The discriminator network includes a physical rule filtering layer, a multimodal discrimination branch, and a temporal analysis module.

[0021] Improvements to the standard GAN model are made through physical constraint mechanisms, including: The generator network includes a physical constraint module. It introduces the physical laws of photovoltaic systems (such as energy conservation and the photoelectric conversion equation) into the generator of a standard GAN as hard constraints to ensure that the data generated by the generator conforms to basic physical rules. Specifically: By defining constraint functions S is the data sample vector (i.e., the generator output). This includes multiple parameters of the photovoltaic system, such as irradiance, temperature, and power. For example: S=[I rr T amb T mod P dc V dc I dc η inv ,...],I rr T is the irradiance (W / m²). amb T represents ambient temperature (°C). mod P represents the component temperature (°C). dc DC power (W), V dc I is the DC voltage (V). dc For direct current (A), η inv For inverter efficiency. i (S) represents the i-th physical constraint. For example, when S is power-related anomalous data, f1(S) = |P dc -(I rr *A array *η pv )|-ε1,A array η is the area of ​​the photovoltaic array (m²). pv Let f be the component's nominal efficiency, ε1 be the allowable error threshold (W), and P be the total number of physical constraints. i (S)≤0 indicates that the data sample S follows this physical rule, when f i If (S)>0, it means that the data sample S violates the physical rules.

[0022] This is not a fixed value; it needs to be calculated based on historical normal data of the photovoltaic system. The calculation method is as follows: ,in and These are historical data The mean and standard deviation.

[0023] A penalty term is introduced into the standard generator loss function to ensure that the data generated by the generator follows anomalous characteristics without violating basic physical laws. When a data sample S violates physical laws (fi(S)>0), a positive penalty value is generated; when a data sample S does not violate physical laws, the penalty value is 0. This penalty term is: ; The penalty coefficient is... The value of can be adjusted according to the specific scale and constraint strength of the photovoltaic system, and the value range is from 0.1 to 10; p is the total number of physical constraint functions; Z is the generated data output by the generator; Z is the random noise vector of the generator's input layer structure, with dimension k; C is the condition vector, with dimension m, containing feature parameters of the fault type. These are the generator network parameters.

[0024] The generator's total loss function becomes: The original generator loss function is used. The generator weights are updated based on the total loss function. Weight After the update, the generator network A slight change has been made to the internal computational logic, so that the next time a data sample S is used to output data, given the same input noise Z and condition C, it will output a "better" data sample S. This "better" is reflected in the fact that this new data sample S is more likely to fool the discriminator and violates fewer physical rules than the old data sample S.

[0025] The data sample S output by the generator is input into the input layer of the discriminator. The intermediate layer of the discriminator processes the data sample S, and finally, the output layer of the discriminator outputs the scalar D(S) ∈ [0,1] corresponding to the data sample S, representing the probability score of the data sample S input to the input layer. d ),in These are the parameters of the discriminator network.

[0026] A physical rule filtering layer is added to the input layer of the discriminator to filter out sample data S that do not conform to physical laws. Specifically, a physical rule function g(S) is defined, and a filtering mechanism is set up: ;D NN The output of the original discriminator is as follows: after passing through the physical rule filtering layer, data samples S that clearly violate the physical rules are directly marked as "false" (output 0), while compliant data samples are allowed to enter the subsequent neural network discriminator D. NN .

[0027] The data processing logic of the discriminator has changed, as shown below. Where g(S) is the rule function, the physical rule filtering layer indirectly affects the subsequent discriminator backbone network D by changing the distribution of effective input data. NN Gradient calculation and weight update, D NN weight θ d After the update, the internal computational logic of the discriminator network underwent a minor change, making its ability to discriminate input data S that can pass physical filtering more accurate. This "accuracy" is reflected in the fact that the updated discriminator D is better at distinguishing between generated data that "seems real but actually violates physical laws" and real data, because it receives more training signals from compliant data that requires fine-grained discrimination, rather than a large amount of obviously invalid data.

[0028] Improvements to the standard GAN model through a temporal correlation preservation mechanism include: The generator network includes a temporal generation module. A Temporal Generative Adversarial Network (TCGAN) is constructed using a Long Short-Term Memory (LSTM) network and the aforementioned GAN model. By introducing a temporal state transfer mechanism, the generator's output data samples S are ensured to have high-fidelity temporal relevance. The temporal state transfer function is implemented through the forward computation process of the LSTM unit. The temporal state transfer function is: Q_t = H(Q_t-1, S_in_t), where Q_t = (h_t, c_t) represents the state at the current time step, consisting of the hidden state h_t and the cell state c_t. Q_t-1 is the state at the previous time step. S_in_t is the input data at the current time step, formed by concatenating the data S_t-1 generated at the previous time step, the random noise vector Z, and the conditional vector C, i.e., S_in_t = [S_t-1; Z; C].

[0029] The specific calculation of the LSTM state transfer function H includes the input gate, forget gate, output gate, and cell state update: Input Gate: ; Forgotten Gate: ; Output gate: ; New cell state: ; Final cell state: ; Final hidden state: ; It is the sigmoid activation function. This represents element-wise multiplication. and These are trainable parameter matrices and vectors. The forward computation process of the LSTM unit is existing technology and will not be described in detail here.

[0030] The LSTM unit is fed with the state Q_{t-1}=(h_{t-1}, c_{t-1}) from the previous time step and the current input vector S_in_t. Through a series of calculations including the input gate, forget gate, and output gate, a new state Q_t=(h_t, c_t) for the current time step is generated. The hidden state h_t is then fed into an output layer (e.g., a fully connected layer) of the generator. This output layer maps h_t to the generator's final output data point S_t at time step t, where S_t=O(h_t; θ_o), where O is the output layer function and θ_o is the output layer parameter. The output S_t of the current step is used as part of the input for the next time step, while the state Q_t is passed to the LSTM unit at the next time step. This process is repeated until a complete time series {S_1, S_2, ..., S_T} of the required length is generated. The generator receives a random noise vector Z and a conditional vector C describing the anomaly, initializing the hidden state h_0 and the cell state c_0 as zero vectors. Starting at t=1, the initial input S_in_1=[0; Z; C] (where 0 is the initialized data vector) is input into the LSTM network to compute the state Q_1 and generate the first data point S_1. Subsequently, S_in_2=[S_1; Z; C] is used as the input for the next time step, recursively calling the LSTM unit to generate the data sequence {S_1, S_2, ..., S_T} sequentially. This process, through the sequential passing of the state Q_t, ensures the continuity and long-range correlation of the output sequence in the time dimension.

[0031] A time-series analysis module is introduced based on the standard discriminator's feature extraction, enabling the discriminator to analyze the rationality of sample data S from the perspective of the overall time series. The discriminator's feature extraction logic is changed, and the final output feature representation is H=D. FeatExtract (S;θ d The loss L calculated based on feature H D To update the discriminator's weights θ d weight θ dAfter the update, the internal feature extraction logic of the discriminator network D undergoes a slight change, enabling it to extract "better" features for the same input sequence S in the next iteration. The new feature representation is more effective than the old feature representation in revealing irrationalities in the sequence in the time dimension, such as abnormal duration of anomalies, excessively high mutation frequency, or abnormal recovery patterns, thereby making the final discrimination result D(S) more accurate.

[0032] Improvements to the standard GAN model through multimodal data fusion include: Building upon standard generator input processing, a multimodal encoder-decoder structure is introduced. Independent encoders map heterogeneous numerical, categorical, and spatiotemporal data to a unified latent space for fusion, followed by a decoder reconstructing the data into consistent multimodal data. This alters the generator's data fusion logic. Represented as: G(Z, C, F) num F cat F spatio ;θ g ), where F is the multimodal feature extracted from the historical operating data of the photovoltaic system.

[0033] Specifically, the numerical data encoder inputs a numerical feature vector F. num , dimension d num (For example: continuous parameters such as irradiance, temperature, voltage, and current), the output implicitly represents H. num ∈R hhidden Where hhidden is the uniform dimension of the latent space; the input categorical feature vector F in the categorical data encoder cat The dimension is dcat (e.g., weather status, device status flags, communication status, etc.), and the output implicit representation is Hcat∈R. hhidden The dimensions are consistent with those of numerical implicit representations; the spatiotemporal data encoder inputs a spatiotemporal feature tensor F. spatio ∈R T×S×C (e.g., GPS tracks, time-series images, spatial distribution maps, etc.), output implicit representation H spatio ∈R hhidden The three types of implicit representations are weighted and summed according to modal importance: ; Where γ1+γ2+γ3=1, the initial suggested values ​​are γ1=0.5, γ2=0.3, and γ3=0.2.

