Photovoltaic panel wind damage resistance prediction and risk assessment method based on self-supervised learning

By employing a self-supervised learning dual-task model and incremental learning mechanism, the accuracy and real-time performance issues of wind damage prediction for photovoltaic panels were resolved, achieving efficient and accurate wind damage risk assessment and improving the safety and stability of photovoltaic systems.

CN121434935APending Publication Date: 2026-01-30ALPHA ESS CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202511610110.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-11-05
Publication Date
2026-01-30

AI Technical Summary

Technical Problem

Existing methods for predicting wind damage to photovoltaic panels are insufficient in terms of accuracy and real-time performance, making it difficult to meet the stability and safety requirements of photovoltaic systems. Traditional physical modeling methods lack precision, machine learning methods rely on labeled data and have poor generalization ability, and CNN methods are weak in modeling temporal and spatial characteristics, making it impossible to comprehensively assess wind damage risks.

Method used

A self-supervised learning-based approach is adopted to construct a dual-task model, including a shared layer and an independent task layer. Multi-dimensional data is processed through a multi-layer Transformer encoder, and training is performed using a joint loss function and an optimizer. This enables wind damage risk assessment and early warning for photovoltaic panels, and supports incremental learning and transfer learning to adapt to different environments.

Benefits of technology

It significantly improves the accuracy and real-time performance of wind damage prediction for photovoltaic panels, enhances the model's adaptability and generalization ability, and can maintain high accuracy in wind damage prediction and risk assessment in unseen environments, providing customized wind disaster prediction services.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121434935A_ABST
    Figure CN121434935A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of photovoltaic system safety, in particular to a photovoltaic panel wind damage resistance prediction and risk assessment method based on self-supervised learning, and the method comprises the steps: firstly collecting multi-dimensional data such as wind speed, wind direction, temperature and humidity and photovoltaic panel material, and carrying out the cleaning, standardization, PCA dimension reduction and data set division preprocessing; a dual-task model containing a shared Transform layer and an independent wind damage resistance prediction and wind disaster risk assessment layer is constructed, and a loss function and Adam are combined for optimization training; according to the photovoltaic panel wind damage resistance prediction and risk assessment method based on self-supervised learning, the problems that in the prior art, precision, real-time performance and generalization are insufficient are solved, wind damage resistance prediction MSE0.035 in an experiment is achieved, the risk assessment accuracy rate is 92.5%, and the risk assessment accuracy rate is 92.5%. And the safety and the stability of the photovoltaic system are obviously improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of photovoltaic system safety technology, specifically a method for predicting and assessing the wind damage resistance of photovoltaic panels based on self-supervised learning. Background Technology

