A new energy equipment fault early warning system for a photovoltaic power station

Abstract

This invention belongs to the technical field of photovoltaic equipment fault early warning systems. It discloses a fault early warning system for new energy equipment in photovoltaic power plants, comprising: first, collecting photovoltaic system parameters at a preset frequency to form a time-series database; then, preprocessing the parameters to generate a physical feature dataset and an analytical feature dataset; next, constructing a dynamic Bayesian network model and a machine learning classification model, respectively performing fault analysis based on the physical feature dataset and the analytical feature dataset, generating two fault probability distributions; then, calculating the consistency index of the two distributions, adopting corresponding decision-making strategies based on the consistency index, and generating a comprehensive fault probability; finally, determining the fault state based on the comparison result of the fault probability value and a preset threshold, and generating an early warning signal. This invention achieves intelligent fusion of the results of the two models by integrating a physical model and a data-driven method, improving the accuracy and reliability of fault early warning.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic equipment fault early warning system technology, and more specifically, to a new energy equipment fault early warning system for photovoltaic power plants. Background Technology

[0002] With the global energy structure transformation and the rapid development of renewable energy, photovoltaic (PV) power generation, as an important component of clean energy, is being widely promoted and applied globally. PV power plants, due to their advantages such as being pollution-free, noise-free, and resource-renewable, have become a major force in the development of new energy. However, with the rapid growth of PV installed capacity, the operation and maintenance management of PV power plants faces increasingly greater challenges, among which fault diagnosis and early warning are crucial links in ensuring the safe and efficient operation of PV power plants.

[0003] Currently, photovoltaic power plant fault diagnosis technology has evolved from traditional manual inspections to model-based intelligent diagnostic methods. Existing fault diagnosis systems typically employ physical model-based or data-driven methods for fault identification. However, these systems suffer from a lack of diversity in their fault diagnosis models, making it difficult to cope with the complex and ever-changing fault scenarios in photovoltaic power plants. Most systems either use physical model-based methods or data-driven methods, lacking an organic integration of the two approaches. While physical model-based methods offer good interpretability, their ability to identify complex nonlinear fault modes is limited; while data-driven methods, although capable of capturing complex patterns, lack physical mechanism support and suffer from insufficient reliability under sparse or abnormal data conditions. Summary of the Invention

[0004] To overcome the aforementioned problems of the prior art, this invention proposes a fault early warning system for new energy equipment in photovoltaic power plants, which is used to solve the above problems.

[0005] This invention provides the following technical solution:

[0006] A fault early warning system for new energy equipment in photovoltaic power plants includes:

[0007] The data acquisition module is used to collect photovoltaic system parameters at a preset frequency to form a time series database;

[0008] The data preprocessing module is used to preprocess the photovoltaic system parameters in the time series database to generate physical feature datasets and analytical feature datasets.

[0009] The Bayesian inference module is used to build dynamic Bayesian network models, perform fault inference based on physical feature datasets, and generate Bayesian fault probability distributions.

[0010] The machine learning module is used to build machine learning classification models, classify faults based on the analyzed feature dataset, and generate a learned fault probability distribution.

[0011] The collaborative decision-making module is used to receive the Bayesian fault probability distribution and the learned fault probability distribution, calculate the consistency index of the two distributions, take corresponding decision-making strategies based on the consistency index, and generate a comprehensive fault probability.

[0012] The early warning generation module is used to extract the probability values ​​of each fault type from the comprehensive fault probability distribution, determine the fault status based on the comparison results of the fault probability values ​​and preset thresholds, and generate corresponding early warning signals.

[0013] Preferably, the photovoltaic system parameters include electrical parameters and environmental parameters; the electrical parameters include voltage, current, and power; the environmental parameters include module temperature, ambient temperature, and irradiance.

[0014] Preferably, the physical feature dataset includes component temperature coefficient, current-voltage curve feature points, and power-voltage curve feature points; the analytical feature dataset includes time-domain features, frequency-domain features, and statistical features.

[0015] Preferably, the preprocessing of photovoltaic system parameters in the time series database includes:

[0016] This is used to standardize photovoltaic system parameters in a time series database. The standardization process includes outlier detection and removal, missing data imputation, and noise filtering.

[0017] Based on the standardized photovoltaic system parameters, a physical feature dataset is extracted, including: calculating the temperature difference between the module temperature and the ambient temperature, recording the temperature difference changes under different irradiance intensities, and obtaining the module temperature response coefficient to irradiance intensity through linear regression analysis as the module temperature coefficient.