[0034] The decoder in the generator receives the fused implicit representation H fusion The data is reconstructed into multimodal data output S through deconvolution or FC layers.

[0035] The total loss function is calculated based on the features obtained after multimodal fusion, and is expressed as: L = γ4L num +γ5L cat +γ6L spatio (γ4+γ5+γ6=1), where L_num is the mean squared error (MSE) loss for numerical data; L_cat is the cross-entropy loss for categorical data; and L_spatio is the composite loss of Huber loss and cosine similarity for spatiotemporal data. γ4, γ5, and γ6 are the corresponding weighting coefficients, which can be dynamically adjusted according to the specific fault type and test objective: if the test focuses more on spatial anomalies (such as GPS drift), γ6 can be appropriately increased; if the test focuses on state transitions (such as communication interruptions), γ5 can be increased, etc.

[0036] Update the generator's weights θ according to the total loss function. g weight θ g After the update, the internal encoding-decoding logic of the generator network G undergoes a slight change, so that during the next forward propagation, for the same input noise Z and condition C, it outputs a "better" data sample S. This new data is more logically consistent within and between modes than the old data. For example, when generating the "GPS coordinate drift" anomaly, the azimuth and tilt data undergo geometrically consistent coordinated changes, and these spatial changes are logically self-consistent with the numerical power drop and the categorical communication interruption flag on the timeline.

[0037] A multimodal discriminant branch is introduced on top of the standard discriminator architecture: three parallel subnetworks D num D cat D spatio The discriminator independently evaluates the authenticity of the input data S across each modality through parallel dedicated subnetworks and performs weighted fusion decisions. The discriminator architecture and loss calculation logic are altered, as shown below: γ4+γ5+γ6=1, and γ4, γ5, and γ6 are consistent with the weight coefficients in the generator to ensure training consistency.

[0038] The discriminator's weights θd are updated by obtaining the multimodal fusion loss. After the weights θd are updated, the internal feature extraction and decision logic of the discriminator network D undergoes a slight change, so that for the same input data S, it will output a more "accurate" discrimination result: the new discrimination result comprehensively and meticulously considers the authenticity of the data in all modal dimensions. For example, the updated discriminator can keenly identify "numerical sequence patterns that seem reasonable but are seriously inconsistent with the labeled weather category 'rainstorm'" or "GPS drift trajectories that are spatially continuous but whose associated string current data do not show the expected shadow occlusion changes", thus greatly improving the ability to detect cross-modal inconsistency defects.

[0039] Furthermore, the initial recommended values ​​for the weighting coefficients are γ4=0.6, γ5=0.2, and γ6=0.2. If the test focuses on numerical anomalies (such as power surges), then increase γ4; if the test focuses on state transitions (such as communication interruptions), then increase γ5; if the test focuses on spatial anomalies (such as GPS drift), then increase γ6.

[0040] The improvement to the construction of GAN models can be made by selecting one or more of the following mechanisms from the data characteristics of the evaluation test samples for photovoltaic fault types as needed: physical constraint mechanism, temporal correlation preservation mechanism, and multimodal data fusion.

[0041] S2: Constructing a fusion framework for knowledge graphs and probabilistic graphical models, including building an ontology model in the photovoltaic field to form a knowledge graph for the photovoltaic field, and establishing the probabilistic graphical model based on the constraint relationships in the ontology model, including: S21: Construct an ontology model Q=(E, R, A) for the photovoltaic system, where E is the entity set, R is the relation set, and A is the attribute set. The constraints of the knowledge graph include the entity set, relation set, and attribute set. The entity set E includes various physical entities in the photovoltaic system, such as E={inverter, photovoltaic module, irradiance sensor, temperature sensor, GPS module, string, combiner box, ...,}; R is the relation set, describing the structure, function, and causal relationships between entities, including input-output relationships, control relationships, physical connection relationships, data dependency relationships, and influence relationships; A is the attribute set, used to define the observable or computable attributes of each entity, such as A={efficiency, temperature, voltage, current, power, coordinates, timestamp, ...,}. Through this ontology model, a structured photovoltaic domain knowledge graph is formed, achieving a standardized representation of the photovoltaic system's composition, operating logic, and anomaly propagation paths.

[0042] S22: Establishing the probabilistic graphical model based on the constraints in the knowledge graph includes: S221: Based on the relation set and attribute set in the ontology model, establish a Bayesian structure of the parameter-index causal network, wherein the nodes of the Bayesian structure represent system parameters X or business indicators I, and the edges represent conditional probability distributions P(I|X) derived from the relation set R.

[0043] The nodes in a Bayesian structure are divided into two categories: system parameter nodes X and business indicator nodes I. System parameter nodes X mainly include environmental parameters X. env With equipment operating parameter X device Business indicator node I is a high-level performance indicator defined from the perspective of photovoltaic power plant operation, maintenance and business decision-making. It is used to quantify the overall health status, performance and economic value of photovoltaic systems. It mainly includes fault detection accuracy, performance evaluation indicators, power generation prediction values ​​and economic benefit indicators.

[0044] The edges of the Bayesian structure represent the conditional probability distribution P(I|X) derived from the relation set R. This conditional probability distribution is learned by maximum likelihood estimation of historical normal operation data in the photovoltaic field, ensuring that it reflects the statistical dependence between parameters and indicators in the actual system.

[0045] Furthermore, S221 includes: S2211: Map the attributes in the attribute set A to the system parameter node X in the Bayesian structure.

[0046] S2212: Map the predefined business metrics to business metric nodes I in the Bayesian structure.

[0047] S2213: Based on the relation set R, determine the existence and direction of directed edges between the system parameter nodes and between the system parameter nodes and the business indicator nodes, and obtain the Bayesian structure.

[0048] S222: Jointly map the Bayesian network structure and the ontology model through shared variable nodes to form a unified factor graph structure G=(V F V B ), where V F V represents the factor node of an entity-relationship in a knowledge graph. B This represents the probability distribution nodes in a Bayesian network; the factor graph achieves deep integration of domain knowledge and statistical models, supporting bidirectional information transmission and dynamic knowledge updates.

[0049] S223: Based on the historical operating data of the photovoltaic system, the factor function parameters corresponding to the factor nodes in the factor graph model are calculated using the maximum likelihood estimation method to obtain the trained probability graph model.

[0050] S3: Based on the photovoltaic fault type, the characteristic parameters corresponding to the photovoltaic fault type, and the key features corresponding to the historical operation data of the photovoltaic monitoring system, the generator network and the discriminator network are iteratively trained until convergence; the generator network that has completed training outputs a preset number of evaluation test samples of predetermined photovoltaic fault types.

[0051] Furthermore, S3 includes: S31: Collect historical operating data of the photovoltaic system, preprocess the historical operating data and extract key features of the photovoltaic system, including time series features, parameter correlation features and spatial distribution features.

[0052] Historical operational data includes: environmental parameters (such as irradiance, ambient temperature, and module temperature), equipment operating parameters (such as DC voltage / current, inverter output power, inverter efficiency, and string current), geographical location information (such as GPS coordinates, azimuth, and tilt angle), and temporal characteristics (such as sampling timestamps, daily variation curves, and seasonal variation characteristics). Historical operational data is collected from the historical database of the target photovoltaic monitoring system over a continuous time period (e.g., 12 months) via standardized data interfaces (such as ODBC, OPCUA, and MQTT). The collected historical operational data requires preprocessing, including handling missing values ​​(using time series interpolation or regression based on correlation parameters), smoothing noise (using moving average or wavelet transform for denoising), and standardization / normalization to ensure the stability of subsequent feature extraction and model training. The cleaned historical data is then analyzed to extract the corresponding key features. The technical solution for preprocessing historical data of the photovoltaic system and extracting key features is existing technology and will not be described in detail here.

[0053] These key features include: time series features, parameter correlation features, and spatial distribution features.

[0054] Time series features can serve as a time dimension reference when the generator outputs data samples S, ensuring that the generated data samples S have reasonable variation patterns over time. Time series feature extraction includes: extracting frequency components such as daily and seasonal cycles of irradiance and power output through Fourier transform (FFT); analyzing the inertia of the time series through autocorrelation function (ACF); and calculating statistics (such as mean, variance, skewness, and kurtosis) within the sliding window to describe local dynamics. Time series features are the foundation for simulating abnormal time patterns (such as abrupt changes, drift, and oscillations).

[0055] Parameter correlation features enable the generator to maintain consistency in the physical laws between parameters when outputting data samples S, making the output data samples S more realistic. Parameter correlation feature extraction includes: calculating irradiance (Irr) and DC power (P).dc Correlation coefficients or fitted regression models under different weather conditions; analysis of ambient temperature (T) amb ) and component temperature (T) mod Models for the heat transfer hysteresis relationship between parameters, etc. Parameter correlation features ensure that the generated data to be evaluated has logical consistency, avoiding contradictory data that violates physical or operational logic.