[0002] With the rapid development of the photovoltaic industry, photovoltaic power generation systems are widely used in different climates and environments. The wind resistance of photovoltaic panels is a key factor in the stability and safety of photovoltaic systems. Especially under extreme weather conditions such as typhoons, the damage caused by wind to photovoltaic panels will seriously affect the performance of photovoltaic systems. Existing methods for predicting wind damage to photovoltaic panels can be broadly categorized as follows: First, physical modeling-based methods. Traditional methods typically rely on physical and mechanical models, such as wind speed, wind direction, and photovoltaic panel angle, to predict wind damage by simulating the interaction between wind pressure and the photovoltaic panel structure. While these methods can provide a theoretical basis for prediction, their accuracy in actual predictions is poor, and they cannot perform real-time wind damage assessment. Second, machine learning-based methods. In recent years, with the development of machine learning technology, some studies have adopted supervised learning methods for predicting wind damage to photovoltaic panels. By training models with a large amount of historical data, machine learning methods can make predictions based on existing data. These methods typically rely on a large amount of labeled data, and more importantly, they lack flexibility and have poor generalization ability when facing new environments or unseen data. Third, wind damage detection methods based on convolutional neural networks (CNNs) and deep learning. These methods utilize image processing and CNNs for wind damage image detection. Although they can effectively identify the location and extent of damage in a localized area, this method is mainly used for image data and has weak modeling capabilities for the temporal and spatial characteristics of wind damage prediction, lacking the ability to comprehensively assess wind damage risk. While physics-based wind damage prediction methods can theoretically provide some predictive basis, these methods rely on complex physical and mechanical models, such as wind speed, wind direction, and photovoltaic panel angles, and make predictions by simulating the interaction between wind pressure and the photovoltaic panel structure. However, in practical applications, these methods often suffer from poor accuracy due to a lack of precise model parameters and an inability to fully consider complex environmental factors, and they cannot perform real-time wind damage assessment. Machine learning-based wind damage prediction methods employ supervised learning algorithms to train models using large amounts of historical data. While these methods can learn certain patterns from historical data, they typically rely on large amounts of labeled data, and the labeling process... The current methods for wind damage detection are cumbersome and time-consuming, with poor generalization ability and difficulty in adapting to environmental changes. While CNN and deep learning-based wind damage detection methods can effectively identify the specific location and extent of wind damage in local areas, these methods mainly rely on image data and have weak modeling capabilities for the temporal and spatial characteristics of wind damage. They cannot comprehensively assess the overall wind damage risk of photovoltaic panels. These shortcomings of existing technologies make it impossible to achieve efficient and accurate wind damage prediction and risk assessment for photovoltaic panels, which is insufficient to meet the stability and safety requirements of photovoltaic systems. Therefore, in view of the above situation, it is urgent to develop a self-supervised learning-based method for wind damage prediction and risk assessment of photovoltaic panels to overcome the shortcomings in current practical applications. Summary of the Invention

[0003] The purpose of this invention is to provide a method for predicting and assessing the wind damage resistance of photovoltaic panels based on self-supervised learning, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting and assessing wind damage to photovoltaic panels based on self-supervised learning includes the following steps: S1: Collect multi-dimensional data related to photovoltaic panels and preprocess the multi-dimensional data; S2: Construct a dual-task model, which includes a shared layer and two independent task layers. The shared layer is used to process the preprocessed multi-dimensional data and generate high-dimensional feature representations. The two independent task layers are the wind damage prediction task layer and the wind disaster risk assessment task layer, respectively. S3: The dual-task model is trained and optimized using a joint loss function and optimizer; S4: The real-time collected multi-dimensional data related to the photovoltaic panel is preprocessed and then input into the trained and optimized dual-task model to output the wind damage risk score of the photovoltaic panel, and to perform early warning operation based on the wind damage risk score. S5: Supports parameter updates for the trained and optimized dual-task model, and adapts to the needs of cross-regional photovoltaic panel wind damage prediction and risk assessment.

[0005] As a further aspect of the present invention: in step S1, the multi-dimensional data includes environmental data, photovoltaic panel data, and historical data; The environmental data includes at least wind speed, wind direction, temperature, humidity, and air pressure; The photovoltaic panel data includes at least the photovoltaic panel material and installation angle; The historical data includes at least historical damage records of photovoltaic panels and historical wind disaster records.

[0006] As a further aspect of the present invention: in step S1, the preprocessing includes data cleaning, missing value imputation, standardization, feature selection and dimensionality reduction, and dataset partitioning; The data cleaning includes detecting and removing outliers in multi-dimensional data; The feature selection and dimensionality reduction were performed using principal component analysis (PCA). The dataset partitioning divides multi-dimensional data into training and testing sets, and K-fold cross-validation is used to evaluate model performance during model training.

[0007] As a further aspect of the present invention: in step S2, the shared layer adopts a multi-layer Transformer encoder architecture; The Transformer encoder processes the preprocessed multi-dimensional data through a self-attention mechanism to capture temporal dependencies in the data, thereby generating the high-dimensional feature representation.