[0018] By collecting voltage and current data, a current-voltage curve is constructed, and feature points of the current-voltage curve are extracted. The feature points of the current-voltage curve include short-circuit current, open-circuit voltage, current and voltage corresponding to the maximum power point, and the slope change points of the current-voltage curve. By collecting power and voltage data, a power-voltage curve is constructed, and feature points of the power-voltage curve are extracted. The feature points of the power-voltage curve include the maximum power point and its corresponding voltage, the rise rate and fall rate of the power curve, and the inflection point position of the power curve.

[0019] The analysis is based on the standardized photovoltaic system parameters, and the extracted feature dataset includes: sliding window analysis of the time series of voltage, current and power, calculating the maximum, minimum, mean, variance and trend slope within each window to form time-domain features; fast Fourier transform of the time series of voltage and current to extract power spectral density and dominant frequency components to form frequency-domain features; and calculation of skewness, kurtosis and interquartile range of voltage, current, power, module temperature and irradiance to form statistical features.

[0020] Preferably, the construction of the dynamic Bayesian network model includes:

[0021] The initial Bayesian network is constructed with a three-layer structure, including a fault node layer to represent the possible fault types of the photovoltaic system; a state parameter layer to represent the intermediate state variables of the photovoltaic system; and an observation parameter layer to represent the parameters in the physical feature dataset.

[0022] Create multiple time slices, each containing a complete three-layer network structure; define the conditional probability relationships between nodes within the same time slice; define the state transition probabilities of state parameter nodes between adjacent time slices; and define the observation probabilities between state parameter nodes and observation parameter nodes in each time slice.

[0023] Based on historical fault data and corresponding physical feature data, a dynamic Bayesian network is trained using the expectation-maximization algorithm to obtain a trained dynamic Bayesian model.

[0024] The step of performing fault reasoning based on the physical feature dataset and generating a Bayesian fault probability distribution includes:

[0025] The dynamic Bayesian network trained by inputting the physical feature dataset as observation evidence is used to calculate the posterior probability of each fault node through the joint tree algorithm, generating a Bayesian fault probability distribution. The Bayesian fault probability distribution is represented as a multidimensional vector of each fault type and its corresponding probability value.

[0026] Preferably, the construction of the machine learning classification model includes:

[0027] A multilayer perceptron neural network is constructed, including an input layer for receiving and analyzing feature datasets; a hidden layer for feature transformation and abstract representation; and an output layer for representing the probability distribution of various faults. A hierarchical feature fusion method is adopted to weight and combine time-domain features, frequency-domain features, and statistical features to form a comprehensive feature vector as the input of the neural network.

[0028] Based on historically labeled fault data, a neural network is trained using the backpropagation algorithm, including initializing network weight parameters; calculating the prediction results of forward propagation; calculating the cross-entropy loss between the prediction results and the true labels; updating network parameters using gradient descent; repeating the above steps until the loss function converges to obtain a trained machine learning classification model.

[0029] The step of classifying faults based on the analyzed feature dataset and generating a learned fault probability distribution includes:

[0030] The feature dataset is input into the trained machine learning classification model, and the probability values ​​of various faults are calculated through forward propagation to generate a learned fault probability distribution, which is represented as a multi-dimensional vector of each fault type and its corresponding probability value.

[0031] Preferably, the calculation of the consistency index between the two distributions includes:

[0032] The reciprocal of the bulldozer distance between the Bayesian fault probability distribution and the learned fault probability distribution is calculated as an index of consistency between the two distributions.

[0033] The step of adopting corresponding decision-making strategies based on consistency indicators to generate a comprehensive failure probability includes:

[0034] When the consistency index is higher than the preset threshold, the arithmetic mean of the Bayesian fault probability distribution and the learned fault probability distribution is directly calculated to generate a comprehensive fault probability distribution.

[0035] When the consistency index is lower than the preset threshold, the stability index of the two models is quantified, and a weighted average is performed based on the stability index to generate a comprehensive failure probability distribution.

[0036] Preferably, the stability indices for quantifying the two models include:

[0037] For dynamic Bayesian network models: calculate the standard deviation matrix of the Bayesian failure probability distribution over at least the two most recent time windows, and take the square root of the sum of the squares of its elements as the volatility index. Calculate the standard deviation of the rate of change of the physical feature dataset relative to its historical average, and use it as an input stability index. According to the formula The stability index of the dynamic Bayesian network model was calculated. In the formula, and These are the weighting coefficients for the preset volatility index and the input stability index, respectively.

[0038] For machine learning models: calculate the standard deviation matrix of the learned fault probability distribution within the most recent two time windows, and take the square root of the sum of squares of its elements as the dissimilarity index. Calculate the standard deviation of the rate of change of the feature dataset relative to the historical average, and use it as an input fit metric. According to the formula Calculate the stability index of machine learning models , and These are the weighting coefficients for the preset difference index and the input adaptability index, respectively.