[0056] Spatial distribution features enable the generator to consider the differences between devices in different locations when outputting data samples S, making the output data samples S more representative. Spatial distribution feature extraction includes: calculating the physical distance between different strings and inverters using Geographic Information System (GIS) coordinates; statistically analyzing the distribution of string current values ​​in different areas at the same time (e.g., Gaussian distribution, skewed distribution); and establishing a location-based irradiance difference model. These features are essential for generating anomalies with spatial characteristics (e.g., local shading, string faults, GPS drift).

[0057] S32: Collect photovoltaic fault data, classify the fault data according to fault type, and define the abnormal feature parameters corresponding to the fault type according to the parameters of the key features in the fault data. The fault types include sensor abnormality, equipment operation abnormality, communication abnormality, location information abnormality, data consistency abnormality, etc.

[0058] Continuously collect fault data in the photovoltaic field, classify the fault data according to fault type, and define the abnormal feature parameters corresponding to the fault type based on the parameters of key characteristics in the fault data. Create a fault database based on the fault type and the corresponding abnormal feature parameters. This fault database is continuously updated and includes more than 200 fault types and their corresponding abnormal feature parameters. When needed, the predetermined photovoltaic fault type and its corresponding abnormal feature parameters can be directly obtained from the fault database.

[0059] For each defined fault type, let the set of fault types be: Analyze the key features in the fault data corresponding to each fault type A. i The corresponding set of feature parameters is: The generator extracts quantifiable abnormal feature parameters from the feature parameter set that define the nature of the fault, forming a condition vector C corresponding to the fault type and inputting it into the generator. For example, for the fault type of irradiance mutation (sensor anomaly): the extracted abnormal feature parameters include the mutation amplitude (e.g., when the irradiance baseline is approximately 800 W / m² based on historical data and environmental conditions (e.g., weather, solar altitude angle) at the current moment, and the peak value after mutation is approximately 2400 W / m² ± 50 W / m², the mutation amplitude is 300% ± 5%), mutation duration (e.g., 15 min ± 2 min), rise time (e.g., 5 min), and occurrence time preference (e.g., UTC time 10:00-15:00); for the fault type of inverter efficiency... For equipment malfunctions (abnormal equipment operation), the extracted abnormal feature parameters include the direction of efficiency deviation (positive / negative), the degree of deviation (e.g., negative values ​​(e.g., -5% to -20%), or extremely high values ​​(e.g., 150% ± 5%)), and the frequency of abnormal fluctuations. For fault types such as GPS coordinate drift (abnormal location information), the extracted abnormal feature parameters include the drift speed (e.g., 0.05° / h to 0.5° / h), the drift direction (e.g., azimuth), and the maximum drift distance. For fault types such as communication interruption (abnormal communication), the extracted abnormal feature parameters include the interruption duration, data loss rate, and reconnection mode.

[0060] S33: The GAN model is iteratively trained until convergence based on the photovoltaic fault type, the abnormal feature parameters corresponding to the photovoltaic fault type, and the key features corresponding to the historical operating data.

[0061] First, the conditional vector C corresponding to the fault type and random noise Z are input into the generator, which outputs sample data S. Then, the key features corresponding to the sample data S and historical operating data are input into the discriminator, which outputs corresponding scalars. The discriminator aims to maximize its ability to distinguish between real and generated data. The discriminator's loss function L... D as follows: L_D(θ d )=E[log(D(S_real))]+E[log(1-D(S_fake))]=E_{S~p_data}[log(D(S;θ d ))]+E_{Z, C~p_z, p_c}[log(1-D(G(Z,C;θ g );θ dE_{S~p_data} represents the expectation of a sample S sampled from the real data distribution (p_data); E_{Z, C~p_z, p_c} represents the expectation of Z and C sampled jointly from the noisy prior distribution (p_z) and the conditional distribution (p_c); S_real=S represents the real historical data sample; S_fake=G(Z, C, θ) g () represents the data sample generated by the generator; By using the backpropagation algorithm, only the discriminator's parameter θ is updated. d This increases D(S_real) (approaching 1) and decreases D(S_fake) (approaching 0), forcing the discriminator D to better learn the characteristics of normal historical data S_real and accurately identify abnormal data S_fake generated by the generator that does not conform to the normal pattern.

[0062] Another batch of Z and C is sampled from the random noise and fault condition distribution. The generator's goal is to minimize the probability that it will be detected by the discriminator, i.e., to generate anomalous data that can "fool" the current discriminator. The generator's loss function L_G is as follows: L_G(θ g )=E_{Z, C~p_z, p_c}[log(1-D(G(Z, C; θ g );θ d E_{Z, C~p_z, p_c} represents the expectation of Z and C jointly sampled from the noise prior distribution (p_z) and the conditional distribution (p_c).

[0063] The backpropagation algorithm is used to update only the generator's parameter θ. g This allows the generated abnormal data S_fake to increase the value of D(S_fake) (approaching 1); it forces the generator G to adjust its parameters so that the generated fault data S_fake retains the specified fault characteristics (controlled by C) while being as close as possible in statistical characteristics to the distribution of the "real" data (i.e. normal data) as the discriminator considers it, thereby generating highly realistic abnormal data that conforms to physical laws.

[0064] Repeat the above steps, alternating between updating the discriminator D and the generator G. Through this adversarial training, the performance of the generator and discriminator continuously competes to improve: the discriminator continuously improves its ability to distinguish between real and fake data, while the generator continuously optimizes the realism of the generated anomalous data. The training process converges when the model reaches Nash equilibrium, i.e., when the discriminator can no longer distinguish between real and generated data with a higher probability than random guessing (D(S_real)≈0.5, D(S_fake)≈0.5). At this point, the generator G is able to output highly realistic anomalous data corresponding to the fault type to be evaluated based on the given fault condition vector C.

[0065] The GAN model in S1 is trained based on the fault type and corresponding abnormal feature parameters and key features. The generator training objective is to minimize the function L_G(θ). g This allows the generator's output data to deceive the discriminator to the greatest extent possible. The discriminator's training objective is to maximize the function L_D(θ). d This improves the discriminator's ability to distinguish between real and generated data. By alternately optimizing these two objective functions, the GAN model gradually reaches Nash Equilibrium, which satisfies the GAN model's convergence condition. This allows the generator to ultimately produce highly realistic evaluation test samples, which are highly realistic anomaly feature parameters corresponding to the fault type. Iterative training of the GAN model until convergence is a current technique and will not be elaborated upon here.

[0066] S34: Control the generator network that has completed training to output a predetermined number of evaluation test samples corresponding to a predetermined photovoltaic fault type.

[0067] According to the test requirements, the tester can select one or more predetermined photovoltaic fault types from the fault type library, and set the condition vector C according to the abnormal characteristic parameters corresponding to the fault type, or directly set the condition vector C, input random noise Z and condition vector C into the generator, and the generator G outputs a predetermined number of evaluation test samples S that meet all parameter settings of the predetermined photovoltaic fault type.

[0068] This embodiment can obtain a predetermined number of evaluation test samples S corresponding to a predetermined photovoltaic fault type, avoiding the problem of poor test results caused by insufficient data in existing photovoltaic monitoring system tests. It realizes the automated and industrialized production of test data, solving the pain point of traditional methods that rely entirely on scarce real fault data or inefficient manual construction. The generated data is highly diverse, highly authentic, and can be precisely controlled, greatly improving the coverage and depth of photovoltaic monitoring system testing.

[0069] S4: Input a preset number of the evaluation test samples into the probabilistic graph model, output the multi-dimensional impact probability corresponding to each evaluation test sample according to the mapping of the constraint relationship of the knowledge graph, and obtain the business impact degree corresponding to the evaluation test sample according to the multi-dimensional impact probability.

[0070] The evaluation test sample S is input into the factor graph model to identify anomalous feature parameters contained in the evaluation test sample (e.g., identifying a sudden change in the irradiance parameter Irr at timestamp t0). Subsequently, the system parameter nodes corresponding to the anomalous parameters are found in the factor graph model (e.g., finding the node X representing irradiance). irr ), and set the observation value of that node to the value in the outlier data (e.g., at time t0, force X). irr =2400W / m²), which is equivalent to "injecting" evidence at this specific node of the probabilistic graphical model.

[0071] The improved BeliefPropagation algorithm is launched, which involves computation of two core message types: Messages from parameter nodes (variable nodes) to function nodes (factor nodes): ; Messages from function nodes (factor nodes) to parameter nodes (variable nodes): .