[0008] As a further aspect of the present invention: In step S2, the wind damage prediction task layer is constructed based on a feedforward neural network, taking the high-dimensional feature representation as input, and outputting the damage probability of the photovoltaic panel through regression calculation; The wind disaster risk assessment task layer takes the high-dimensional feature representation as input and outputs the wind damage risk score of the photovoltaic panel through regression calculation.

[0009] As a further aspect of the present invention: in step S3, the joint loss function is a weighted sum of the loss from the wind damage prediction task and the loss from the wind disaster risk assessment task; The loss from both the wind damage prediction task and the wind disaster risk assessment task uses the mean squared error (MSE) loss function; the optimizer used is the Adam optimizer.

[0010] As a further aspect of the present invention: the expression for the joint loss function is: in, To predict mission losses in response to wind damage, λ represents the loss from the wind disaster risk assessment task, and λ is a hyperparameter that adjusts the contribution ratio of the losses from the two tasks to the total loss.

[0011] As a further aspect of the present invention: in step S4, the warning operation includes: Set at least two risk scoring thresholds, compare the real-time output wind injury risk score with the risk scoring thresholds, and classify the risk into at least three risk levels: low risk, medium risk, and high risk. Execute corresponding operations for different risk levels: normal monitoring is performed at low risk levels, enhanced monitoring is performed at medium risk levels, and notifications are sent to relevant personnel and emergency response is initiated at high risk levels.

[0012] As a further aspect of the present invention: In step S5, the parameter updates of the dual-task model adopt an incremental learning method. When new wind disaster-related data is acquired, the existing parameters of the model are fine-tuned based on the new data to optimize the model's predictive capabilities.

[0013] As a further aspect of the present invention: In step S5, the adaptation to cross-regional requirements adopts a transfer learning approach. Based on the wind disaster characteristic data of the target area, the parameters of the trained and optimized dual-task model are fine-tuned to adapt the model to the scenario of wind damage prediction and risk assessment of photovoltaic panels in the target area.

[0014] Compared with the prior art, the beneficial effects of the present invention are: The self-supervised learning-based method for predicting and assessing the wind damage of photovoltaic panels provided by this invention can effectively solve the problems of low prediction accuracy, poor real-time performance and insufficient generalization ability in the existing technology, and significantly improve the safety and stability of photovoltaic systems. First, by introducing a self-supervised learning framework, wind damage patterns can be automatically learned from a large amount of unlabeled historical data, no longer relying on a large amount of labeled data. This not only improves data utilization efficiency but also enhances the adaptability and prediction accuracy of the model. Compared with traditional supervised learning methods, this method can still maintain high generalization ability in unseen environments. Secondly, by integrating and comprehensively analyzing various environmental data such as wind speed, wind direction, temperature, humidity, air pressure, photovoltaic panel material and installation angle, it can more accurately capture the influencing factors of wind disasters, realize real-time dynamic prediction, and provide more comprehensive and accurate wind damage prediction and risk assessment, making up for the shortcomings of existing technologies in dealing with sudden wind disasters. Furthermore, the model supports incremental learning. Whenever new wind disaster data arrives, the model can be fine-tuned based on the new data to optimize its predictive ability. At the same time, it adopts transfer learning technology, which can make the model parameters better adapt to the wind disaster characteristics of different geographical regions by fine-tuning the existing model, and provide customized wind disaster prediction services. In addition, the actual experimental results show that the method has the smallest mean squared error (MSE) of only 0.035 in the wind damage prediction task, which is better than the 0.089 of the physical model and the 0.045 of the SVM-based supervised learning method. In the wind disaster risk assessment task, the accuracy reaches 92.5%, which is significantly higher than the 78.3% of the physical model and the 85.2% of the SVM method, which fully verifies its significant advantages in prediction accuracy and risk assessment accuracy. Attached Figure Description

[0015] Figure 1 This is a flowchart of a method for predicting and assessing wind damage to photovoltaic panels based on self-supervised learning, as described in an embodiment of the present invention.

[0016] Figure 2 This is a structural diagram of the dual-task model for predicting and assessing wind damage to photovoltaic panels based on self-supervised learning, as described in an embodiment of the present invention.

