Photovoltaic module fault prediction method and system based on deep learning

Abstract

The invention discloses a photovoltaic module fault prediction method and system based on deep learning, and relates to the technical field of photovoltaic power generation, and the method comprises the steps: building a battery parameter data set through a photovoltaic system simulation model, and analyzing the changes of a photovoltaic characteristic curve under different fault types; based on the change of the photovoltaic characteristic curve, constructing a fault prediction model based on Transform, and performing correlation analysis on extracted change parameters in combination with an attention mechanism; performing parameter optimization and learning rate control by adopting a U-Net decoder, a ReLU activation function and an Adam optimizer on the basis of correlation analysis of variable parameters; a SoftMax function is introduced to classify the severity of the faults; the system comprises a multi-source heterogeneous data acquisition module, a dynamic feature extraction module, a space-time double-flow Transform prediction model, an edge calculation deployment module and an online incremental learning module. The method has better adaptability and generalization ability when facing diversified fault conditions of an actual photovoltaic module, and can identify and predict faults more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and specifically relates to a photovoltaic component fault prediction method and system based on deep learning. Background Art

[0002] The construction of large-scale photovoltaic power stations and the wide application of distributed photovoltaics have posed great challenges to the operation and maintenance of photovoltaic components, and effective fault prediction technologies are required to ensure the stable operation of the system.

[0003] Traditional fault diagnosis methods are mostly based on expert experience and physical models, and it is difficult to adapt to the complex and changeable operating environment of photovoltaic systems and the processing requirements of a large amount of data. Due to the differences in equipment, environment, and operating conditions of different photovoltaic power stations, when a fault prediction model trained based on a specific dataset is applied to other scenarios, there may be problems with poor generalization ability and it cannot effectively adapt to new operating environments and equipment. Deep learning has unique advantages in processing complex data patterns and non-linear relationships, can automatically learn the features and laws in the data, and does not require manual feature extraction, greatly improving the accuracy and efficiency of fault prediction. Summary of the Invention

[0004] To solve the above technical problems, a photovoltaic component fault prediction method and system based on deep learning are provided. This technical solution solves the problems of being difficult to adapt to the complex and changeable operating environment of photovoltaic systems and the processing requirements of a large amount of data, and the problem of poor generalization ability that occurs when a fault prediction model trained based on a specific dataset is applied to other scenarios and cannot effectively adapt to new operating environments and equipment.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A photovoltaic component fault prediction method based on deep learning, comprising:

[0007] Establish a battery parameter dataset through a photovoltaic system simulation model, and analyze the changes in photovoltaic characteristic curves under different fault types;

[0008] Based on the changes in photovoltaic characteristic curves, construct a fault prediction model based on Transformer, and perform correlation analysis on the extracted change parameters in combination with the attention mechanism;

[0009] Based on the correlation analysis of the change parameters, use a U-Net decoder, ReLU activation function, and Adam optimizer for parameter optimization and learning rate control;

[0010] Introduce the SoftMax function to classify the severity of faults.

[0011] Preferably, establishing a battery parameter dataset through the photovoltaic system simulation model and analyzing the changes in photovoltaic characteristic curves under different fault types specifically includes:

[0012] The core parameters for basic electrical parameter acquisition include open-circuit voltage, short-circuit current, maximum power point voltage and current; simulating data under different irradiances and temperatures through the simulation model;

[0013] Introducing the Python single-diode model to quantify the influence of series resistance and parallel resistance parameters on the output;

[0014] Labeling the fault type, fault location and severity, and generating a multi-dimensional dataset in combination with the simulation model;

[0015] Extracting time-frequency domain features through wavelet packet decomposition, and extracting infrared thermal imaging spatial features in combination with a deep learning model; quantifying the coupling effect between environmental parameters and fault severity;

[0016] Quantifying prediction uncertainty based on Monte Carlo Dropout, and augmenting small-sample fault data in combination with a generative adversarial network;

[0017] Measuring the IV / PV curve through the built experimental platform to verify the error rate of the simulation model.

[0018] Preferably, based on the changes in the photovoltaic characteristic curve, constructing a fault prediction model based on Transformer, and specifically including performing correlation analysis on the extracted change parameters in combination with the attention mechanism:

[0019] Collecting electrical parameters, mechanical signals, and environmental parameters to form multiple data sources, using wavelet packet decomposition to extract frequency domain energy entropy for time series data, analyzing the frequency components of the signals, and using a convolutional network to extract local features for spatial data to identify local spatial structure information;

[0020] Normalizing numerical parameters and performing One-Hot encoding on discrete state parameters to make them the input of the model; using the sliding window technique to divide time series data into multiple time windows to generate time series samples.