[0072] The belief propagation algorithm starts from the node (Xirr) where the injected evidence is located. It propagates the uncertainty represented by this change along the conditional probability relationships defined by the edges in the factor graph to the connected functional nodes (F) and business indicator nodes (I). After the propagation calculation is completed, the belief propagation algorithm calculates the posterior marginal probability of the abnormal parameter on the state of each business indicator node in the factor graph model. The posterior marginal probability directly quantifies the possibility that the abnormality will cause the business indicator to deviate from the normal range, obtaining the multi-dimensional impact probability. Then, based on the multi-dimensional impact probability, the business impact corresponding to the test sample S is calculated and evaluated.

[0073] S5: Filter the evaluation test samples according to the business impact to obtain test data for testing the anomaly detection capability of the photovoltaic monitoring system.

[0074] The test samples S are selected based on their level of business impact. The test samples S with high business impact are selected to form test data for testing the anomaly detection capability of the photovoltaic monitoring system. This ensures that the test data has a significant impact on the key business indicators of the photovoltaic monitoring system, thereby effectively improving the relevance and efficiency of the photovoltaic monitoring system test. This ensures that the test data not only has high authenticity and diversity, but also has clear business orientation and fault coverage depth. It can systematically verify the anomaly response and fault tolerance handling capabilities of the photovoltaic monitoring system in multiple dimensions such as fault detection, performance evaluation, power generation prediction, and economic operation and maintenance.

[0075] Furthermore, S222 includes: S2221: Map each of the conditional probability distributions in the Bayesian structure to a factor node in the factor graph model, and connect the factor node to all variable nodes involved in the conditional probability distribution; For example, the nodes of a Bayesian structure are irradiance Irr and component temperature T. mod DC power P dc The corresponding edge is: Irr→T mod (Irradiance affects module temperature), Irr→P dc (Irradiance affects DC power), T mod →P dc (Component temperature also affects DC power), so the conditional probability distribution is P(Tmod|Irr) (the probability distribution of component temperature given irradiance), P(P dc |Irr,T mod P(Irr) (the probability distribution of DC power given irradiance and component temperature), and P(Irr) (the prior probability distribution of irradiance), represent P. dc Simultaneously affected by Irr and T mod The decision was made.

[0076] For P(T) mod |Irr) Create a factor node: f T (Representing temperature function), is P(P) dc |Irr,T mod Create a factor node: f P (Representing power function), create a factor node for P(Irr): f I (Represents prior irradiance).

[0077] Connect each newly created factor node to all variable nodes involved in the conditional probability distribution: for factor node f T (From P(T)) mod |Irr)), involving T mod (Node) and Irr(Parent), therefore, fT With variable node T mod Connect Irr and Irr with edges respectively.

[0078] S2222: Map the relations in the relation set R to the factor nodes in the factor graph model, and connect the factor nodes to the variable nodes corresponding to the entities participating in the relations; wherein, the set of variable nodes V B The factor graph model is constructed by consisting of the system parameter nodes and the business indicator nodes.

[0079] In this embodiment, knowledge graphs and probabilistic graphical models are combined. Knowledge graphs provide deterministic constraints, resolving the potential absurd reasoning that pure probabilistic models may exhibit in sparse data regions. Probabilistic graphical models offer the ability to quantify uncertainty and learn complex relationships from data, compensating for the often lack of numerical strength in knowledge graphs. This combination makes the evaluation model both reliable and flexible.

[0080] The fusion framework of the trained knowledge graph and the probabilistic graphical model becomes a "digital twin" evaluation system. When the fault data (evaluation test sample S) corresponding to the photovoltaic fault is injected, the reasoning algorithm can calculate how the fault data will affect the final business indicators (I), thereby realizing the quantitative mapping from "data anomaly" to "business impact" and obtaining the multi-dimensional impact probability corresponding to the fault data (evaluation test sample S).

[0081] Furthermore, in S4, an improved belief propagation algorithm is executed in the factor graph model to obtain the multi-dimensional probability and influence path of the evaluation test sample on the business indicator node; the improvement of the belief propagation algorithm includes: An attenuation factor is introduced into the message passing rules of the belief propagation algorithm for improvement. The improved message passing rules include: An attenuation factor α is introduced into the message passing rules from the system parameter node to the functional node to improve the message passing rules: ; An attenuation factor is introduced into the message passing rules from the functional node to the system parameter node to improve the message passing rules: ; The attenuation factor is defined as: α(d)=1 / (1+βd)β>0, where d is the shortest path hop count between the two nodes in the factor graph G, and β is an adjustable attenuation coefficient. The recommended range for β is [0.1, 5.0], and the initial recommended value is β=1.0.

[0082] By introducing an attenuation factor into the message passing rules of the belief propagation algorithm, the problem of biased impact probability estimation caused by the infinite propagation or excessive diffusion of messages in the large-scale factor graph of photovoltaic systems by traditional algorithms is effectively solved. This embodiment can accurately quantify the attenuation characteristics of abnormal impact with network topology distance, significantly improve the accuracy and reliability of multi-dimensional impact probability and impact path calculation, and provide a solid theoretical basis and calculation foundation for the objective assessment of subsequent business impact.

[0083] Furthermore, the improvement to the belief propagation algorithm also includes: introducing an anomaly feature weight function w(e) into the belief propagation algorithm: Based on the photovoltaic fault type e corresponding to the evaluation test sample S, the corresponding abnormal feature weight w(e) is obtained. For example, for sensor anomalies (e.g., when the irradiance suddenly changes to 300%), w(e) = 1.2; for equipment operation anomalies (e.g., inverter efficiency anomalies), w(e) = 1.5; for communication anomalies (e.g., communication interruption), w(e) = 0.8; for location information anomalies (e.g., GPS drift), w(e) = 1.0; and for data consistency anomalies, w(e) = 1.1, etc. When calculating the message passed from the functional node to the parameter node, the message is weighted according to the weight w(e) of the fault type e: .

[0084] By introducing an anomaly feature weight function w(e) closely related to photovoltaic fault types into the belief propagation algorithm, the precise quantification and dynamic modulation of the differentiated impact intensity of different types of anomalies are achieved. This embodiment effectively overcomes the limitations of traditional algorithms for various types of faults, significantly improves the distinguishability and business relevance of multi-dimensional impact probability calculation, and makes the final evaluation results more objectively reflect the true hazard level of specific faults, providing a more reliable basis for the accurate screening of test data and system vulnerability analysis.

[0085] Furthermore, the improvement to the belief propagation algorithm also includes: employing a hierarchical iterative strategy to execute the improved belief propagation algorithm to obtain the influence probability and influence path, including: Input the device parameters in the evaluation test sample S, perform belief propagation between nodes within a single photovoltaic device, and output device layer messages; The device layer message is input into the system layer, belief propagation is performed between cross-device nodes, and the system layer message is output. The system-level message is input, globally transmitted to the business indicator node, and the impact path and the impact probability are output.

[0086] By employing a hierarchical iterative strategy to execute an improved belief propagation algorithm, the complex global impact propagation calculation is decomposed into multi-level refined processing at the device layer, system layer, and global business layer. This effectively solves the problems of high computational complexity, slow convergence speed, and large path noise encountered when directly performing global propagation in large-scale photovoltaic monitoring systems. This hierarchical mechanism significantly improves the computational efficiency and accuracy of impact probability and impact path, and can clearly reveal the complete propagation chain of anomalies from local device parameters to system-level performance indicators, and finally to global business impact. This provides efficient and reliable algorithmic support for accurately assessing the scope of fault impact and screening test data.

[0087] Furthermore, based on the factor graph structure, an improved belief propagation algorithm is used to calculate the multi-dimensional probability and impact path of the evaluation test sample S on business indicator I. The improved belief propagation algorithm introduces an attenuation factor to control the attenuation of the impact with network distance on the basis of standard belief propagation. At the same time, an anomaly feature weight function w(e) is introduced to dynamically adjust the propagation weight according to different fault types e, so as to more accurately reflect the impact characteristics of different types of anomalies. The propagation process adopts a hierarchical iterative strategy, and message passing is executed sequentially at the device layer, system layer and business layer. Finally, the impact probability value of each business indicator node and its propagation path map are output, providing a quantitative basis for subsequent business impact calculation and test report generation.