[0017] Figure 3 This is an architecture diagram of the photovoltaic panel wind damage risk assessment and early warning system in an embodiment of the present invention. Detailed Implementation

[0018] 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.

[0019] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0020] Please see Figures 1-3 This invention provides a method for predicting and assessing the wind damage resistance of photovoltaic panels based on self-supervised learning. This method combines self-supervised learning, multimodal data fusion, and incremental learning mechanisms to accurately and in real-time predict the wind damage resistance and damage area of ​​photovoltaic panels under different environmental conditions, thereby improving the safety and stability of photovoltaic systems. Specifically, it includes the following steps: Step 1: Photovoltaic panel data acquisition and preprocessing; 1. Data Acquisition: Obtain wind speed ( ),wind direction( Temperature (T), humidity (H), air pressure (P), photovoltaic panel material (M), installation angle (α), and historical damage records of the photovoltaic panel ( The data comprehensively describes the damage to photovoltaic panels under different environmental conditions and wind speeds using multi-source data.

[0021] 2. Data Cleaning and Missing Value Imputation: First, outlier detection is performed to remove potentially erroneous or unusual data points. The Z-score method is used for outlier detection, calculating the Z-value of each data point and determining if it exceeds a threshold of 2. The formula is as follows: in, For a single data point, μ is the mean and σ is the standard deviation. If the value is not found, it is considered an outlier and should be removed.

[0022] Then, missing value imputation is performed. Missing value imputation uses KNN interpolation (calculates the similarity between the missing value sample and other samples, and selects K nearest neighbor sample values ​​to fill the missing value), linear interpolation (used for time series data, which fills the missing value by using the linear relationship between the data points before and after), or mean interpolation (used for cases with a single feature value, which fills the missing value by using the feature mean).

[0023] To eliminate dimensional differences between different features, all input features are standardized so that the mean of each feature is 0 and the standard deviation is 1. The standardization formula is: in, The original data set for a certain feature. The mean, Standard deviation, This is a standardized dataset, which helps improve the efficiency and stability of subsequent model training.

[0024] 3. Feature selection and dimensionality reduction: Combining domain knowledge and data analysis, the most relevant features for resisting wind damage, such as wind speed fluctuations, wind direction changes, air pressure changes, photovoltaic panel material and installation angle, are selected; principal component analysis (PCA) is used to reduce the dimensionality of the features, thereby reducing redundant information and computational complexity of model training.

[0025] 4. Dataset Partitioning: Divide the dataset into a training set (70%–80%) and a test set (20%–30%). During training, use K-fold cross-validation to evaluate model performance. Divide the training set into five subsets, selecting one subset as the validation set each time, and using the rest as the training set. Train and evaluate five times. The final evaluation result is the average of the K evaluation results, thus reducing overfitting and improving model stability.

[0026] Step 2: Design of the dual-task model; The dual-task model comprises a shared layer and independent task layers. The shared layer consists of multiple Transformer encoders, responsible for processing time-series data and generating high-dimensional feature representations. The independent task layers are divided into separate output layers for wind damage prediction and wind disaster risk assessment, used to perform damage prediction and risk assessment respectively. Figure 2 As shown, the details are as follows: Shared Layer (Basic Model Architecture): Employs a Transformer architecture, with self-attention as its core mechanism. This mechanism captures long-term dependencies from the input time-series data. The input includes multi-dimensional features such as wind speed, wind direction, temperature, humidity, photovoltaic panel material, and installation angle. These features are processed through multiple Transformer encoders to generate a high-dimensional feature representation. The self-attention mechanism is calculated using the following formula: in, It is a query matrix. It is a key matrix. It is a value matrix. For the dimensions of query and key, The weights for each time step are used to calculate the weights, and through this process, the model is able to capture the complex dependencies between the time steps.