[0021] Preferably, based on the changes in the photovoltaic characteristic curve, constructing a fault prediction model based on Transformer, and specifically including performing correlation analysis on the extracted change parameters in combination with the attention mechanism:

[0022] Based on the time series samples, learnable position embeddings are used in the Transformer model to represent time information, so as to capture the characteristic positions of each moment in the time series;

[0023] Adopt a multi-head attention mechanism to calculate the correlation weights between different feature parameters, analyze the temporal variation correlations of the parameters, and capture the physical correlations between device components;

[0024] Visualize the attention weights through the LRP algorithm to identify key fault indication parameters, and automatically identify the parameters and signals most relevant to the fault through the attention mechanism;

[0025] Calculate the correlation coefficient between the main frequency of the vibration spectrum and the current harmonic components, construct a parameter correlation matrix, and explore the internal relationships between various fault-related physical quantities; Trace the abnormal conduction chain of parameters based on the attention weights, and according to the attention weights, trace the conduction chain of the fault to identify the problem parameters where the fault starts to spread and the problem factors causing the root cause of the fault;

[0026] Combine the sliding window statistics, and judge the fault triggering condition by setting a change rate threshold to trigger attention recalibration and adjust the model's focus on the current data.

[0027] Preferably, for the correlation analysis based on changing parameters, the parameter optimization and learning rate control using a U-Net decoder, ReLU activation function, and Adam optimizer specifically include:

[0028] U-Net decoder optimization, introducing residual connections in the decoding stage, and performing channel splicing on the high-resolution features in the encoder and the semantic features generated in the decoder;

[0029] Replace traditional convolutions with depthwise separable convolutions, perform per-channel convolutions first, and then pointwise convolutions to reduce the computational amount and the number of parameters;

[0030] Parallelly output the fault type classification and remaining life regression branches, which are used to identify the fault types occurring in the current system and predict the remaining service life of the system, sharing the underlying feature extraction network;

[0031] Fuse the outputs of different decoding layers through a feature pyramid network to help the network synthesize information from a multi-scale perspective and obtain a richer feature representation.

[0032] Preferably, for the correlation analysis based on changing parameters, the parameter optimization and learning rate control using a U-Net decoder, ReLU activation function, and Adam optimizer specifically include:

[0033] ReLU activation function optimization, using Leaky ReLU in the deep layers of the encoder to avoid the gradient vanishing problem, enhance the sparse feature expression ability, and maintain effective gradient flow;

[0034] Adopt Sigmoid at the last layer of the U-Net decoder to constrain the output range and match the percentage numerical characteristics of life prediction;

[0035] Adam optimizer parameter settings, including the initial learning rate, first-order momentum, second-order momentum, and weight decay; set the initial learning rate of the Adam optimizer to control the magnitude of parameter updates at the beginning of training; adjust the learning rate of each parameter according to the historical information of the gradient by combining the first-order momentum and the second-order momentum; add L2 regularization during training to avoid overfitting of the model;

[0036] The learning rate is dynamically controlled using the exponential decay method, and the learning rate is gradually reduced as the training process progresses; introduce a warm restart mechanism to periodically reset the learning rate to a larger value during training and then decay it;

[0037] According to the importance of different parameters, the learning rate is set in layers to optimize the training process, enabling the network to adjust different parameters at different stages.

[0038] Preferably, the introduction of the SoftMax function to classify the severity of faults specifically includes:

[0039] The SoftMax function maps the output of the hidden layer to a four-dimensional probability vector, representing the probability distribution of different fault categories; adopt a weighted cross-entropy loss function, and in the case of unbalanced fault categories, reduce the impact of class imbalance on the training results by assigning different weights to different fault categories;

[0040] Introduce a channel attention module after the hidden layer to calculate the weights of feature channels, determine the importance of feature channels for the final prediction, and automatically adjust the network's attention to different features;

[0041] Integrate Monte Carlo Dropout, calculate the prediction confidence interval through multiple forward propagations, identify low-confidence samples and trigger manual review;

[0042] Deploy a sliding window mechanism to capture the distribution shift of new data. When the distribution shift of new data exceeds a predefined threshold, initiate model fine-tuning and update the parameters of the SoftMax layer;

[0043] Use a confusion matrix to analyze the classification effect, calculate the precision, recall, and F1-Score, focus on the recall of severe faults to identify serious faults; draw an ROC curve to verify the AUC values of different severity levels;

[0044] Based on the classification results of the model, when the predicted fault severity reaches a certain probability threshold, trigger the corresponding hierarchical warning mechanism. When the probability of severe faults reaches the predefined threshold, trigger an emergency stop instruction, and when the probability of moderate faults reaches the predefined threshold, trigger an operation and maintenance work order.