[0039] The step of generating a comprehensive failure probability distribution by weighted averaging based on stability indices includes:

[0040] According to the stability indices of the two models and ,use As weights in a Bayesian model, use As weights in machine learning models;

[0041] Using the obtained weights, a weighted average is performed on the Bayesian fault probability distribution and the learned fault probability distribution to generate a comprehensive fault probability distribution.

[0042] Preferably, the step of determining the fault state based on the comparison result between the fault probability value and a preset threshold, and generating a corresponding early warning signal and diagnostic report includes:

[0043] Extract the probability value of each fault type from the comprehensive fault probability distribution; compare the probability value of each fault type with the corresponding warning threshold; for fault types whose probability value exceeds the warning threshold, generate the corresponding fault warning signal.

[0044] This invention provides a fault early warning system for new energy equipment in photovoltaic power plants, which has the following beneficial effects:

[0045] By employing a Bayesian inference module based on a physical feature dataset for fault reasoning, the system can fully leverage the physical mechanisms and causal relationships of photovoltaic systems, demonstrating high interpretability and accuracy for faults with clear physical manifestations. Conversely, a machine learning module, based on an analytical feature dataset for fault classification, can capture complex nonlinear relationships and hidden fault patterns, exhibiting strong identification capabilities for faults sensitive to data patterns. The complementary advantages of these two models significantly enhance the system's ability to identify various faults, making fault warnings more comprehensive and reliable. By calculating the consistency index between the Bayesian fault probability distribution and the learned fault probability distribution, and adopting corresponding decision-making strategies based on this index, intelligent fusion of the results from the two models is achieved. This adaptive fusion mechanism dynamically adjusts the weights of each model in the final decision based on the reliability of different scenarios, effectively improving the overall prediction accuracy and robustness of the system. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of a module of a new energy equipment fault early warning system for photovoltaic power plants according to the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1

[0049] Please see Figure 1 In this embodiment, a fault early warning system for new energy equipment in a photovoltaic power station includes:

[0050] The data acquisition module is used to collect photovoltaic system parameters at a preset frequency to form a time series database;

[0051] The photovoltaic system parameters include electrical parameters and environmental parameters; the electrical parameters include voltage, current, and power; the environmental parameters include module temperature, ambient temperature, and irradiance.

[0052] In this embodiment, multiple sensors can be installed in the photovoltaic power station to collect system parameters. Electrical parameter acquisition includes installing voltage and current sensors at the photovoltaic strings and inverters, and power data is obtained through real-time calculation of voltage and current.

[0053] Environmental parameter acquisition includes installing temperature sensors on the back of the photovoltaic modules to measure module temperature, installing temperature sensors around the array to measure ambient temperature, and installing irradiance meters on the array plane to measure solar irradiance.

[0054] The data collection frequency can be set to once per minute, and all data is timestamped. The data can be organized and stored according to device ID, parameter type, and timestamp to form a complete time-series database, providing basic data support for subsequent analysis.

[0055] The data preprocessing module is used to preprocess the photovoltaic system parameters in the time series database to generate physical feature datasets and analytical feature datasets.

[0056] The physical feature dataset includes component temperature coefficient, current-voltage curve feature points, and power-voltage curve feature points; the analytical feature dataset includes time-domain features, frequency-domain features, and statistical features.

[0057] The preprocessing of photovoltaic system parameters in the time series database includes:

[0058] This is used to standardize photovoltaic system parameters in a time series database. The standardization process includes outlier detection and removal, missing data imputation, and noise filtering.

[0059] Based on the standardized photovoltaic system parameters, a physical feature dataset is extracted, including: calculating the temperature difference between the module temperature and the ambient temperature, recording the temperature difference changes under different irradiance intensities, and obtaining the module temperature response coefficient to irradiance intensity through linear regression analysis as the module temperature coefficient.

[0060] By collecting voltage and current data, a current-voltage curve is constructed, and feature points of the current-voltage curve are extracted. The feature points of the current-voltage curve include short-circuit current, open-circuit voltage, current and voltage corresponding to the maximum power point, and the slope change points of the current-voltage curve. By collecting power and voltage data, a power-voltage curve is constructed, and feature points of the power-voltage curve are extracted. The feature points of the power-voltage curve include the maximum power point and its corresponding voltage, the rise rate and fall rate of the power curve, and the inflection point position of the power curve.

[0061] The analysis is based on the standardized photovoltaic system parameters, and the extracted feature dataset includes: sliding window analysis of the time series of voltage, current and power, calculating the maximum, minimum, mean, variance and trend slope within each window to form time-domain features; fast Fourier transform of the time series of voltage and current to extract power spectral density and dominant frequency components to form frequency-domain features; and calculation of skewness, kurtosis and interquartile range of voltage, current, power, module temperature and irradiance to form statistical features.