[0027] The independent task layer is divided into separate output layers for wind damage prediction tasks and wind disaster risk assessment tasks, as described in detail below: Task 1: Wind Damage Prediction Task; The goal of the wind damage prediction task is to predict the potential damage level of photovoltaic panels in the future based on input features such as wind speed and direction. The model outputs the damage probability through a regression task, and the calculation formula is as follows: in, For a shared Transformer encoder, the output high-dimensional feature representation for a single input sample is... and It is a weight matrix. and It is a bias term. For the feedforward neural network, the mean squared error (MSE) is used as the loss function for the wind damage prediction task, and the calculation formula is as follows: in, For true damage labeling, To predict the probability of damage.

[0028] Task Two: Wind Disaster Risk Assessment; The goal of the wind disaster risk assessment task is to evaluate the damage risk of photovoltaic panels under given environmental conditions such as wind speed and direction. The model outputs a risk score through a regression task, representing the probability of damage to the photovoltaic panels under different environmental conditions. The calculation formula is as follows: in, For wind speed, For wind direction, For temperature, , ,... For the weighting coefficients, the mean squared error (MSE) loss function is also used: in, To score the true risk, A risk score for prediction.

[0029] Step 3: Model training and optimization; A joint loss function is used to balance the optimization objectives of the two tasks, and the formula is as follows: Where λ=0.5 is a hyperparameter that adjusts the contribution ratio of the two tasks to the total loss; the model is trained using the Adam optimizer, and the optimizer update formula is: in, For learning rate, and These are the first and second moment estimates of the gradient, respectively. It is a small constant; during training, cross-validation is used to evaluate model performance and adjust hyperparameters. Evaluation metrics include mean squared error (MSE) and the accuracy of risk prediction.

[0030] Step Four: Wind Injury Risk Assessment and Early Warning (e.g., Figure 3 (as shown) 1. Wind Damage Risk Scoring and Real-time Prediction: Real-time collected data on wind speed, wind direction, temperature, humidity, air pressure, photovoltaic panel material, and installation angle are input into a pre-trained dual-task model. The model generates feature representations through a shared layer and outputs a real-time wind damage risk score, using the following formula: in, For wind speed, For wind direction, For temperature, For humidity, For air pressure, It is made of photovoltaic panel material. The installation angle is determined by the model, which uses feature representations generated through a shared layer to predict risk scores.

[0031] 2. Wind injury early warning mechanism: Set risk scoring thresholds. ,like < It is then classified as low risk (normal monitoring). < < This is considered a medium risk (strengthen monitoring). ≥ This indicates a high risk; the early warning formula is: The system notifies relevant personnel via SMS and email and initiates necessary emergency responses.

[0032] Step 5: Model Update and Cross-Regional Transfer Learning; 1. Incremental Learning: When new wind disaster-related data is acquired, incremental learning is used to fine-tune the model in order to enable it to learn new knowledge and adapt to changes in data distribution. Specifically, the loss is calculated based on the new data, and the existing parameters θ of the model are updated using the same optimizer defined in step three, thereby optimizing the model's predictive ability while retaining historical knowledge.

[0033] 2. Transfer learning: Based on the characteristics of wind disasters in different geographical regions, the parameters of the existing model are fine-tuned to adapt the model to the data characteristics of the new region, so as to achieve cross-regional prediction and assessment. II. Specific Implementation Methods To verify the effectiveness of the method of the present invention, an experiment was conducted in a real photovoltaic power station scenario, as follows: 1. Experimental setup and data preparation; A photovoltaic power station located in Taishan City, Guangdong Province, with a total installed capacity of 50MW, was selected. Situated in a windy area, it is frequently affected by extreme weather events such as typhoons and thunderstorms. Experimental data was collected at a minute-by-minute time granularity, covering a six-month period, totaling over one million data points. Data types included: Environmental data: wind speed (m / s), wind direction (°), air temperature (°C), humidity (%), and air pressure (hPa); Photovoltaic panel data: photovoltaic panel material, installation angle, and historical damage records; Historical records of wind disasters: Damage to photovoltaic panels caused by past typhoons and strong winds and other extreme weather conditions.