[0045] Furthermore, a photovoltaic module fault prediction system based on deep learning is used to implement a photovoltaic module fault prediction method based on deep learning as described above, including:

[0046] Multi-source heterogeneous data acquisition module: Integrate electrical parameter sensors, infrared thermal imagers, vibration accelerometers, and meteorological monitoring units to synchronously obtain the three-dimensional spatio-temporal data stream of the working state of photovoltaic modules;

[0047] Dynamic feature extraction module: Adopt an improved ResNet-18 network to extract the spatial distribution features of thermal images, combine the wavelet packet decomposition algorithm to extract the frequency-domain energy entropy of current and voltage signals, and establish a cross-modal feature fusion pyramid;

[0048] Spatio-temporal dual-stream Transformer prediction model: Include a time attention sub-module: Analyze the time-varying correlation of the electrical parameter sequence and calculate the time-step weight matrix; a spatial attention sub-module: Model the physical topological relationship between components through a graph convolutional network; a federated learning optimizer: Adopt a differential privacy mechanism to achieve secure sharing of model parameters across power stations;

[0049] Edge computing deployment module: Integrate the TensorRT acceleration engine and the dynamic pruning algorithm to achieve quantization compression of model parameters;

[0050] Online incremental learning module: Update model parameters in real time based on a sliding time window, and trigger cloud joint training when the prediction confidence is lower than the threshold.

[0051] Optionally, the multi-source heterogeneous data acquisition module specifically includes:

[0052] Meteorological data acquisition unit: The light intensity is collected using a light sensor, the temperature and humidity are obtained through a temperature and humidity sensor, and the wind speed and direction are collected using an anemometer;

[0053] Electrical data acquisition unit: The voltage and current are measured through a voltage sensor and a current sensor;

[0054] Image data acquisition unit: Visible light images use a camera to collect visible light images of photovoltaic modules to detect physical damage on the surface of the modules, and infrared thermal images detect the temperature distribution on the surface of the modules through infrared thermal imaging technology;

[0055] Adopt Beidou dual-mode time synchronization to ensure the time synchronization error of multi-source data; Record the longitude, latitude, and installation height of each module through the GIS coordinate system, generate timestamped data packets at a per-second reference interval; Use the Kalman filter algorithm to eliminate sensor noise and achieve feature compression through PCA dimensionality reduction.

[0056] Optionally, the spatio-temporal dual-stream Transformer prediction model specifically includes:

[0057] Temporal Attention Sub-module: Adopt the multi-head self-attention mechanism to calculate the time-step weight matrix of the electrical parameter sequence, introduce the sliding window mechanism to capture periodic features; use sine function encoding and learnable position embedding to enhance the model's perception ability of the time order; monitor the convergence of attention weights through the recurrent neural network;

[0058] Spatial Attention Sub-module: Construct the photovoltaic array graph structure, model the physical topological relationship between components through the graph convolutional network, calculate the thermal radiation influence factor between components, and generate the spatial attention heat map; splice the shallow local features and the deep global features in channels to improve the detection accuracy of small-area faults;

[0059] Federated Learning Optimizer: Adopt differential privacy protection, add Gaussian noise in the parameter update stage to meet differential privacy, limit the amplitude of model parameter updates, and prevent privacy leakage; apply the Paillier homomorphic encryption algorithm to achieve the secure upload of the power station's local model parameters, automatically adjust the aggregation weights according to the power station data volume, and achieve the secure sharing of cross-station model parameters;

[0060] Model Cooperative Working Sub-module: Spatiotemporal feature interaction, set the cross-attention gate, and fuse the time flow output and the space flow output through the gating mechanism;

[0061] In the spatiotemporal dual-stream Transformer prediction model, the temporal attention sub-module includes:

[0062]

[0063] where Q is the query vector, representing the query information in the input sequence; K is the key vector, representing the key information in the input sequence; V is the value vector, representing the value information in the input sequence; d k is the dimension of the key vector, used to scale the dot product to avoid gradient vanishing; K T is the transpose of the key vector matrix, used to calculate the dot product between the query vector and the key vector to obtain their similarity; softmax is the activation function, used to convert the result of the dot product into a probability distribution;

[0064] The spatial station attention sub-module includes:

[0065]

[0066] where Ms(F) is the spatial attention weight matrix; σ is the activation function, the Sigmoid function; is a 7×7 convolution kernel; is to splice the global average pooling feature and the maximum pooling feature together;

[0067] The federated learning optimizer includes:

[0068]

[0069] In the formula, is the gradient after adding noise, Gradient is the original gradient, represents Gaussian noise with a mean of 0 and a variance of .

