Photovoltaic power station fault diagnosis method based on digital twinning and transfer learning

Abstract

The invention discloses a photovoltaic power station fault diagnosis method based on digital twinning and transfer learning, and belongs to the technical field of photovoltaic power generation fault detection. According to the method, a photovoltaic power station digital twin model based on a physical mechanism is constructed, operation behaviors in normal and multiple fault states are simulated, and a simulation data set with an accurate label is generated; in combination with the acquired real operation data, a fault diagnosis model is trained by adopting a transfer learning framework containing a domain adaptation loss item, so that the distribution difference between simulation data and real data is effectively reduced; a physical consistency loss item is innovatively introduced, and a physical rule is used as a constraint embedded model, so that the reliability and interpretability of a diagnosis result are improved; and finally, performing accurate fault diagnosis on the real-time operation data by using the trained model. According to the method, the model training problem caused by scarcity of real fault data is solved, the diagnosis precision and generalization ability are remarkably improved, and the method is suitable for intelligent operation and maintenance of the photovoltaic power station.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation fault detection technology, and in particular to a photovoltaic power plant fault diagnosis method based on digital twins and transfer learning. Background Technology

[0002] As a clean and renewable energy source, photovoltaic (PV) power generation has seen continuous growth in installed capacity. Ensuring the safe and efficient operation of PV power plants is crucial. However, PV power plants are exposed to complex environments for extended periods, making them susceptible to various faults such as hot spots, wiring failures, and potential induced degradation, leading to decreased power generation efficiency and even safety incidents.

[0003] Currently, mainstream fault diagnosis methods include threshold alarms and machine learning models based on historical data. Threshold alarms are simple but have a high false alarm rate and cannot identify early or complex faults. While machine learning models can achieve intelligent diagnosis, their performance heavily relies on a large amount of high-quality, labeled historical fault data. In actual power plants, severe fault samples are scarce, and labeling costs are high, resulting in insufficient model generalization ability. Furthermore, purely data-driven models lack physical mechanism support, and their reliability decreases when operating conditions change.

[0004] Digital twin technology can construct virtual models of physical entities, making it possible to generate fault data. Transfer learning can transfer knowledge learned from one domain (source domain) to another related but different domain (target domain). However, how to effectively combine the two, especially how to solve the problem of data distribution differences (i.e., domain differences) between simulated data generated by digital twins and real data, and ensure that the transfer process conforms to physical laws, is a key challenge for achieving high-precision fault diagnosis. Currently, there is no publicly available effective solution. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a photovoltaic power station fault diagnosis method based on digital twins and transfer learning, so as to solve the problem of difficulty in training the diagnostic model and poor generalization ability due to the scarcity of real fault data, and to improve the accuracy and reliability of diagnosis by introducing physical constraints.

[0006] In a first aspect, the present invention provides a method for fault diagnosis of photovoltaic power plants based on digital twins and transfer learning, comprising the following steps: S1. Construct a digital twin model of the target photovoltaic power station. The digital twin model is based on the physical mechanism of the photovoltaic system and is used to simulate the operation behavior of the photovoltaic power station under normal conditions and at least one fault condition. S2. Run the digital twin model to generate a simulated operation dataset containing fault labels; S3. Obtain the actual operation dataset of the target photovoltaic power station; S4. Based on the transfer learning framework, a fault diagnosis model is trained using the simulated running dataset as the source domain data and the real running dataset as the target domain data; wherein, the loss function of the transfer learning framework includes a domain adaptation loss term for reducing the difference in feature distribution between the source domain data and the target domain data. S5. Use the trained fault diagnosis model to diagnose faults in the real-time operating data of the target photovoltaic power station.

[0007] Preferably, in step S1, the digital twin model includes a photovoltaic module model, an inverter model, and an environmental model; the photovoltaic module model adopts a dual-diode model, and its IV characteristic relationship is defined by a physical equation including photocurrent, diode saturation current, series resistance, parallel resistance, and environmental parameters; the simulation of the fault state is achieved by dynamically modifying at least one parameter in the physical equation.