[0035] 2. Experimental procedures; Step 1: Photovoltaic panel data acquisition and preprocessing; Data collection: Collect environmental data of photovoltaic power plants, data on photovoltaic panels themselves, and historical wind disaster records.

[0036] Data cleaning and missing value imputation: missing data were imputed using KNN interpolation and linear interpolation; outlier detection was performed on environmental data to remove obvious erroneous data; and standardization was applied to standardize all input features.

[0037] Feature selection and dimensionality reduction: Through domain knowledge and data analysis, features highly correlated with wind damage (such as wind speed fluctuations and wind direction changes) were selected. Principal component analysis (PCA) was then used to reduce the dimensionality of the data, preserving more than 95% of the variance.

[0038] Dataset splitting: The dataset was split into a training set (70%) and a test set (30%), and 5-fold cross-validation was performed.

[0039] Step 2: Design of the dual-task model; This model comprises two main tasks: wind damage prediction and wind disaster risk assessment. The entire model consists of a shared layer and independent task layers, with the specific steps as follows: Shared layer: Adopting the Transformer architecture, it processes the input multidimensional time series data through a self-attention mechanism and generates high-dimensional feature representations.

[0040] Independent task layer: Includes wind damage prediction task and wind disaster risk assessment task. The wind disaster risk assessment task uses a regression model to predict the damage probability of photovoltaic panels; the wind disaster risk assessment task assesses the damage risk of photovoltaic panels under given environmental conditions.

[0041] Step 3: Model training and optimization; (1) Joint loss function: The joint loss function is used to combine the losses of the two tasks (i.e., wind damage prediction loss and wind disaster risk assessment loss) to optimize the objective to maintain balance during training; the Adam optimizer is used to train the model and adjust the hyperparameters, and finally the model performance is evaluated by 5-fold cross-validation.

[0042] Step 4: Wind injury risk assessment and early warning; Wind injury risk scoring and real-time prediction: Input real-time data (wind speed and wind direction, etc.) into a trained dual-task model and output a wind injury risk score.

[0043] Step 5: Model Update and Cross-Regional Transfer Learning; Incremental learning: When new typhoon data arrives, the model is fine-tuned to enhance its predictive ability.

[0044] Transfer learning: To adapt to the characteristics of wind disasters in different geographical regions, the model supports transfer learning and fine-tunes the model based on data from new regions.

[0045] The method of this invention is compared with two existing methods: (1) Wind damage prediction method based on physical model: This method relies on physical parameters such as wind speed, wind direction and photovoltaic panel installation angle to simulate wind pressure to predict damage.

[0046] (2) Wind damage prediction method based on machine learning (SVM): It is trained using historical labeled data and uses the support vector machine (SVM) model for prediction.

[0047] The experimental results are shown in the table below: The self-supervised learning-based method proposed in this invention exhibits the smallest mean squared error (MSE) of 0.035 in wind damage prediction tasks, outperforming the physical model (0.089) and the SVM method (0.045). In wind disaster risk assessment tasks, the accuracy of this invention reaches 92.5%, significantly higher than the 78.3% of the physical model and the 85.2% of the SVM method.