[0088] In one implementation, the probability that anomalies will cause business metrics to deviate from the normal range is directly quantified based on the posterior marginal probability, resulting in multi-dimensional impact probabilities, including: Obtain the posterior marginal probabilities, denoted as: fault detection anomaly probability P_fd, performance evaluation anomaly probability P_pe, power prediction anomaly probability P_ef, and economic benefit anomaly probability P_eb; An evaluation matrix is ​​constructed based on the obstacle sensitivity function, performance deviation function, prediction error accumulation function, and economic loss estimation model, including: Construct a fault sensitivity function Y_f(S, A). If the probability P_FD of fault detection-related nodes is too high, call the fault detection impact function Y_f(S, A). Input the evaluation test sample S and the fault detection algorithm set A, and output the fault detection impact quantification value FDI, including: ; Indicates the fault detection time. This represents normal data, where wᵢ represents the weight coefficient of algorithm Aᵢ, satisfying... To evaluate the test sample S, algorithm A... i The time required to detect the fault This represents the time required for algorithm Ai to detect a fault under normal data S0. According to the set of key performance indicators for photovoltaic power plants (Including performance ratio, capacity utilization, inverter efficiency, etc.), define the performance deviation function.

[0089] ,in Let vᵢ represent the i-th performance index value calculated under data S, and let vᵢ represent the weight coefficient of this index, satisfying the following condition. ; Under the evaluation test sample S, the value of the i-th performance index, The value of the i-th performance metric under normal data S0; Construct a prediction error accumulation function E_f(S, F). If the probability of an anomaly in the electricity prediction P_ef exceeds a preset threshold, then call the prediction error accumulation function E_f(S, F) for calculation. The cumulative impact of the test sample S on the power generation prediction function F in the time interval [t1, t2] is evaluated quantitatively. F(t,S) is the predicted value of the power generation prediction model at time t under abnormal data S input; F(t;S0): the predicted value of the power generation prediction model at time t under normal data S0 input; [t1,t2]: the time interval for evaluation.

[0090] An economic loss estimation model L_e(X) is established. If the probability of abnormal economic benefit P_eb exceeds a preset threshold, the economic loss estimation model L_e(X) is called for calculation. The model L_e(X) integrates the economic consequences brought about by the performance evaluation impact (PEI, calculated by D_p), the power prediction impact (EFI, calculated by E_f), and the fault detection impact (FDI, calculated by S_f), and maps them to specific economic loss values.

[0091] EBI=L_e(X)=∑ i ᵏcᵢ·f(D_p,E_f)ᵢ+dᵢ·g(S_f)ᵢ, f(D_p,E_f)ᵢ: The loss mapping function of the i-th type caused by the combined performance deviation (D_p) and prediction error (E_f). For example, f1(D_p,E_f) can represent the loss of power generation revenue caused by the combined system performance degradation (PEI) and inaccurate power generation prediction (EFI); g(S_f)ᵢ: The loss mapping function of the i-th type caused by fault detection delay or misjudgment (S_f). For example, g1(S_f) can represent the additional maintenance costs, fines or power loss caused by the failure of faults to be detected in a timely and accurate manner; cᵢ, dᵢ: Economic coefficients for the corresponding loss types, respectively, and k is the total number of economic loss types considered.

[0092] An evaluation matrix is ​​constructed by using a fault sensitivity function, a performance deviation function, a prediction error accumulation function, and an economic loss estimation model. The posterior marginal probability is input into the evaluation matrix to calculate the anomaly probability of each dimension. If the value exceeds a preset threshold, the anomaly probability is used as a trigger signal. The original evaluation test sample S and normal baseline data S0 are used as inputs to calculate the specific impact level of that dimension. If P_fd exceeds the threshold, the fault detection impact function is called, taking S and S0 as input, and calculating and outputting the fault detection impact degree FDI. If P_pe exceeds the threshold, the performance evaluation impact function is called, taking S and S0 as input, and calculating and outputting the performance evaluation impact level PEI. If P_ef exceeds the threshold, the power prediction impact function is called, with S and S0 as input, to calculate and output the power prediction impact level EFI; If P_eb exceeds the threshold, the economic benefit impact function is called, taking S and S0 as inputs, and combining the calculation results of PEI and EFI to calculate and output the degree of economic benefit impact EBI.

[0093] The probability of output fault detection impact P(FDI), the probability of performance evaluation impact P(PEI), the probability of power prediction impact P(EFI), and the probability of economic benefit impact P(EBI).

[0094] Furthermore, the subsequent calculation of the business impact corresponding to the test sample based on multi-dimensional impact probabilities includes: Obtain the multi-dimensional impact probabilities corresponding to the evaluation test samples. The multi-dimensional impact probabilities include: fault detection impact probability P(FDI), performance evaluation impact probability P(PEI), power prediction impact probability P(EFI), and economic benefit impact probability P(EBI). For each probability of impact, a corresponding weight coefficient is assigned. Based on the weight coefficients and the probability of impact, the business impact of the evaluation test sample is obtained. , where t1, t2, t3, and t4 are weighting coefficients and satisfy t1+t2+t3+t4=1.

[0095] t1, t2, t3, and t4 are pre-set weighting coefficients based on the photovoltaic system's operation strategy and testing preferences. For example, if the current test focuses more on fault detection capabilities, t1 can be set higher. The probability value of the impact directly reflects the "hazard level" or "test value" of the abnormal parameters included in the evaluation test sample to the overall photovoltaic monitoring.

[0096] By constructing an evaluation matrix that integrates multi-dimensional quantitative functions and achieving precise mapping and triggering calculations from fault data to specific business impacts based on posterior marginal probabilities, a complete, objective, and configurable abnormal data business value assessment system is established through weighted fusion to obtain the comprehensive business impact. This embodiment effectively solves the problems of strong subjectivity, single dimensions, and difficulty in quantification in the assessment of abnormal data impacts in traditional methods. It can accurately quantify the differentiated impact degree and probability of specific anomalies on four key dimensions: fault detection, performance evaluation, power generation prediction, and economic benefits. It also supports dynamic adjustment of evaluation weights according to actual operation strategies and test objectives, thereby ensuring that the generated test data is not only realistic and diverse but also has clear business relevance and high test value, greatly improving the coverage depth and evaluation efficiency of photovoltaic monitoring system testing.

[0097] In one implementation of this embodiment, step S4 further includes: The evaluation test sample is input into the probabilistic graphical model, and the influence path corresponding to the evaluation test sample is output.

[0098] The influence path is mapped to domain entity relationships based on the constraints of the ontology model, and the influence path graph corresponding to the evaluation test sample is output based on the domain entity relationships.

[0099] Based on the impact path diagram, the multi-dimensional impact probability, and the business impact degree, output a test report for the evaluation test sample. The test report includes: fault type, impact path diagram, multi-dimensional impact probability, and business impact degree.

[0100] Specifically, for example, when the probabilistic graphical model output shows the impact path of the evaluation test sample as: irradiance sensor → string current → inverter DC input power → inverter efficiency → total system output power, based on the constraints defined in the pre-built photovoltaic domain ontology model Q=(E, R, A), such as inverter and string in entity set E, direct influence and membership in relation set R, and current and efficiency in attribute set A, the abstract impact path output by the above probabilistic model is mapped into a specific entity relation chain rich in domain semantics. The system automatically generates a visualized impact path diagram based on this relation chain. This diagram clearly shows the complete chain of anomalies starting from the source device (such as the irradiance sensor), through intermediate components and parameters, and ultimately affecting high-level business indicators (such as power generation). The system integrates the impact path diagram, calculates the multi-dimensional impact probabilities (i.e., P(FDI), P(PEI), P(EFI), P(EBI)), and finally quantifies the business impact, and automatically generates a structured test report. The core contents of the test report include: fault type, impact path diagram, multi-dimensional impact probabilities, and business impact.

[0101] This invention deeply integrates the output of a probabilistic graphical model with a domain knowledge graph and automatically generates a detailed test report containing a visualized impact path diagram and multi-dimensional quantitative indicators, achieving transparent, explainable, and accurate tracing of the impact from underlying data anomalies to top-level business operations.

[0102] Meanwhile, the test report serves as a feedback signal for the GAN model training process, providing a novel, objective, and quantifiable external verification mechanism for evaluating the convergence of generative adversarial networks. This approach effectively overcomes the limitations of traditional GAN ​​training, which relies solely on subjective judgment based on fluctuations in the generator and discriminator loss functions. By continuously analyzing the business impact, multi-dimensional impact probabilities, and the stability and rationality of impact paths in the test report, it can accurately determine from the business logic level whether the model has learned real, effective, and physically consistent abnormal patterns. This significantly improves the controllability of model training, the reliability of convergence results, and the quality and practical value of the final generated data. Example 2

[0103] Based on the photovoltaic monitoring system abnormal test data generation method provided in Embodiment 1 above, this embodiment generates test data corresponding to the specific fault type of "irradiance mutation" according to the photovoltaic monitoring system abnormal test data generation method.