[0008] Preferably, the dynamic modification includes: to simulate hot spot faults, the parallel resistance is modified to be a function of time and spatial location; to simulate potential induced decay faults, a leakage current term related to ground voltage and ambient humidity is introduced.

[0009] Preferably, in step S4, the transfer learning framework adopts a deep neural network structure that includes a feature extractor, a fault classifier, and a domain discriminator; the domain adaptation loss term is calculated based on the discrimination result of the domain discriminator on the source domain features and target domain features output by the feature extractor.

[0010] Preferably, the loss function further includes a physical consistency loss term, which is used to constrain the difference between the physical quantities that the features extracted by the feature extractor can decode through a preset physical computation graph and the theoretical physical quantities calculated by the real-time environmental parameters through physical formulas to minimize the difference.

[0011] Preferably, the loss function is expressed as: ; in, Classify the task loss for fault classification. For the domain adaptation loss term, For the physical consistency loss term, and To balance hyperparameters.

[0012] Preferably, after step S3 and before step S4, the method further includes a step of data preprocessing and feature engineering on the simulated running dataset and the real running dataset. The feature engineering includes extracting time-domain features, frequency-domain features, and operating condition-aware features, wherein the operating condition-aware features include the deviation rate between the actual power and the ideal power calculated by the digital twin model.

[0013] Secondly, the present invention provides a system for implementing the above method, including a digital twin module, a data acquisition and processing module, a model training module, and a fault diagnosis module.

[0014] Thirdly, the present invention provides a computing device and a computer-readable storage medium for performing the above-described method.

[0015] Therefore, the photovoltaic power plant fault diagnosis method based on digital twins and transfer learning using the above structure, as described in this invention, has the following beneficial effects: (1) This invention uses a mechanism-driven high-fidelity digital twin model and a dynamic fault injection mechanism to generate massive amounts of simulated fault data with precise labels, thus completely overcoming the bottleneck of difficulty in obtaining real fault samples.

[0016] (2) By introducing a transfer learning framework with a domain adaptation loss term, this invention forces the model to learn domain-invariant features, effectively bridging the distribution differences between simulated data and real data, and enabling the model to be robustly applied to real power plant environments.

[0017] (3) The present invention innovatively adds a physical consistency loss term to the loss function, embedding physical laws as constraints into the model, avoiding the misjudgment that may occur in a pure data-driven black box model that violates physical common sense, and enhancing the interpretability of the results.

[0018] (4) The present invention uses the "diagnosis-feedback-update" mechanism to continuously calibrate the digital twin model and fine-tune the diagnostic model using online data, so that the system can adapt to the aging and changes of the power station, ensuring the accuracy and reliability of long-term application.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a photovoltaic power station fault diagnosis method based on digital twins and transfer learning according to the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0023] Example like Figure 1 As shown, this invention provides a photovoltaic power plant fault diagnosis method based on digital twins and transfer learning, including the following steps: S1. Construct a high-fidelity digital twin model of a photovoltaic power station. S11. Photovoltaic module model: A five-parameter dual-diode model is adopted, and its IV characteristics satisfy the following formula: ; I ph Photocurrent, and irradiance G Battery temperature T c Satisfying , ( Photocurrent under standard test conditions K I The temperature coefficient of current. T STC For standard test temperature, G STC (Standard test irradiance); and This represents the saturation current of the two diodes; and This is the diode ideality factor; Thermoelectric voltage ( k Boltzmann's constant, q (elementary charge); and These are series resistors and parallel resistors, respectively. Model parameters are extracted from manufacturer data sheets and particle swarm optimization algorithms to ensure they match the characteristics of real components.

[0024] Inverter Model: Integrating Efficiency Curve Model and Maximum Power Point Tracking (MPPT) Dynamic Model MPPT employs an improved perturbation observation method, and the voltage reference value update formula is as follows: Where, Δ V For the perturbation step size, For symbolic functions, Let be the power and voltage at time k, respectively; Inverter conversion efficiency For load rate A function that satisfies: ( , , (These are the parameters fitted experimentally).