[0062] In this embodiment, the data preprocessing module first standardizes the photovoltaic system parameters in the time series database. The standardization process includes three steps: outlier detection and removal, missing data imputation, and noise filtering.

[0063] After standardization, the data preprocessing module further extracts the physical feature dataset. First, it calculates the temperature difference between the module and the ambient temperature, recording the temperature difference changes under different irradiation intensities. Then, by performing linear regression analysis on these data, it obtains the module temperature response coefficient to irradiation intensity, i.e., the module temperature coefficient. This coefficient reflects the module's thermal characteristics and is an important indicator for judging module faults such as hot spots and thermal degradation.

[0064] Next, the data preprocessing module constructs current-voltage curves using the collected voltage and current data and extracts their feature points. These feature points include short-circuit current, open-circuit voltage, the current and voltage corresponding to the maximum power point, and the slope changes of the current-voltage curve. The short-circuit current reflects the photoelectric conversion capability of the module, while the maximum power point reflects the energy conversion efficiency of the system. The slope changes of the curve correspond to changes in the internal resistance of the module, which can indicate potential local faults. Similarly, by constructing power-voltage curves using power and voltage data, feature points such as the maximum power point and its corresponding voltage, the rise and fall rates of the power curve, and the inflection point of the power curve are extracted. Changes in these feature points can reflect performance degradation and abnormal operating states of the module or system.

[0065] When extracting features from the dataset, the data preprocessing module first performs sliding window analysis on the time series of voltage, current, and power, calculating the maximum, minimum, mean, variance, and trend slope within each window to form time-domain features. These time-domain features reflect the dynamic changes in system parameters and are of great significance for detecting intermittent faults and performance degradation.

[0066] Secondly, a Fast Fourier Transform (FFT) is performed on the time series of voltage and current to extract the power spectral density and dominant frequency components, forming frequency domain features. Frequency domain analysis can reveal periodic changes and hidden patterns in the time series, making it particularly suitable for detecting problems caused by oscillations, harmonic interference, and grid interactions.

[0067] Finally, the skewness, kurtosis, and interquartile range of voltage, current, power, component temperature, and irradiance are calculated to form statistical characteristics. These statistical characteristics reflect the morphological features of the data distribution and are suitable for detecting abnormal changes in system behavior.

[0068] The Bayesian inference module is used to build dynamic Bayesian network models, perform fault inference based on physical feature datasets, and generate Bayesian fault probability distributions.

[0069] The construction of the dynamic Bayesian network model includes:

[0070] The initial Bayesian network is constructed with a three-layer structure, including a fault node layer to represent the possible fault types of the photovoltaic system; a state parameter layer to represent the intermediate state variables of the photovoltaic system; and an observation parameter layer to represent the parameters in the physical feature dataset.

[0071] Create multiple time slices, each containing a complete three-layer network structure; define the conditional probability relationships between nodes within the same time slice; define the state transition probabilities of state parameter nodes between adjacent time slices; and define the observation probabilities between state parameter nodes and observation parameter nodes in each time slice.

[0072] Based on historical fault data and corresponding physical feature data, a dynamic Bayesian network is trained using the expectation-maximization algorithm to obtain a trained dynamic Bayesian model.

[0073] The step of performing fault reasoning based on the physical feature dataset and generating a Bayesian fault probability distribution includes:

[0074] The dynamic Bayesian network trained by inputting the physical feature dataset as observation evidence is used to calculate the posterior probability of each fault node through the joint tree algorithm, generating a Bayesian fault probability distribution. The Bayesian fault probability distribution is represented as a multidimensional vector of each fault type and its corresponding probability value.

[0075] In this embodiment, the Bayesian inference module first constructs a three-layer structure of the initial Bayesian network. The fault node layer, located at the top layer, contains the main fault types that may occur in the photovoltaic system, such as module hotspot faults, bypass diode faults, module breakage, module PID effects, string connection faults, and inverter faults. The state parameter layer, located in the middle layer, represents the intermediate state variables of the system, such as the module temperature distribution state, module electrical characteristic state, and inverter operating state. These state variables cannot be directly observed but can be inferred from observed parameters. The observation parameter layer, located at the bottom layer, corresponds to parameters in the physical feature dataset, including the module temperature coefficient, current-voltage curve feature points, and power-voltage curve feature points.