[0104] A GAN model is constructed. The generator in the GAN model adopts a 5-layer LSTM-CNN hybrid architecture, with 128 neurons in each LSTM layer and 64 convolutional kernels in each CNN layer. It is used to learn and generate data with high-fidelity temporal characteristics. The discriminator adopts a 4-layer ResNet structure as a temporal analysis module to judge the rationality of the data from the perspective of the overall sequence.

[0105] Define the data sample vector S=[I rr T amb T mod P dc V dc I dc η inv [, ...]. For this embodiment, constraint functions are defined. A array For the area of ​​the photovoltaic array, η pv For the component's nominal efficiency, ε1 is set to 50W, and the penalty coefficient λ is 0.5.

[0106] A photovoltaic system ontology model Q=(E, R, A) is constructed, where the entity set E includes {irradiance sensor, photovoltaic string, inverter, ...}; the relation set R includes {direct influence, membership, input / output}; and the attribute set A includes {irradiance, module temperature, DC power, ...}. Based on the constraints in this ontology model, a probabilistic graphical model is established, specifically by mapping attributes to system parameter nodes, determining directed edges between nodes based on the relation set, and finally merging them into a factor graph model G=(V F V B Using historical operating data, the factor function parameters are calculated through the maximum likelihood estimation method to complete the training of the probabilistic graphical model.

[0107] Historical operational data from January to December (12 months in total) of a 500kW photovoltaic power station were collected using a standardized interface. The sampling granularity was 5 minutes, resulting in approximately 105,120 data entries. This included data on the irradiance time-series data. rr (t) Perform feature extraction, including time-domain analysis (calculating the intraday variation feature function f_d(t)=E[I rr The analysis, including the use of frequency domain analysis (identifying periodic patterns through Fourier transform F[Irr(t)]) and conditional probability modeling (constructing P(Irr|W,t), where W is the weather state vector), yielded an average peak irradiance of 842 W / m², a standard deviation σ = 76 W / m², and a short-term variation rate typically not exceeding 10% / 5 min.

[0108] Select the fault type "Irradiance mutation" from the preset fault library and the corresponding abnormal feature parameters of the fault type. Based on the feature parameters of the anomaly, construct the condition vector C: Amplitude: 300% ± 5%, Duration: 15 min ± 2 min, Rise Time: 5 min.

[0109] The GAN model was trained using 80% of the historical data as the training set and 20% as the validation set. The conditional vector C and key features from the historical data were used to iteratively train the GAN model constructed in S1. Training parameters were set as follows: learning rate η = 0.0001, batch size b = 64, and training epochs e = 500. The loss function was the sum of the Wasserstein distance and the physical constraint penalty term. After 500 training epochs, the discriminator accuracy reached 92.7%, the generator loss decreased to 0.068, and the model converged.

[0110] After training, the abnormal feature parameters corresponding to the specific fault type "irradiance mutation" are used as the condition vector C input to the generator to generate evaluation test samples containing irradiance mutation features. The evaluation test samples are input into the probabilistic graphical model to output the corresponding multi-dimensional impact probabilities and impact path diagram. The multi-dimensional impact probabilities include: Fault Detection Impact (FDI): 0.876 (high impact), Performance Evaluation Impact (PEI): 0.724 (medium-high impact), Power Prediction Impact (EFI): 0.913 (high impact), Economic Benefit Impact (EBI): estimated daily power generation loss of about 3.2%, EBI=0.032; Impact path diagram: the anomaly is from the irradiance sensor → string current → inverter DC power → total system output power → power generation revenue.

[0111] The business impact of the evaluation test sample is calculated based on the multi-dimensional impact probability, including: setting corresponding weight coefficients, t1=0.4, t2=0.2, t3=0.1, t4=0.4, and obtaining the business impact of the evaluation test sample based on the weight coefficients and the impact probability. P(CI(S))=t1·P(FDI)+t2·P(PEI)+t3·P(EFI)+t4·P(EBI)= (0.4×0.876)+(0.2×0.724)+(0.1×0.913)+(0.3×0.032)=0.5961.

[0112] The overall business impact score is 0.596, indicating that the fault type is "sudden change in irradiance" and the abnormal event with the characteristic parameters of "sudden change in irradiance of 300% lasting for 15 minutes and a rise time of 5 minutes" has an overall business impact of 59.6% on the photovoltaic monitoring system. This data shows that this abnormal event is a highly valuable test case that can effectively test the core data processing and fault tolerance capabilities of the photovoltaic monitoring system.

[0113] Based on the number of test data used to test the anomaly detection capability of the photovoltaic monitoring system, the generator outputs a corresponding number of evaluation test samples. The evaluation test samples are input into the probabilistic graphical model to obtain a corresponding multi-dimensional influence probability and a corresponding influence path diagram. The corresponding comprehensive business influence is calculated, and the evaluation test samples are selected based on the comprehensive business influence. The selected evaluation test samples are the test data used to test the anomaly detection capability of the photovoltaic monitoring system.

[0114] Based on the selected evaluation test samples, the corresponding impact path diagram, multi-dimensional impact probabilities, and business impact are output, forming a test report corresponding to the test data. The test report includes: Anomaly type: Irradiance mutation; Characteristic parameters: "Irradiance mutation of 300% lasting for 15 minutes, mutation rise time of 5 minutes"; Multi-dimensional impact probabilities include: FDI=0.876, PEI=0.724, EFI=0.913, EBI=0.032; Comprehensive business impact P(CI(S))=0.596; Main impact path: Irradiance sensor → string current → inverter DC power → total system output power → power generation revenue; Test value rating: High. Example 3

[0115] like Figure 3 As shown, this embodiment 3 provides a schematic diagram of the photovoltaic monitoring system abnormal test data generation device, based on the photovoltaic monitoring system abnormal test data generation method provided in Embodiment 1. Figure 3 As shown, the device includes: an adversarial generative network module 100, an evaluation framework construction module 200, a test data generation module 300, a business impact assessment module 400, and a test data acquisition module 500.

[0116] The Generative Adversarial Network (GAN) module 100 is used to construct a GAN model, including constructing a generator network and a discriminator network. The generator network includes a physical constraint module, a multimodal encoder-decoder structure, and a temporal generation module. The discriminator network includes a physical rule filtering layer, a multimodal discriminator branch, and a temporal analysis module. The evaluation framework construction module 200 is used to construct a fusion framework of knowledge graph and probabilistic graphical model, including constructing an ontology model in the photovoltaic field to form a knowledge graph in the photovoltaic field, and establishing the probabilistic graphical model according to the constraint relationship in the ontology model; The test data generation module 300 is used to iteratively train the generator network and the discriminator network until convergence based on the photovoltaic fault type, the characteristic parameters corresponding to the photovoltaic fault type, and the key features corresponding to the historical operation data of the photovoltaic monitoring system; the generator network that has completed training outputs a preset number of evaluation test samples of predetermined photovoltaic fault types. The business impact assessment module 400 is used to input a preset number of assessment test samples into the probability graph model, output the multi-dimensional impact probability corresponding to each assessment test sample according to the mapping of the constraint relationship of the knowledge graph, and obtain the business impact degree corresponding to the assessment test sample according to the multi-dimensional impact probability. The test data acquisition module 500 is used to filter the evaluation test samples according to the business impact to obtain test data for testing the anomaly detection capability of the photovoltaic monitoring system.

[0117] In one alternative approach, the evaluation framework building module 200 includes: an ontology model building unit and a probabilistic graphical model building unit.

[0118] The ontology model building unit is used to construct the ontology model Q=(E, R, A) in the photovoltaic system, where E is the entity set, R is the relation set, and A is the attribute set. The constraints of the knowledge graph include the entity set, relation set, and attribute set. The entity set E includes various physical entities in the photovoltaic system, such as E={inverter, photovoltaic module, irradiance sensor, temperature sensor, GPS module, string, combiner box, ...,}; R is the relation set, describing the structure, function, and causal relationships between entities, including input-output relationships, control relationships, physical connection relationships, data dependency relationships, influence relationships, etc.; A is the attribute set, used to define the observable or computable attributes of each entity, such as A={efficiency, temperature, voltage, current, power, coordinates, timestamp, ...,}. Through this ontology model, a structured photovoltaic domain knowledge graph is formed, realizing a standardized representation of the photovoltaic system composition, operating logic, and anomaly propagation paths.