[0025] S12, System-level Integration and Environment Coupling: According to the actual layout of the physical power station (number of series connections) Parallel number Extend the component model by introducing parameter inconsistencies between strings (such as differences in component attenuation). Integrated meteorological model, input local historical / real-time meteorological data (irradiance) G Ambient temperature Wind speed v ), calculate battery temperature using a thermal model : ; in: The rated battery operating temperature. The ambient temperature corresponding to the rated operating temperature. Irradiance corresponding to the rated operating temperature; For maximum power point efficiency, It is the product of the component's transmittance and absorptivity; This is the temperature cooling term caused by wind speed v.

[0026] S13, Dynamic random fault injection mechanism: Simulate the physical evolution of a fault to generate realistic fault data: Hot spot fault: introduces time-varying degradation to the parallel resistance Rsh of random cells in the module, as shown in the formula: ; in, This is a normal parallel resistor. The severity factor is 0~1. It is a two-dimensional Gaussian kernel spatial distribution function (simulating the local effect of hot spots). The fault time constant (simulating incremental evolution); PID fault: Based on the component-to-ground voltage Vpid and ambient humidity H, the leakage current is simulated, and the formula is: ; in, This is normal leakage current. and For PID sensitivity parameters; Wiring fault: Introducing time-varying contact resistance to random connection points ; ( This is normal contact resistance. This refers to the resistance fluctuation range. For fluctuation frequency, (The initial phase) simulates the intermittent electric arc caused by loosening.

[0027] By combining fault type, location, start time, and severity, a synthetic dataset with fault labels is generated. . ( It is a sequence of current, voltage, power, and temperature. (Includes a label for fault type and severity).

[0028] S2, Big Data Preprocessing and Multimodal Feature Engineering S21. Multi-source data cleaning and alignment: Real data cleaning: For the current, voltage, and power data collected by the SCADA system of the physical power plant, noise is removed by a combination algorithm of "Isolation Forest (identifying outliers) + Change Point Detection (locating data mutations)"; for missing data, ARIMA time series model interpolation is used (to replace simple mean filling). Time alignment: To address clock drift from different sensors, a dynamic time warping algorithm is employed to calibrate the time axis of multi-source data (such as component temperature and inverter power) within the same time window.

[0029] S22. Extract features from the cleaned time series data using a sliding window (window size can be set to 5-15 minutes, step size to 1-3 minutes): Time-domain characteristics: mean, variance, skewness, kurtosis, waveform factor (RMS / mean), impulse factor (peak value / mean); Frequency domain features: The time domain data is converted into the frequency domain by Fast Fourier Transform (FFT) to extract the main frequency amplitude, the spectral centroid (frequency domain mean), and the spectral width (frequency domain standard deviation). Operating condition perception characteristics: Actual power deviation rate from ideal power: ; The computational power of a normal digital twin model under the current G and T conditions; Fill factor decay: ; in, , The maximum power point current / voltage, , Given the short-circuit current / open-circuit voltage, calculate its deviation from the standard fill factor. MPPT efficiency: The ratio of real-time inverter output power to the theoretical maximum power under current operating conditions.

[0030] The features of each window are combined into a high-dimensional feature vector f to form the source domain feature set (the features corresponding to the synthetic data). The target domain feature set (features corresponding to the real data). .

[0031] S23. Initial Alignment of Source and Target Domains: A geodesic streaming basis learning method is employed. and The common subspace is learned within the Zhangcheng space, minimizing the mean difference of features between the two domains in the common subspace, thus reducing the domain adaptation difficulty of subsequent transfer learning.

[0032] S3. Construction and Training of Transfer Learning Framework with Physical Information Constraints S31, Network Architecture Design A multi-branch deep neural network is used, with the following structure: Shared feature encoder It consists of 1D convolutional layers (extracting local temporal features) and LSTM layers (extracting long-term dependent features), taking a feature vector f as input and outputting a high-level spatiotemporal feature representation h=E(f); Fault classifier Fully connected layer, input h, output fault category probability distribution .