[0070] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0071] The present invention proposes to establish a battery parameter dataset through a photovoltaic system simulation model, which can simulate the operating states of photovoltaic modules under different fault types and generate rich training data. This enables the model to have better adaptability and generalization ability when facing diverse fault situations of actual photovoltaic modules, and can more accurately identify and predict faults; this method does not completely rely on the understanding of the precise physical characteristics of photovoltaic modules, but learns fault features in a data-driven manner, which enables it to better adapt to complex actual operating environments, such as different lighting, temperature, humidity and other conditions, effectively solving the problem of the decrease in diagnostic accuracy caused by the mismatch between the model assumptions and the actual situation in the traditional method in a complex environment; introducing the SoftMax function to classify the severity of faults can convert multiple numerical values output by the model into a probability distribution, thereby realizing accurate classification of different fault types. It can not only predict whether a fault occurs, but also further determine the severity of the fault, providing more detailed decision-making support information for maintenance personnel, helping to take corresponding maintenance measures in a timely manner and reducing the impact of faults on the operation of the photovoltaic system. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 is a flowchart of a photovoltaic module fault prediction method based on deep learning;

[0073] Figure 2 is an internal framework diagram of a photovoltaic module fault prediction system based on deep learning. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0074] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments in the following description are only examples, and those skilled in the art can think of other obvious variations.

[0075] Referring to Figure 1 shown, a photovoltaic module fault prediction method based on deep learning includes:

[0076] Establishing a battery parameter dataset through a photovoltaic system simulation model and analyzing the changes in photovoltaic characteristic curves under different fault types;

[0077] Based on the changes in the photovoltaic characteristic curve, a fault prediction model based on Transformer is constructed, and the attention mechanism is combined to conduct correlation analysis on the extracted change parameters;

[0078] Based on the correlation analysis of the change parameters, a U-Net decoder, ReLU activation function, and Adam optimizer are used for parameter optimization and learning rate control;

[0079] The SoftMax function is introduced to classify the severity of the fault.

[0080] It should be noted that the analysis of the changes in the photovoltaic characteristic curve under different fault types includes:

[0081] Feature extraction is to extract key features from the operation curves of photovoltaic modules (such as power-time curve, current-voltage curve, etc.). For example, calculate the slope, peak value, valley value, area, etc. of the curve. The slope of the power-time curve in the morning reflects the power rising speed of the module when the light intensity increases.

[0082] Curve fitting is to fit the operation curve using a suitable mathematical model, such as polynomial fitting, exponential fitting, etc. By fitting the curve, the change trend of the data can be better captured to judge whether the operation state of the module is normal.

[0083] Anomaly detection is to establish a reference model based on the normal operation curve and compare the real-time curve with the reference model. If the deviation between the real-time curve and the reference model exceeds a certain threshold, it is considered that the module may have a fault. For example, if the real-time value of the power curve is lower than 80% of the reference curve, a fault warning is issued.

[0084] The construction of the Transformer model includes:

[0085] The Transformer model is mainly composed of an encoder and a decoder. The encoder is responsible for feature extraction and encoding of the input sequence, and the decoder makes predictions based on the output of the encoder. In the fault prediction of photovoltaic modules, usually only the encoder part is used. The model contains multiple multi-head self-attention layers and feed-forward neural network layers, and captures the long-range dependencies in the data through the self-attention mechanism;

[0086] The Transformer model is trained using the training set, and a suitable loss function (such as cross-entropy loss function) and optimizer (such as Adam optimizer) are selected. During the training process, by continuously adjusting the parameters of the model, the value of the loss function is minimized to improve the prediction performance of the model.

[0087] Refer to Figure 2 As shown, a photovoltaic module fault prediction system based on deep learning includes:

[0088] Multi-source heterogeneous data acquisition module: Integrate electrical parameter sensors, infrared thermal imagers, vibration accelerometers, and meteorological monitoring units to synchronously obtain three-dimensional spatio-temporal data streams of the working status of photovoltaic modules;

[0089] Dynamic feature extraction module: Adopt an improved ResNet-18 network to extract the spatial distribution features of thermal imaging, combine the wavelet packet decomposition algorithm to extract the frequency-domain energy entropy of current and voltage signals, and establish a cross-modal feature fusion pyramid;

[0090] Spatio-temporal dual-stream Transformer prediction model: Include a time attention sub-module: Analyze the time-varying correlation of electrical parameter sequences and calculate the time-step weight matrix; A spatial attention sub-module: Model the physical topological relationship between components through a graph convolutional network; A federated learning optimizer: Use the differential privacy mechanism to achieve secure sharing of model parameters across power stations;

[0091] Edge computing deployment module: Integrate the TensorRT acceleration engine and the dynamic pruning algorithm to achieve quantization compression of model parameters;

[0092] Online incremental learning module: Based on a sliding time window, update model parameters in real time, and trigger cloud joint training when the prediction confidence is lower than the threshold.

[0093] In the dynamic feature extraction module, the combined wavelet packet decomposition algorithm includes:

[0094]

[0095] In the formula, is the wavelet packet coefficient of the j-th layer, is the high-pass filter coefficient, is the low-frequency coefficient of the (j - 1)-th layer.

[0096] In the online incremental learning module, the sliding time window formula includes:

[0097]

[0098] In the formula, is the updated model parameter, is the current model parameter, is the learning rate, which controls the step size of parameter update, is the gradient of the loss function on the new data.