[0076] To capture the dynamic changes in system state over time, the Bayesian inference module creates multiple time slices, each containing a complete three-layer network structure. The time slice intervals are set to one hour, enabling the model to capture short-term changes in system state. Within the same time slice, the Bayesian inference module defines conditional probability relationships between nodes. For example, a conditional probability relationship exists between the component hotspot fault node and the component temperature distribution state node, representing the probability distribution of the component temperature distribution state given that a hotspot fault has occurred.

[0077] Between adjacent time slices, the Bayesian inference module defines the state transition probabilities for the state parameter nodes. These state transition probabilities describe how the system state evolves over time; for example, how the current electrical characteristic state of a component affects its electrical characteristic state at the next time step. This time dependency allows the model to account for the development of system faults and the influence of historical factors. Furthermore, the Bayesian inference module also defines the observation probabilities between the state parameter nodes and observation parameter nodes within each time slice, describing the probability of observing a specific parameter value given a system state.

[0078] Training the dynamic Bayesian network is a crucial step in the Bayesian inference module. Based on historical fault data and corresponding physical feature data, the Bayesian inference module uses the Expectation-Maximization (EM) algorithm to train the dynamic Bayesian network. The training process consists of two alternating steps: an expectation step and a maximization step. In the expectation step, the posterior distribution of the latent variables is estimated based on the current model parameters; in the maximization step, the model parameters are updated based on the estimated latent variable distribution. Through multiple iterations, the EM algorithm progressively optimizes the model parameters, enabling the model to better fit the historical data.

[0079] After training, the Bayesian inference module uses the trained dynamic Bayesian network for fault reasoning. When a new physical feature dataset arrives, the Bayesian inference module inputs it as observational evidence into the dynamic Bayesian network. Precise inference is performed using the joint tree algorithm to calculate the posterior probability of each faulty node. The joint tree algorithm first transforms the Bayesian network into an acyclic undirected graph, then constructs a joint tree structure, and performs probability propagation on the tree through a message-passing mechanism, ultimately calculating the marginal posterior probability of the target node. The result of Bayesian inference is a Bayesian fault probability distribution, represented as a multidimensional vector of each fault type and its corresponding probability value.

[0080] The machine learning module is used to build machine learning classification models, classify faults based on the analyzed feature dataset, and generate a learned fault probability distribution.

[0081] The construction of the machine learning classification model includes:

[0082] A multilayer perceptron neural network is constructed, including an input layer for receiving and analyzing feature datasets; a hidden layer for feature transformation and abstract representation; and an output layer for representing the probability distribution of various faults. A hierarchical feature fusion method is adopted to weight and combine time-domain features, frequency-domain features, and statistical features to form a comprehensive feature vector as the input of the neural network.

[0083] Based on historically labeled fault data, a neural network is trained using the backpropagation algorithm, including initializing network weight parameters; calculating the prediction results of forward propagation; calculating the cross-entropy loss between the prediction results and the true labels; updating network parameters using gradient descent; repeating the above steps until the loss function converges to obtain a trained machine learning classification model.

[0084] The step of classifying faults based on the analyzed feature dataset and generating a learned fault probability distribution includes:

[0085] The feature dataset is input into the trained machine learning classification model, and the probability values ​​of various faults are calculated through forward propagation to generate a learned fault probability distribution, which is represented as a multi-dimensional vector of each fault type and its corresponding probability value.

[0086] In this embodiment, the machine learning module first constructs a multilayer perceptron neural network. This neural network comprises three main parts: an input layer, hidden layers, and an output layer. The input layer receives and analyzes the feature dataset; its number of nodes equals the dimension of the feature vectors. The hidden layers use the ReLU activation function, and the output layer has nodes equal to the number of fault types. The output layer uses the Softmax activation function to transform the neural network's output into a probability distribution.

[0087] The neural network training process can use backpropagation to train the network based on historically labeled fault data. First, the network weight parameters are initialized. Next, training samples are input into the network, and the prediction results, i.e., the probability distribution of various fault types, are calculated through forward propagation. Then, the cross-entropy loss between the prediction results and the true labels is calculated. After the loss calculation is complete, the machine learning module updates the network parameters using gradient descent. To prevent overfitting, L2 regularization and Dropout techniques can also be applied, with the Dropout rate set to 0.3. The training process uses a mini-batch training method, with a batch size of 64, and repeats the above steps until the loss function converges or the preset maximum number of training epochs (200 epochs) is reached.

[0088] After training is complete, when a new dataset of analytical features arrives, the machine learning module inputs it into the trained machine learning classification model. Through forward propagation calculations, the model generates probability values ​​for various types of faults, forming a learned fault probability distribution.

[0089] The collaborative decision-making module is used to receive the Bayesian fault probability distribution and the learned fault probability distribution, calculate the consistency index of the two distributions, take corresponding decision-making strategies based on the consistency index, and generate a comprehensive fault probability.