[0033] Domain discriminator : Fully connected layer, input h, output the probability of the feature belonging to the domain (source domain / target domain), used for adversarial domain adaptation; Physical consistency constraint module Differentiable physical computation graph, decoding key physical quantities (such as...) from h , ), used to calculate physical consistency loss.

[0034] S32. Design of Coupling Loss Function The overall loss function is a weighted sum of three parts: ;in, , The domain adaptation loss term is determined through cross-validation and typically has a value of 0.1 to 1.0.

[0035] Mission loss Cross-entropy loss, calculated only on the source domain data, ensures the model's ability to classify faults. ; in The number of samples in the source domain. C Number of fault categories For the first i The true labels of each sample (one-hot encoded) To predict probabilities.

[0036] Adversity domain adaptation loss : Connect the encoder via a gradient inversion layer (GRL) E Domain discriminator D The encoder is trained to generate "domain-indistinguishable" features: ; in The number of samples in the target domain. , These are the sample features of the source domain and the target domain, respectively.

[0037] Physical consistency loss The constraint feature representation conforms to the physical laws of photovoltaic systems. The mean square error between the calculated decoded physical quantities and the theoretical physical quantities is: ; in, To decode physical quantity vectors (such as , ), Based on current environmental parameters ( , (a) A vector of theoretical physical quantities calculated by standard physical formulas.

[0038] S33, Alternating Iterative Training Strategy (1) Training domain discriminator D Fixed encoder E , minimize ,make D It can accurately distinguish between source domain and target domain features; (2) Training the encoder E With classifier C : Fixed D, minimize make E Generate "deception" D The domain-invariant feature of "" allows C to accurately classify faults; (3) Repeat steps (1) to (2) until the loss function converges (e.g., the loss decreases by less than 1 for 10 consecutive epochs). e -4).

[0039] S4, System Integration, Online Diagnostics and Continuous Learning S41, Cloud-Edge Collaborative Deployment Architecture, is as follows: Edge side: Deploy a lightweight feature extractor and model inference engine at the photovoltaic subarray gateway to collect subarray data (current, voltage, temperature) in real time, input the data into the model after preprocessing and feature extraction, and output preliminary diagnostic results (fault type / no fault), with response latency controlled within 1 second; In the cloud: Deploy a complete digital twin model, model training platform and database, receive diagnostic results and raw data uploaded from the edge, perform centralized analysis (such as multi-subarray fault correlation analysis), and store historical data.

[0040] S42. Online Diagnosis and Uncertainty Assessment: Fault Diagnosis: After edge-side preprocessing and feature extraction, online data is input into the transfer learning model to obtain the fault category probability distribution. If the maximum probability value is greater than the preset threshold (e.g., 0.85), it is determined to be the corresponding fault; if all probability values ​​are less than the threshold, it is determined to be a "suspected fault"; uncertainty quantification: using the MCDropout method, 10 to 20 inferences are performed on the same input sample (each inference randomly shuts down some network nodes), and the variance of the prediction results is calculated: if the variance is greater than the preset value (e.g., 0.1), an advanced alarm is triggered, and the operation and maintenance personnel are notified to manually review it.

[0041] S43, Continuous Learning and Digital Twin Updates Feedback loop: Maintenance personnel send the results of manual review (confirming the fault / eliminating false alarms) back to the cloud database as labeled data; Model update: The fault diagnosis model is fine-tuned monthly based on newly labeled data, and an elastic weight consolidation algorithm is used to avoid "catastrophic forgetting" (i.e., retaining knowledge learned from old data). Digital twin calibration: Calculate the deviation between the predicted data (such as power and temperature) of the digital twin model and the actual data, and update the twin model parameters (such as component attenuation coefficient and inverter efficiency parameters) in reverse through gradient descent to ensure dynamic matching between the twin model and the physical power plant.