[0099] The spatio-temporal dual-stream Transformer prediction model specifically includes:

[0100] Temporal Attention Sub-module: Adopt the multi-head self-attention mechanism to calculate the time-step weight matrix of the electrical parameter sequence, introduce the sliding window mechanism to capture periodic features; use sine function encoding and learnable position embedding to enhance the model's perception ability of time order; monitor the convergence of attention weights through the recurrent neural network;

[0101] Spatial Attention Sub-module: Construct the photovoltaic array graph structure, model the physical topological relationship between components through the graph convolutional network, calculate the thermal radiation influence factor between components, and generate the spatial attention heat map; perform channel splicing on the shallow local features and deep global features to improve the detection accuracy of small-area faults;

[0102] Federated Learning Optimizer: Adopt differential privacy protection, add Gaussian noise in the parameter update stage to satisfy differential privacy, limit the amplitude of model parameter updates, and prevent privacy leakage; apply the Paillier homomorphic encryption algorithm to realize the secure upload of the local model parameters of the power station, automatically adjust the aggregation weight according to the power station data volume, and realize the secure sharing of cross-station model parameters;

[0103] Model Cooperative Working Sub-module: Spatiotemporal feature interaction, set the cross-attention gate, and fuse the time flow output and spatial flow output through the gating mechanism;

[0104] In the spatio-temporal dual-stream Transformer prediction model, the temporal attention sub-module includes:

[0105]

[0106] In the formula, Q is the query vector, representing the query information in the input sequence; K is the key vector, representing the key information in the input sequence; V is the value vector, representing the value information in the input sequence; is the dimension of the key vector, used to scale the dot product to avoid gradient vanishing; is the transpose of the key vector matrix, used to calculate the dot product between the query vector and the key vector to obtain their similarity; softmax is the activation function, used to convert the result of the dot product into a probability distribution;

[0107] The spatial station attention sub-module includes:

[0108]

[0109] In the formula, Ms(F) is the spatial attention weight matrix; σ is the activation function, the Sigmoid function; is a 7×7 convolutional kernel; is to splice the global average pooling feature and the maximum pooling feature together;

[0110] The federated learning optimizer includes:

[0111]

[0112] In the formula, is the gradient after adding noise, and Gradient is the original gradient. represents Gaussian noise with a mean of 0 and a variance of .

[0113] It should be noted that the spatio-temporal dual-stream Transformer prediction model includes:

[0114] Temporal attention sub-module: Adopting the multi-head self-attention mechanism, it can simultaneously focus on the correlations between different time steps in the electrical parameter sequence. By calculating the time-step weight matrix, the model can determine which time-point data is more important for fault prediction, thereby highlighting the features of key time steps and improving the model's ability to model time-series data;

[0115] Introducing the sliding window mechanism enables the model to locally focus on the continuous window data in the time series, thereby capturing the periodic features in the data;

[0116] Encoding through the sine function provides the model with prior information on time order, and at the same time introducing learnable position embeddings enables the model to better adapt to specific fault prediction tasks;

[0117] Utilizing the recursive characteristics of the recurrent neural network to continuously monitor and update the attention weights, ensuring that the attention weights can converge stably during the training process, which helps to improve the training stability and prediction accuracy of the model;

[0118] Spatial attention sub-module: Photovoltaic array graph structure and graph convolutional network: Constructing the photovoltaic array graph structure, representing the physical connection relationships between photovoltaic components as nodes and edges in the graph, and modeling this physical topological relationship through the graph convolutional network, the model can learn the ways in which components influence each other, thereby more accurately predicting the spatial distribution and propagation of faults;

[0119] Federated learning optimizer: Adding Gaussian noise in the parameter update stage to meet the requirements of differential privacy, restricting the amplitude of model parameter updates, which effectively prevents the leakage of privacy information, protects the confidentiality of data of each photovoltaic power station, and makes cross-station data sharing more secure and reliable.

[0120] Adopting the Paillier homomorphic encryption algorithm to achieve the secure upload of local model parameters of the power station. This algorithm allows specific calculation operations to be performed on encrypted data without decrypting the data, thereby ensuring the security of model parameters during transmission and aggregation.

[0121] Model collaborative working sub-module: By setting cross-attention gates, the fusion of the time flow output and the spatial flow output is achieved. The cross-attention mechanism allows the features in the time flow and the spatial flow to pay attention to and complement each other, captures the correlation between the time features and the spatial features, and further improves the model's comprehensive judgment ability for faults.

[0122] Embodiment:

[0123] 1. Data preparation: Establish a photovoltaic system simulation model: Use professional photovoltaic system simulation software (such as PVsyst) to establish a simulation model of photovoltaic modules, and simulate the output characteristics of photovoltaic modules under different environmental conditions (such as light intensity, temperature, etc.) and different fault types (such as hot spots, cracks, aging, etc.).