[0090] The consistency index for calculating the two distributions includes:

[0091] The reciprocal of the bulldozer distance between the Bayesian fault probability distribution and the learned fault probability distribution is calculated as an index of consistency between the two distributions.

[0092] The step of adopting corresponding decision-making strategies based on consistency indicators to generate a comprehensive failure probability includes:

[0093] When the consistency index is higher than the preset threshold, the arithmetic mean of the Bayesian fault probability distribution and the learned fault probability distribution is directly calculated to generate a comprehensive fault probability distribution.

[0094] When the consistency index is lower than the preset threshold, the stability index of the two models is quantified, and a weighted average is performed based on the stability index to generate a comprehensive failure probability distribution.

[0095] The stability metrics for the two models include:

[0096] For dynamic Bayesian network models: calculate the standard deviation matrix of the Bayesian failure probability distribution over at least the two most recent time windows, and take the square root of the sum of the squares of its elements as the volatility index. Calculate the standard deviation of the rate of change of the physical feature dataset relative to its historical average, and use it as an input stability index. According to the formula The stability index of the dynamic Bayesian network model was calculated. In the formula, and These are the weighting coefficients for the preset volatility index and the input stability index, respectively.

[0097] For machine learning models: calculate the standard deviation matrix of the learned fault probability distribution within the most recent two time windows, and take the square root of the sum of squares of its elements as the dissimilarity index. Calculate the standard deviation of the rate of change of the feature dataset relative to the historical average, and use it as an input fit metric. According to the formula Calculate the stability index of machine learning models , and These are the weighting coefficients for the preset difference index and the input adaptability index, respectively.

[0098] The step of generating a comprehensive failure probability distribution by weighted averaging based on stability indices includes:

[0099] According to the stability indices of the two models and ,use As weights in a Bayesian model, use As weights in machine learning models;

[0100] Using the obtained weights, a weighted average is performed on the Bayesian fault probability distribution and the learned fault probability distribution to generate a comprehensive fault probability distribution.

[0101] In this embodiment, the collaborative decision-making module first calculates the reciprocal of the bulldozer distance between the Bayesian fault probability distribution and the learned fault probability distribution, using it as a consistency index between the two distributions. The bulldozer distance is an effective measure of the difference between two probability distributions; taking its reciprocal as the consistency index means that the smaller the distance, the higher the consistency. A larger consistency index value indicates that the two distributions are closer, meaning that the two models based on different principles are more consistent in their judgments of fault types.

[0102] When the prediction results of two models show high consistency, i.e., the consistency index is higher than a preset threshold such as 0.75, it indicates that the two models have a high degree of consistency in their judgment of the current fault, which usually means that the prediction results are highly reliable. In this case, the collaborative decision-making module directly performs an arithmetic mean on the Bayesian fault probability distribution and the learned fault probability distribution to generate a comprehensive fault probability distribution. Arithmetic averaging is a simple and effective method that can combine the advantages of the two models and improve the stability of the prediction.

[0103] When the prediction results of two models show low consistency, i.e., the consistency index is below a preset threshold, it indicates a significant discrepancy in their judgments of the current fault. This may mean that the prediction accuracy of one model is reduced under the current circumstances. In this case, a simple arithmetic average may introduce errors and reduce prediction accuracy. Therefore, the collaborative decision-making module needs to further quantify the stability index of the two models and perform a weighted average based on the stability index to favor the more stable and reliable model.

[0104] To quantify the stability of dynamic Bayesian network models, the collaborative decision-making module calculates the standard deviation matrix of the Bayesian failure probability distribution over at least two recent time windows, and takes the square root of the sum of squares of its elements as a volatility index. The volatility index reflects the temporal stability of the Bayesian model's prediction results; a smaller value indicates a more stable model output. Simultaneously, the standard deviation of the rate of change of the physical feature dataset relative to its historical average is calculated as an input stability index. The input stability index reflects the degree of change in the input data; a smaller value indicates a more stable input data. A larger stability index for the dynamic Bayesian network model calculated by the formula indicates a higher probability of accurate predictions; similarly, a larger stability index for the machine learning model calculated by the formula indicates a higher probability of accurate predictions.

[0105] Based on the stability indices of the two models obtained from the calculation, the collaborative decision-making module further calculates the normalized weights.

[0106] Finally, the collaborative decision-making module uses the calculated weights to perform a weighted average of the Bayesian fault probability distribution and the learned fault probability distribution to generate a comprehensive fault probability distribution.