[0119] The probabilistic graphical model building unit is used to build the probabilistic graphical model based on the constraints in the knowledge graph, including: Based on the relation set and attribute set in the ontology model, a Bayesian structure of a parameter-index causal network is established, wherein the nodes of the Bayesian structure represent system parameters X or business indicators I, and the edges represent conditional probability distributions P(I|X) derived from the relation set R. This includes: mapping attributes in the attribute set A to system parameter nodes X in the Bayesian structure, and mapping predefined business indicators to business indicator nodes I in the Bayesian structure; determining the existence and direction of directed edges between system parameter nodes and between system parameter nodes and business indicator nodes based on the relation set R, thereby obtaining the Bayesian structure; and jointly mapping the Bayesian structure and the ontology model through shared variable nodes to form a factor graph model G=(V F V B V F V represents the factor node of the relation-attribute relationship in the knowledge graph. B The factor graph model is the probability distribution node in the Bayesian network, comprising: mapping each conditional probability distribution in the Bayesian structure to a factor node in the factor graph model, and connecting the factor node to all variable nodes involved in the conditional probability distribution; mapping relations in the relation set R to factor nodes in the factor graph model, and connecting the factor nodes to the variable nodes corresponding to the entities participating in the relations; wherein, the set of variable nodes V BThe factor graph model is constructed by the system parameter nodes and the business indicator nodes. Based on the historical operating data of the photovoltaic system, the factor function parameters corresponding to the factor nodes in the factor graph model are calculated by the maximum likelihood estimation method to obtain the trained probability graph model.

[0120] In one alternative embodiment, the test data generation module 300 includes: Historical operating data of the photovoltaic system is collected, preprocessed, and key features of the photovoltaic system are extracted. These key features include time series features, parameter correlation features, and spatial distribution features. Fault data of the photovoltaic system is collected and classified according to fault type. Anomaly feature parameters corresponding to each fault type are defined based on the parameters of the key features in the fault data. The fault types include sensor anomalies, equipment operation anomalies, communication anomalies, location information anomalies, and data consistency anomalies. The GAN model is iteratively trained until convergence based on the photovoltaic fault type, the anomaly feature parameters corresponding to the photovoltaic fault type, and the key features corresponding to the historical operating data. The trained generator network is controlled to output a predetermined number of evaluation test samples corresponding to a predetermined photovoltaic fault type.

[0121] In one alternative approach, the business impact assessment module 400 includes: executing an improved belief propagation algorithm in the factor graph model to obtain the multi-dimensional impact probability and impact path of the assessment test sample on the business indicator node; the improvement of the belief propagation algorithm includes: An attenuation factor is introduced into the message passing rules of the belief propagation algorithm for improvement. The improved message passing rules include: An attenuation factor α is introduced into the message passing rules from the system parameter node to the functional node to improve the message passing rules: ; An attenuation factor is introduced into the message passing rules from the functional node to the system parameter node to improve the message passing rules: ; The attenuation factor is defined as: d is the number of hops in the shortest path between the two nodes in the factor graph G. This is an adjustable attenuation coefficient.

[0122] The improvement to the belief propagation algorithm further includes: introducing an anomaly feature weight function w(e) into the belief propagation algorithm: Based on the photovoltaic fault type e corresponding to the evaluation test sample, obtain the corresponding abnormal feature weight w(e); When calculating the message passed from the functional node to the parameter node, the message is weighted according to the weight w(e) of the fault type e: .

[0123] The improvement to the belief propagation algorithm further includes: employing a hierarchical iterative strategy to execute the improved belief propagation algorithm to obtain the influence probability and influence path, including: Input the device parameters in the evaluation test sample, perform belief propagation between nodes within a single photovoltaic device, and output device layer messages; The device layer message is input into the system layer, belief propagation is performed between cross-device nodes, and the system layer message is output. The system-level message is input, globally transmitted to the business indicator node, and the impact path and the impact probability are output.

[0124] In one implementation, the business impact assessment module 400 further includes: directly quantifying the possibility that anomalies cause business indicators to deviate from the normal range based on posterior marginal probabilities, obtaining multi-dimensional impact probabilities, including: obtaining posterior marginal probabilities, denoted as: fault detection anomaly probability P_fd, performance assessment anomaly probability P_pe, power prediction anomaly probability P_ef, and economic benefit anomaly probability P_eb; constructing an assessment matrix based on the fault sensitivity function, performance deviation function, prediction error accumulation function, and economic loss estimation model; inputting the posterior marginal probabilities into the assessment matrix, calculating the anomaly probability of each dimension, and if its value exceeds a preset threshold, using the anomaly probability as a trigger signal, and reverting the original assessment measurement... Using the test sample S and normal baseline data S0 as input, the specific impact value of this dimension is calculated: If P_fd exceeds the threshold, the fault detection impact function is called, inputting S and S0, and the fault detection impact degree FDI is calculated and output; if P_pe exceeds the threshold, the performance evaluation impact function is called, inputting S and S0, and the performance evaluation impact degree PEI is calculated and output; if P_ef exceeds the threshold, the power prediction impact function is called, inputting S and S0, and the power prediction impact degree EFI is calculated and output; if P_eb exceeds the threshold, the economic benefit impact function is called, inputting S and S0, and the economic benefit impact degree EBI is calculated and output, taking into account the calculation results of PEI and EFI. The outputs are the fault detection impact probability P(FDI), the performance evaluation impact probability P(PEI), the power prediction impact probability P(EFI), and the economic benefit impact probability P(EBI).

[0125] The process then calculates the business impact of the evaluation test sample based on the multi-dimensional impact probabilities, including: obtaining the multi-dimensional impact probabilities corresponding to the evaluation test sample, wherein the multi-dimensional impact probabilities include: fault detection impact probability P(FDI), performance evaluation impact probability P(PEI), power prediction impact probability P(EFI), and economic benefit impact probability P(EBI); setting corresponding weight coefficients for each impact probability, and obtaining the business impact of the evaluation test sample based on the weight coefficients and the impact probabilities. , where t1, t2, t3, and t4 are weighting coefficients and satisfy t1+t2+t3+t4=1.

[0126] In one implementation, the business impact assessment module 400 further includes: inputting the assessment test sample into the probabilistic graphical model and outputting the impact path corresponding to the assessment test sample; mapping the impact path to domain entity relations according to the constraint relations of the ontology model, and outputting the impact path graph corresponding to the assessment test sample according to the domain entity relations; and outputting a test report of the assessment test sample according to the impact path graph, the multi-dimensional impact probability, and the business impact degree, wherein the test report includes: fault type, impact path graph, multi-dimensional impact probability, and business impact degree.

[0127] The apparatus provided in this invention constructs a Generative Adversarial Network (GAN) model that integrates physical constraints, a multimodal encoder-decoder structure, and a temporal generation module through a generative adversarial network module. Combined with a deep fusion framework of knowledge graphs and probabilistic graphical models, it can quickly and automatically generate test data covering more than several photovoltaic (PV) fault types. This test data is input into a PV monitoring system to test the system's data processing capabilities. This solves the problems of incomplete coverage, low authenticity, and poor consistency caused by traditional test data relying on limited real-world anomaly scenarios or manual construction. By introducing physical constraint mechanisms, temporal correlation preservation strategies, and multimodal data fusion to improve the standard GAN model, the apparatus ensures that the generated evaluation test samples conform to the operating rules of the PV system and possess temporal continuity and cross-modal consistency. Through joint reasoning using knowledge graphs and probabilistic graphical models, it achieves quantitative assessment and path tracing of the business impact of the evaluation test samples, significantly improving the relevance, practicality, and evaluation efficiency of the test data. This invention can be widely applied to the testing and verification of data cleaning, fault tolerance processing, and fault detection functions in PV monitoring systems, possessing significant engineering application value and promising prospects for wider application. Example 4

[0128] Figure 4The diagram shows a structural schematic of a computer device provided in Embodiment 4 of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computer device.

[0129] like Figure 3 As shown, the computer device may include: a processor 402, a communications interface 404, a memory 406, and a communications bus 408.

[0130] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408. Communication interface 404 is used to communicate with other network elements, such as clients or other servers. Processor 402 executes program 410, specifically performing the relevant steps described above in any of the photovoltaic monitoring system abnormal test data generation methods provided in Embodiment 1.

[0131] Specifically, program 410 may include program code, which includes computer-executable instructions.

[0132] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The device may include one or more processors of the same type, such as one or more CPUs; or it may include processors of different types, such as one or more CPUs and one or more ASICs.

[0133] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0134] Specifically, program 410 can be called by processor 402 to cause the computer device to perform the following operations: This invention provides a computer-readable storage medium storing at least one executable instruction that, when executed on a computer device, causes the computer device to perform the photovoltaic monitoring system abnormal test data generation method in any of the above method embodiments.