[0124] 2. Generate a battery parameter dataset: Generate the output data of photovoltaic modules under various fault types and normal states through the simulation model, including parameters such as voltage, current, and power, and construct a battery parameter dataset.

[0125] 3. Transformer model construction:

[0126] Input layer: Take the output data of photovoltaic modules (such as voltage, current, power, etc.) as input and perform normalization processing;

[0127] Encoder: Adopt multiple layers of Transformer encoders to extract the feature representations of the input data. Each layer of the encoder includes a multi-head self-attention mechanism and a feed-forward neural network;

[0128] Decoder: Adopt a U-Net decoder to upsample and fuse the features extracted by the encoder to enhance the model's ability to capture local features;

[0129] Output layer: Output the fault prediction results, including the fault type and severity;

[0130] Attention mechanism: Introduce a multi-head self-attention mechanism in the encoder to perform correlation analysis on the extracted features and highlight the key features related to faults.

[0131] 4. Model training:

[0132] Loss function: Adopt a cross-entropy loss function to measure the difference between the model prediction results and the true labels;

[0133] Optimizer: Use the Adam optimizer to optimize the parameters, set the initial learning rate to 0.001, and adopt a learning rate decay strategy to gradually reduce the learning rate as the number of training rounds increases;

[0134] Activation function: Use the ReLU activation function in the middle layer of the model to accelerate the convergence of the model;

[0135] SoftMax function: Introduce the SoftMax function in the output layer to convert the output of the model into a probability distribution for the classification of fault severity.

[0136] 5. Model evaluation:

[0137] Evaluation metrics: Use metrics such as accuracy, recall, and F1-score to evaluate the performance of the model;

[0138] Cross-validation: Adopt the k-fold cross-validation method to train and test the model multiple times to ensure the stability and generalization ability of the model.

[0139] 6. Result application:

[0140] Fault prediction: Apply the trained model to the fault prediction of actual photovoltaic modules, monitor the output characteristics of photovoltaic modules in real time, and predict the fault type and severity;

[0141] Maintenance suggestions: Provide maintenance suggestions for the photovoltaic system based on the prediction results, such as cleaning the surface of photovoltaic modules in time, replacing aging components, etc., to improve the operation efficiency and reliability of the photovoltaic system.

[0142] In summary, the advantages of the present invention are as follows: By establishing a battery parameter dataset through a photovoltaic system simulation model, it is possible to simulate the operating states of photovoltaic modules under different fault types and generate rich training data; By learning fault features in a data-driven manner, it can better adapt to complex actual operating environments, effectively solving the problem of reduced diagnostic accuracy caused by the mismatch between model assumptions and actual situations in traditional methods in complex environments; Introducing the SoftMax function to classify fault severity can convert multiple numerical values output by the model into a probability distribution, thereby achieving accurate classification of different fault types. This can not only predict whether a fault occurs but also further determine the severity of the fault, providing more detailed decision-making support information for operation and maintenance personnel, helping to take corresponding maintenance measures in a timely manner, and reducing the impact of faults on the operation of the photovoltaic system.

[0143] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification is only the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A photovoltaic component fault prediction method based on deep learning, characterized in that, Including: Establish a battery parameter dataset through a photovoltaic system simulation model, and analyze the changes in photovoltaic characteristic curves under different fault types; Based on the changes in photovoltaic characteristic curves, construct a fault prediction model based on Transformer, and conduct correlation analysis on the extracted change parameters by combining the attention mechanism; Based on the correlation analysis of the change parameters, use a U-Net decoder, ReLU activation function, and Adam optimizer for parameter optimization and learning rate control; Introduce the SoftMax function to classify the severity of faults.

2. The photovoltaic module fault prediction method based on deep learning according to claim 1, wherein, The establishment of a battery parameter dataset through a photovoltaic system simulation model and the analysis of the changes in photovoltaic characteristic curves under different fault types specifically include: The core parameters for basic electrical parameter acquisition include open-circuit voltage, short-circuit current, maximum power point voltage and current; simulate data under different irradiances and temperatures through the simulation model; Introduce the Python single-diode model to quantify the influence of series resistance and parallel resistance parameters on the output; Label the fault type, fault location and severity, and generate a multi-dimensional dataset in combination with the simulation model; Extract time-frequency domain features through wavelet packet decomposition, and extract infrared thermal imaging spatial features through a deep learning model; quantify the coupling effect between environmental parameters and fault degree; Quantify prediction uncertainty based on Monte Carlo Dropout, and expand small-sample fault data by combining generative adversarial networks; Measure the IV / PV curve through the built experimental platform to verify the error rate of the simulation model.