[0107] This weighted averaging strategy based on stability indices adaptively adjusts the influence of different models in the final decision, leveraging the complementary strengths of the Bayesian inference module and the machine learning module. The Bayesian inference module reasons based on physical features and causal relationships, offering high interpretability and accuracy for faults with clear physical mechanisms; while the machine learning module performs data-driven classification, capturing complex nonlinear relationships and hidden fault modes. Through the calculation of consistency and stability indices, the collaborative decision-making module intelligently assesses the reliability of the two models and generates a more accurate comprehensive fault probability distribution, thereby improving the overall predictive accuracy and reliability of the system.

[0108] The early warning generation module is used to extract the probability values ​​of each fault type from the comprehensive fault probability distribution, determine the fault status based on the comparison results of the fault probability values ​​and preset thresholds, and generate corresponding early warning signals.

[0109] The step of determining the fault state based on the comparison result between the fault probability value and the preset threshold, and generating the corresponding early warning signal and diagnostic report includes:

[0110] Extract the probability value of each fault type from the comprehensive fault probability distribution; compare the probability value of each fault type with the corresponding warning threshold; for fault types whose probability value exceeds the warning threshold, generate the corresponding fault warning signal.

[0111] In this embodiment, the early warning generation module extracts the probability values ​​of each fault type from the comprehensive fault probability distribution and compares them with preset early warning thresholds. Different early warning thresholds are set for different fault types, such as 0.6 for component hot spot faults, 0.7 for bypass diode faults, and 0.75 for component PID effects.

[0112] When the probability value of a certain type of fault exceeds its corresponding warning threshold, the warning generation module generates a corresponding warning signal. Relevant personnel are then notified through various means, such as system notifications, SMS messages, and telephone calls.

[0113] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only one method, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0114] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

[0115] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault early warning system for new energy equipment in photovoltaic power plants, characterized in that, include: The data acquisition module is used to collect photovoltaic system parameters at a preset frequency to form a time series database; The data preprocessing module is used to preprocess the photovoltaic system parameters in the time series database to generate physical feature datasets and analytical feature datasets. The physical feature dataset includes component temperature coefficient, current-voltage curve feature points, and power-voltage curve feature points; the analytical feature dataset includes time-domain features, frequency-domain features, and statistical features. Based on the standardized photovoltaic system parameters, a physical feature dataset is extracted, including: calculating the temperature difference between the module temperature and the ambient temperature, recording the temperature difference changes under different irradiance intensities, and obtaining the module temperature response coefficient to irradiance intensity through linear regression analysis as the module temperature coefficient. By collecting voltage and current data, a current-voltage curve is constructed, and feature points of the current-voltage curve are extracted. The feature points of the current-voltage curve include short-circuit current, open-circuit voltage, current and voltage corresponding to the maximum power point, and the slope change points of the current-voltage curve. By collecting power and voltage data, a power-voltage curve is constructed, and feature points of the power-voltage curve are extracted. The feature points of the power-voltage curve include the maximum power point and its corresponding voltage, the rise rate and fall rate of the power curve, and the inflection point position of the power curve. The analysis is based on the standardized photovoltaic system parameters, and the extracted feature dataset includes: sliding window analysis of the time series of voltage, current, and power, calculating the maximum, minimum, mean, variance, and trend slope within each window to form time-domain features; fast Fourier transform of the time series of voltage and current to extract the power spectral density and dominant frequency components to form frequency-domain features; and calculation of the skewness, kurtosis, and interquartile range of voltage, current, power, module temperature, and irradiance to form statistical features. The Bayesian inference module is used to build dynamic Bayesian network models, perform fault inference based on physical feature datasets, and generate Bayesian fault probability distributions. The machine learning module is used to build machine learning classification models, classify faults based on the analyzed feature dataset, and generate a learned fault probability distribution. The collaborative decision-making module is used to receive the Bayesian fault probability distribution and the learned fault probability distribution, calculate the consistency index of the two distributions, take corresponding decision-making strategies based on the consistency index, and generate a comprehensive fault probability. The consistency index for calculating the two distributions includes: The reciprocal of the bulldozer distance between the Bayesian fault probability distribution and the learned fault probability distribution is calculated as an index of consistency between the two distributions. The step of adopting corresponding decision-making strategies based on consistency indicators to generate a comprehensive failure probability includes: When the consistency index is higher than the preset threshold, the arithmetic mean of the Bayesian fault probability distribution and the learned fault probability distribution is directly calculated to generate a comprehensive fault probability distribution. When the consistency index is lower than the preset threshold, the stability index of the two models is quantified, and a weighted average is performed based on the stability index to generate a comprehensive failure probability distribution. The early warning generation module is used to extract the probability values ​​of each fault type from the comprehensive fault probability distribution, determine the fault status based on the comparison results of the fault probability values ​​and preset thresholds, and generate corresponding early warning signals.