[0042] Therefore, the present invention provides a photovoltaic power plant fault diagnosis method based on digital twins and transfer learning, which significantly improves the accuracy and real-time performance of photovoltaic fault diagnosis, enables early fault warning, adapts to equipment aging and changes in operating conditions, and is suitable for intelligent operation and maintenance of large-scale photovoltaic power plants.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for fault diagnosis of photovoltaic power plants based on digital twins and transfer learning, characterized in that, Includes the following steps: S1. Construct a digital twin model of the target photovoltaic power station. The digital twin model is based on the physical mechanism of the photovoltaic system and is used to simulate the operation behavior of the photovoltaic power station under normal conditions and at least one fault condition. S2. Run the digital twin model to generate a simulated operation dataset containing fault labels; S3. Obtain the actual operation dataset of the target photovoltaic power station; S4. Based on the transfer learning framework, a fault diagnosis model is trained using the simulated running dataset as the source domain data and the real running dataset as the target domain data; wherein, the loss function of the transfer learning framework includes a domain adaptation loss term for reducing the difference in feature distribution between the source domain data and the target domain data. S5. Use the trained fault diagnosis model to diagnose faults in the real-time operating data of the target photovoltaic power station.

2. The photovoltaic power station fault diagnosis method based on digital twin and transfer learning according to claim 1, characterized in that, In step S1, the digital twin model includes a photovoltaic module model, an inverter model, and an environmental model; the photovoltaic module model adopts a dual-diode model, and its IV characteristic relationship is defined by a physical equation including photocurrent, diode saturation current, series resistance, parallel resistance, and environmental parameters; the simulation of the fault state is achieved by dynamically modifying at least one parameter in the physical equation.

3. The photovoltaic power station fault diagnosis method based on digital twin and transfer learning according to claim 2, characterized in that, The dynamic modifications include: to simulate hot spot faults, the parallel resistance is modified to be a function of time and spatial location; to simulate potential induced decay faults, a leakage current term related to ground voltage and ambient humidity is introduced.

4. The photovoltaic power station fault diagnosis method based on digital twin and transfer learning according to claim 1, characterized in that, In step S4, the transfer learning framework adopts a deep neural network structure that includes a feature extractor, a fault classifier, and a domain discriminator; the domain adaptation loss term is calculated based on the discrimination result of the domain discriminator on the source domain features and target domain features output by the feature extractor.

5. A photovoltaic power station fault diagnosis method based on digital twin and transfer learning according to claim 4, characterized in that, The loss function also includes a physical consistency loss term, which is used to constrain the difference between the physical quantities that the features extracted by the feature extractor can decode through a preset physical computation graph and the theoretical physical quantities calculated by physical formulas from real-time environmental parameters to minimize the difference.

6. The photovoltaic power station fault diagnosis method based on digital twin and transfer learning according to claim 5, characterized in that, The loss function is expressed as: ； in, Classify the task loss for fault classification. For the domain adaptation loss term, For the physical consistency loss term, and To balance hyperparameters.

7. The photovoltaic power station fault diagnosis method based on digital twin and transfer learning according to claim 1, characterized in that, After step S3 and before step S4, the method further includes data preprocessing and feature engineering of the simulated running dataset and the real running dataset. The feature engineering includes extracting time-domain features, frequency-domain features, and operating condition-aware features, wherein the operating condition-aware features include the deviation rate between the actual power and the ideal power calculated by the digital twin model.

8. A photovoltaic power plant fault diagnosis system based on digital twins and transfer learning for implementing the method of any one of claims 1-7, characterized in that, include: A digital twin module is used to build and run a digital twin model of the target photovoltaic power station and generate the simulation operation dataset; The data acquisition and processing module is used to acquire and process the actual running dataset and real-time running data; The model training module is used to implement the transfer learning framework and train the fault diagnosis model. The fault diagnosis module is used to perform real-time fault diagnosis and early warning using a trained model.

9. A computing device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

Images