3. The photovoltaic module fault prediction method based on deep learning according to claim 2, wherein, The construction of a fault prediction model based on Transformer based on the changes in photovoltaic characteristic curves and the correlation analysis of the extracted change parameters by combining the attention mechanism specifically include: Collect electrical parameters, mechanical signals, and environmental parameters to form multiple data sources. Use wavelet packet decomposition to extract frequency domain energy entropy from time series data, analyze the frequency components of the signals, and extract local features from spatial data through a convolutional network to identify local spatial structure information; Normalize numerical parameters and perform One-Hot encoding on discrete state parameters to make them the input of the model; use the sliding window technique to divide time series data into multiple time windows to generate time series samples.

4. A photovoltaic component fault prediction method based on deep learning according to claim 3, characterized in that The construction of a fault prediction model based on Transformer based on the changes in photovoltaic characteristic curves and the correlation analysis of the extracted change parameters by combining the attention mechanism specifically include: Based on the time series samples, learnable position embeddings are used in the Transformer model to represent time information, so as to capture the characteristic positions of each moment in the time series; Adopt the multi-head attention mechanism to calculate the correlation weights between different characteristic parameters, analyze the change correlation of parameters over time, and capture the physical correlation between equipment components; Visualize the attention weights through the LRP algorithm to identify key fault indication parameters, and automatically identify the parameters and signals most relevant to the fault through the attention mechanism; Calculate the correlation coefficient between the main frequency of the vibration spectrum and the current harmonic components, construct a parameter correlation matrix, and explore the internal relationships between the physical quantities related to each fault; based on the attention weights, trace the abnormal conduction chain of parameters, and according to the attention weights, trace the conduction chain of faults, identify the problem parameters where the fault starts to spread, and the problem factors causing the root cause of the fault; Combined with the sliding window statistic, judge the fault triggering condition by setting the change rate threshold, trigger the attention recalibration, and adjust the focus of the model on the current data.

5. A photovoltaic module fault prediction method based on deep learning according to claim 4, characterized in that, The correlation analysis based on the changing parameters, using the U-Net decoder, ReLU activation function and Adam optimizer for parameter optimization and learning rate control specifically includes: U-Net decoder optimization, introducing residual connections in the decoding stage, and performing channel splicing on the high-resolution features in the encoder and the semantic features generated in the decoder; Replace the traditional convolution with depthwise separable convolution, perform per-channel convolution and then pointwise convolution to reduce the amount of computation and the number of parameters; Parallelly output the fault type classification and remaining life regression branches, which are used to identify the fault types occurring in the current system and predict the remaining service life of the system, and share the underlying feature extraction network; Fuse the outputs of different decoding layers through the feature pyramid network to help the network synthesize information from a multi-scale perspective and obtain a richer feature representation.

6. The photovoltaic module fault prediction method based on deep learning according to claim 5, characterized in that, The correlation analysis based on the changing parameters, using the U-Net decoder, ReLU activation function and Adam optimizer for parameter optimization and learning rate control specifically includes: ReLU activation function optimization, using Leaky ReLU in the deep layer of the encoder to avoid the gradient vanishing problem, improve the ability to express sparse features, and maintain effective gradient flow; Adopt Sigmoid at the last layer of the U-Net decoder to constrain the output range and match the percentage numerical characteristics of life prediction; Adam optimizer parameter settings, including the initial learning rate, first-order momentum, second-order momentum, weight decay; set the initial learning rate of the Adam optimizer to control the amplitude of parameter update at the beginning of training; by combining the first-order momentum and the second-order momentum, adjust the learning rate of each parameter according to the historical information of the gradient; add L2 regularization during training to avoid overfitting of the model; The dynamic control of the learning rate adopts the exponential decay method, and gradually reduces the learning rate as the training process progresses; introduce the warm restart mechanism, and periodically reset the learning rate to a larger value during training and then decay it; Set the learning rate hierarchically according to the importance of different parameters to optimize the training process, so that the network adjusts different parameters at different stages.

7. The photovoltaic module fault prediction method based on deep learning according to claim 6, wherein, The introduction of the SoftMax function for classifying the fault severity specifically includes: The SoftMax function maps the output of the hidden layer to a four-dimensional probability vector, representing the probability distribution of different fault categories; adopt the weighted cross-entropy loss function, and in the case of unbalanced fault categories, reduce the impact of class imbalance on the training results by assigning different weights to different fault categories; After the hidden layer, a channel attention module is introduced to calculate the weights of feature channels, determine the importance of feature channels for the final prediction, and automatically adjust the network's attention to different features; Integrate Monte Carlo Dropout to calculate the prediction confidence interval through multiple forward propagations, identify low-confidence samples, and trigger manual review; Deploy a sliding window mechanism to capture the distribution shift of new data. When the distribution shift of new data exceeds a predefined threshold, initiate model fine-tuning and update the parameters of the SoftMax layer; Use a confusion matrix to analyze the classification effect, calculate precision, recall, and F1-Score, with a focus on the recall rate of severe faults to identify serious faults; plot the ROC curve and verify the AUC values for different severity levels; Based on the classification results of the model, when the predicted fault severity reaches a certain probability threshold, trigger the corresponding hierarchical warning mechanism. When the probability of a severe fault reaches the predefined threshold, trigger an emergency shutdown command, and when the probability of a moderate fault reaches the predefined threshold, trigger an operation and maintenance work order.