2. The fault early warning system for new energy equipment in a photovoltaic power station according to claim 1, characterized in that, The photovoltaic system parameters include electrical parameters and environmental parameters; the electrical parameters include voltage, current, and power; the environmental parameters include module temperature, ambient temperature, and irradiance.

3. The fault early warning system for new energy equipment in a photovoltaic power station according to claim 1, characterized in that, The preprocessing of photovoltaic system parameters in the time series database includes: This is used to standardize photovoltaic system parameters in a time series database. The standardization process includes outlier detection and removal, missing data imputation, and noise filtering.

4. A fault early warning system for new energy equipment in a photovoltaic power station according to claim 1, characterized in that, The construction of the dynamic Bayesian network model includes: The initial Bayesian network is constructed with a three-layer structure, including a fault node layer to represent the possible fault types of the photovoltaic system; a state parameter layer to represent the intermediate state variables of the photovoltaic system; and an observation parameter layer to represent the parameters in the physical feature dataset. Create multiple time slices, each containing a complete three-layer network structure; define the conditional probability relationships between nodes within the same time slice; define the state transition probabilities of state parameter nodes between adjacent time slices; and define the observation probabilities between state parameter nodes and observation parameter nodes in each time slice. Based on historical fault data and corresponding physical feature data, a dynamic Bayesian network is trained using the expectation-maximization algorithm to obtain a trained dynamic Bayesian model. The step of performing fault reasoning based on the physical feature dataset and generating a Bayesian fault probability distribution includes: The dynamic Bayesian network trained by inputting the physical feature dataset as observation evidence is used to calculate the posterior probability of each fault node through the joint tree algorithm, generating a Bayesian fault probability distribution. The Bayesian fault probability distribution is represented as a multidimensional vector of each fault type and its corresponding probability value.

5. A fault early warning system for new energy equipment in a photovoltaic power station according to claim 4, characterized in that, The construction of the machine learning classification model includes: Construct a multilayer perceptron neural network, including an input layer for receiving and analyzing feature datasets; hidden layers for feature transformation and abstract representation; and an output layer for representing the probability distribution of various faults. A hierarchical feature fusion method is adopted to weight and combine time-domain features, frequency-domain features, and statistical features to form a comprehensive feature vector as the input of the neural network; Based on historically labeled fault data, a neural network is trained using the backpropagation algorithm, including initializing network weight parameters; calculating the prediction results of forward propagation; calculating the cross-entropy loss between the prediction results and the true labels; updating network parameters using gradient descent; repeating the above steps until the loss function converges to obtain a trained machine learning classification model. The step of classifying faults based on the analyzed feature dataset and generating a learned fault probability distribution includes: The feature dataset is input into the trained machine learning classification model, and the probability values ​​of various faults are calculated through forward propagation to generate a learned fault probability distribution, which is represented as a multi-dimensional vector of each fault type and its corresponding probability value.

6. A fault early warning system for new energy equipment in a photovoltaic power station according to claim 1, characterized in that, The stability metrics for the two models include: For dynamic Bayesian network models: calculate the standard deviation matrix of the Bayesian failure probability distribution over at least the two most recent time windows, and take the square root of the sum of the squares of its elements as the volatility index. Calculate the standard deviation of the rate of change of the physical feature dataset relative to its historical average, and use it as an input stability index. According to the formula The stability index of the dynamic Bayesian network model was calculated. In the formula, and These are the weighting coefficients for the preset volatility index and the input stability index, respectively. For machine learning models: calculate the standard deviation matrix of the learned fault probability distribution within the most recent two time windows, and take the square root of the sum of squares of its elements as the dissimilarity index. Calculate the standard deviation of the rate of change of the feature dataset relative to the historical average, and use it as an input fit metric. According to the formula Calculate the stability index of machine learning models , and These are the weighting coefficients for the preset difference index and the input adaptability index, respectively. The step of generating a comprehensive failure probability distribution by weighted averaging based on stability indices includes: According to the stability indices of the two models and ,use As weights in a Bayesian model, use As weights in machine learning models; Using the obtained weights, a weighted average is performed on the Bayesian fault probability distribution and the learned fault probability distribution to generate a comprehensive fault probability distribution.

7. A fault early warning system for new energy equipment in a photovoltaic power station according to claim 1, characterized in that, The step of determining the fault state based on the comparison result between the fault probability value and the preset threshold, and generating the corresponding early warning signal and diagnostic report includes: Extract the probability value of each fault type from the comprehensive fault probability distribution; compare the probability value of each fault type with the corresponding warning threshold; for fault types whose probability value exceeds the warning threshold, generate the corresponding fault warning signal.

Images