[0135] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of the present invention are not directed to any particular programming language. It should be understood that the content of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0136] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0137] Similarly, it should be understood that, in order to streamline the invention and aid in understanding one or more of the various aspects of the invention, features of the embodiments of the invention are sometimes grouped together in a single embodiment, figure, or description thereof in the above description of exemplary embodiments of the invention. However, this disclosure should not be construed as reflecting an intention that the claimed invention requires more features than are expressly recited in each claim.

[0138] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0139] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

Claims

1. A method for generating abnormal test data of a photovoltaic monitoring system, characterized in that, The method includes: Constructing a GAN model includes building a generator network and a discriminator network. The generator network includes a physical constraint module, a multimodal encoder-decoder structure, and a temporal generation module. The discriminator network includes a physical rule filtering layer, a multimodal discrimination branch, and a temporal analysis module. A framework for integrating knowledge graphs and probabilistic graphical models is constructed, including constructing an ontology model in the photovoltaic field to form a knowledge graph in the photovoltaic field, and establishing the probabilistic graphical model based on the constraint relationships in the knowledge graph; The generator network and the discriminator network are iteratively trained until convergence based on the photovoltaic fault type, the characteristic parameters corresponding to the photovoltaic fault type, and the key features corresponding to the historical operation data of the photovoltaic monitoring system; the generator network that has completed training outputs a preset number of evaluation test samples of predetermined photovoltaic fault types. The preset number of evaluation test samples are input into the probabilistic graph model. The multi-dimensional impact probability corresponding to each evaluation test sample is output according to the mapping of the constraint relationship of the knowledge graph. The business impact degree corresponding to the evaluation test sample is obtained according to the multi-dimensional impact probability. The evaluation test samples are selected based on the business impact to obtain test data for testing the anomaly detection capability of the photovoltaic monitoring system.

2. The method according to claim 1, characterized in that, The step of inputting a preset number of evaluation test samples into the probabilistic graphical model, outputting the multi-dimensional impact probability corresponding to each evaluation test sample according to the mapping of the constraint relationship of the knowledge graph, and obtaining the business impact degree corresponding to the evaluation test sample according to the multi-dimensional impact probability, further includes: The evaluation test sample is input into the probabilistic graphical model, and the influence path corresponding to the evaluation test sample is output. Based on the constraints of the ontology model, the influence path is mapped to domain entity relationships, and the influence path graph corresponding to the evaluation test sample is output based on the domain entity relationships. Based on the impact path diagram, the multi-dimensional impact probability, and the business impact degree, output a test report for the evaluation test sample. The test report includes: fault type, impact path diagram, multi-dimensional impact probability, and business impact degree.

3. The method according to claim 2, characterized in that, The framework for integrating knowledge graphs and probabilistic graphical models includes constructing an ontology model in the photovoltaic field to form a knowledge graph in the photovoltaic field, and establishing the probabilistic graphical model based on the constraint relationships in the knowledge graph, including: Construct an ontology model Q=(E, R, A) in the photovoltaic system, where E is the entity set, R is the relation set, and A is the attribute set. The constraint relations of the knowledge graph include the entity set, relation set, and attribute set. Building the probabilistic graphical model based on the constraints in the knowledge graph includes: Based on the relation set and attribute set in the ontology model, a Bayesian structure of a parameter-index causal network is established, wherein the nodes of the Bayesian structure represent system parameters X or business indicators I, and the edges represent conditional probability distributions P(I|X) derived from the relation set R. The Bayesian structure and the ontology model are jointly mapped through shared variable nodes to form a factor graph model G=(V F V B V F V represents the relation-attribute factor node in the knowledge graph. B For the probability distribution nodes in the Bayesian network, the factor graphical model is the probability graphical model; Based on the historical operating data of the photovoltaic system, the factor function parameters corresponding to the factor nodes in the factor graph model are calculated using the maximum likelihood estimation method, thus obtaining the trained probabilistic graphical model.

4. The method according to claim 3, characterized in that, The step of establishing a Bayesian structure for a parameter-index causal network based on the relation set and attribute set in the ontology model, wherein the nodes of the Bayesian structure represent system parameters X or business indicators I, and the edges represent conditional probability distributions P(I|X) derived from the relation set R, includes: Map the attributes in the attribute set A to the system parameter nodes in the Bayesian structure; Map the predefined business metrics to business metric nodes in the Bayesian structure; Based on the relation set R, determine the existence and direction of directed edges between the system parameter nodes and between the system parameter nodes and the business indicator nodes, and obtain the Bayesian structure.

5. The method according to claim 3, characterized in that, The Bayesian structure and the ontology model are jointly mapped through shared variable nodes to form a factor graph model G=(V F V B V F V represents the factor node of the relation-attribute relationship in the knowledge graph. B For the probability distribution nodes in a Bayesian network, the factor graphical model is the probabilistic graphical model, including: Each conditional probability distribution in the Bayesian structure is mapped to a factor node in the factor graph model, and the factor node is connected to all variable nodes involved in the conditional probability distribution. The relations in the relation set R are mapped to factor nodes in the factor graph model, and the factor nodes are connected to the variable nodes corresponding to the entities participating in the relations; wherein, the set of variable nodes V B The factor graph model is constructed by consisting of the system parameter nodes and the business indicator nodes.

6. The method for generating abnormal test data of a photovoltaic monitoring system according to claim 5, characterized in that, The process of inputting a preset number of evaluation test samples into the probabilistic graphical model, outputting the multi-dimensional impact probability corresponding to each evaluation test sample based on the mapping of the constraint relationships of the knowledge graph, and obtaining the business impact degree corresponding to the evaluation test sample based on the multi-dimensional impact probability includes: An improved belief propagation algorithm is executed in the factor graph model to obtain the multi-dimensional probability and influence path of the evaluation test sample on the business indicator node; The improvements to the belief propagation algorithm include: An attenuation factor is introduced into the message passing rules of the belief propagation algorithm for improvement. The improved message passing rules include: An attenuation factor α is introduced into the message passing rules from the system parameter node to the functional node to improve the message passing rules: ； An attenuation factor is introduced into the message passing rules from the functional node to the system parameter node to improve the message passing rules: ； The attenuation factor is defined as: β>0, d is the shortest path hop count between the two nodes in the factor graph G, and β is an adjustable attenuation coefficient.

7. The photovoltaic monitoring system abnormal test data generation system according to claim 6, characterized in that, The improvement to the belief propagation algorithm further includes: introducing an anomaly feature weight function w(e) into the belief propagation algorithm: Based on the photovoltaic fault type e corresponding to the evaluation test sample, obtain the corresponding abnormal feature weight w(e); When calculating the message passed from the functional node to the parameter node, the message is weighted according to the weight w(e) of the fault type e: 。 8. The method for generating abnormal test data of a photovoltaic monitoring system according to claim 6 or 7, characterized in that, The improvement to the belief propagation algorithm further includes: employing a hierarchical iterative strategy to execute the improved belief propagation algorithm to obtain the influence probability and influence path, including: Input the device parameters in the evaluation test sample, perform belief propagation between nodes within a single photovoltaic device, and output device layer messages; The device layer message is input into the system layer, belief propagation is performed between cross-device nodes, and the system layer message is output. The system-level message is input, globally transmitted to the business indicator node, and the impact path and the impact probability are output.

9. A device for generating abnormal test data of a photovoltaic monitoring system, characterized in that, The device includes: The Generative Adversarial Network (GAN) module is used to construct GAN models, including building a generator network and a discriminator network. The generator network includes a physical constraint module, a multimodal encoder-decoder structure, and a temporal generation module. The discriminator network includes a physical rule filtering layer, a multimodal discrimination branch, and a temporal analysis module. The evaluation framework construction module is used to construct a fusion framework of knowledge graph and probabilistic graphical model, including constructing an ontology model in the photovoltaic field to form a knowledge graph in the photovoltaic field, and establishing the probabilistic graphical model based on the constraint relationships in the knowledge graph; The test data generation module is used to iteratively train the generator network and the discriminator network until convergence based on the photovoltaic fault type, the characteristic parameters corresponding to the photovoltaic fault type, and the key features corresponding to the historical operation data of the photovoltaic monitoring system; the generator network that has completed training outputs a preset number of evaluation test samples of a predetermined photovoltaic fault type. The business impact assessment module is used to input a preset number of assessment test samples into the probabilistic graph model, output the multi-dimensional impact probability corresponding to each assessment test sample according to the mapping of the constraint relationship of the knowledge graph, and obtain the business impact degree corresponding to the assessment test sample according to the multi-dimensional impact probability. The test data acquisition module is used to filter the evaluation test samples according to the business impact to obtain test data for testing the anomaly detection capability of the photovoltaic monitoring system.

10. A computer device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation of the photovoltaic monitoring system abnormal test data generation method as described in any one of claims 1-8.

Images