8. A photovoltaic module fault prediction system based on deep learning, characterized in that, For implementing a photovoltaic module fault prediction method based on deep learning as described in claims 1-7, including: Multi-source heterogeneous data acquisition module: Integrate electrical parameter sensors, infrared thermal imagers, vibration accelerometers, and meteorological monitoring units to synchronously obtain a three-dimensional spatio-temporal data stream of the working state of photovoltaic modules; Dynamic feature extraction module: Use an improved ResNet-18 network to extract the spatial distribution features of thermal images, combine the wavelet packet decomposition algorithm to extract the frequency-domain energy entropy of current and voltage signals, and establish a cross-modal feature fusion pyramid; Spatio-temporal dual-stream Transformer prediction model: Includes a time attention sub-module: Analyze the time-varying correlation of electrical parameter sequences and calculate the time-step weight matrix; a spatial attention sub-module: Model the physical topological relationship between components through a graph convolutional network; a federated learning optimizer: Use differential privacy mechanism to achieve secure sharing of model parameters across power stations; Edge computing deployment module: Integrate the TensorRT acceleration engine and dynamic pruning algorithm to achieve quantization compression of model parameters; Online incremental learning module: Update model parameters in real-time based on a sliding time window, and trigger cloud joint training when the prediction confidence is lower than the threshold.

9. The photovoltaic module fault prediction system based on deep learning according to claim 8, characterized in that The multi-source heterogeneous data acquisition module specifically includes: Meteorological data acquisition unit: The light intensity is collected using a light sensor, the temperature and humidity are obtained through a temperature and humidity sensor, and the wind speed and direction are collected using an anemometer; Electrical data acquisition unit: The voltage and current are measured through a voltage sensor and a current sensor; Image data acquisition unit: Visible light images of photovoltaic modules are collected using a camera to detect physical damage on the surface of the modules, and infrared thermal images are used to detect the temperature distribution on the surface of the modules through infrared thermal imaging technology; Use Beidou dual-mode time synchronization to ensure the time synchronization error of multi-source data; record the longitude, latitude, and installation height of each component through the GIS coordinate system, generate timestamped data packets at a per-second reference interval; use the Kalman filter algorithm to eliminate sensor noise and achieve feature compression through PCA dimensionality reduction.

10. The photovoltaic module fault prediction system based on deep learning according to claim 9, characterized in that, The spatio-temporal dual-stream Transformer prediction model specifically includes: Temporal Attention Sub-module: Adopt the multi-head self-attention mechanism to calculate the time-step weight matrix of the electrical parameter sequence, introduce the sliding window mechanism to capture periodic features; use sine function encoding and learnable position embedding to enhance the model's perception ability of time order; monitor the convergence of attention weights through the recurrent neural network; Spatial Attention Sub-module: Construct the photovoltaic array graph structure, model the physical topological relationship between components through the graph convolutional network, calculate the thermal radiation influence factor between components, and generate a spatial attention heat map; perform channel splicing on shallow local features and deep global features to improve the detection accuracy of small-area faults; Federated Learning Optimizer: Adopt differential privacy protection, add Gaussian noise in the parameter update stage to satisfy differential privacy, limit the update amplitude of model parameters, and prevent privacy leakage; apply the Paillier homomorphic encryption algorithm to realize the secure upload of the local model parameters of the power station, automatically adjust the aggregation weight according to the power station data volume, and realize the secure sharing of cross-station model parameters; Model Collaborative Working Sub-module: Spatiotemporal feature interaction, set cross-attention gates, and fuse the time-stream output and the space-stream output through the gating mechanism; The temporal attention sub-module in the spatiotemporal dual-stream Transformer prediction model includes: Wherein, Q is a query vector, representing the query information in the input sequence; K is a key vector, representing the key information in the input sequence; V is a value vector, representing the value information in the input sequence; d k is the dimension of the key vector, used to scale the dot product to avoid vanishing gradients; K T is the transpose of the key vector matrix, used to calculate the dot product between the query vector and the key vector, so as to obtain the similarity between them; softmax is an activation function, used to convert the result of the dot product into a probability distribution; The space station attention sub-module includes: where $M_s(F)$ is the spatial attention weight matrix; is the activation function, the Sigmoid function; is a 7×7 convolutional kernel; is to concatenate the global average pooling feature and the maximum pooling feature together; The federated learning optimizer includes: In the formula, is the gradient after adding noise, and Gradient is the original gradient. represents Gaussian noise with a mean of 0 and a variance of .

Images