[0048] The experimental results show that the self-supervised learning-based method for predicting and assessing wind damage to photovoltaic panels has significant advantages in the following aspects: Prediction accuracy: Because it adopts self-supervised learning, this method does not rely on a large amount of labeled data and can learn features from unlabeled data, thus improving prediction accuracy; Real-time performance and adaptability: Compared with traditional physical models, models based on self-supervised learning can adapt to new environmental data more quickly and make predictions in real time; Cross-regional adaptability: This method has strong cross-regional adaptability, capable of transfer learning based on new geographical areas, and provides customized wind disaster prediction services. It should be noted that, in this invention, although this specification describes embodiments, not every embodiment contains only one independent technical solution. This descriptive method is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for wind damage prediction and risk assessment of photovoltaic panels based on self-supervised learning, characterized by, The method comprises the following steps: S1: collecting multi-dimensional data related to photovoltaic panels and preprocessing the multi-dimensional data; S2: constructing a double-task model, the double-task model comprising a shared layer and two independent task layers, the shared layer being used for processing the preprocessed multi-dimensional data and generating high-dimensional feature representation, the two independent task layers being an anti-wind damage prediction task layer and a wind disaster risk assessment task layer respectively; S3: training and optimizing the double-task model using a joint loss function and an optimizer; S4: inputting the real-time collected multi-dimensional data related to photovoltaic panels after preprocessing into the double-task model after training and optimization, outputting the wind damage risk score of the photovoltaic panel, and performing a warning operation according to the wind damage risk score; S5: supporting parameter updating of the double-task model after training and optimization, and adapting to the anti-wind damage prediction and risk assessment requirements of photovoltaic panels across regions.

2. The self-supervised learning based photovoltaic panel wind damage resistance prediction and risk assessment method according to claim 1, characterized in that, In step S1, the multi-dimensional data comprises environmental data, photovoltaic panel body data, and historical record data; The environmental data at least includes wind speed, wind direction, temperature, humidity, and air pressure; The photovoltaic panel body data at least includes photovoltaic panel material and installation angle; The historical record data at least includes photovoltaic panel historical damage record and wind disaster historical record.

3. The self-supervised learning based photovoltaic panel wind damage resistance prediction and risk assessment method according to claim 1, characterized in that, In step S1, preprocessing includes data cleaning, missing value filling, standardization processing, feature selection and dimension reduction, and data set division; The data cleaning includes detecting and removing outliers in the multi-dimensional data; The feature selection and dimension reduction use principal component analysis (PCA); The data set division divides the multi-dimensional data into a training set and a test set, and K-fold cross-validation is used to evaluate the model performance during model training.

4. The self-supervised learning based photovoltaic panel wind damage resistance prediction and risk assessment method according to claim 1, characterized in that, In step S2, the shared layer uses a multi-layer Transformer encoder architecture; The Transformer encoder processes the preprocessed multi-dimensional data through a self-attention mechanism to capture the temporal dependence in the data, and then generates the high-dimensional feature representation.

5. The self-supervised learning based photovoltaic panel wind damage resistance prediction and risk assessment method according to claim 1, characterized in that, In step S2, the anti-wind damage prediction task layer is constructed based on a feedforward neural network, taking the high-dimensional feature representation as input, and outputs the damage probability of the photovoltaic panel through regression operation; The wind disaster risk assessment task layer takes the high-dimensional feature representation as input and outputs the wind damage risk score of the photovoltaic panel through regression operation.

6. The self-supervised learning based photovoltaic panel wind damage resistance prediction and risk assessment method according to claim 1, characterized in that, In step S3, the joint loss function is the weighted sum of the anti-wind damage prediction task loss and the wind disaster risk assessment task loss; Both the anti-wind damage prediction task loss and the wind disaster risk assessment task loss use mean square error (MSE) loss function; the optimizer uses Adam optimizer.

7. The self-supervised learning based photovoltaic panel wind damage resistance prediction and risk assessment method according to claim 6, characterized in that, The expression of the joint loss function is: wherein, is the loss for the wind damage prediction task, is the loss for the wind damage risk assessment task, and λ is a hyperparameter that regulates the contribution of the losses of the two tasks to the total loss.

8. The self-supervised learning based photovoltaic panel wind damage resistance prediction and risk assessment method according to claim 1, characterized in that, In step S4, the warning operation includes: Setting at least two risk score thresholds, comparing the real-time output wind damage risk score with the risk score thresholds, and dividing at least three risk levels of low risk, medium risk, and high risk; Performing corresponding operations for different risk levels, performing normal monitoring under low risk level, performing enhanced monitoring under medium risk level, and sending notifications to relevant personnel and starting emergency response under high risk level.

9. The self-supervised learning based photovoltaic panel wind damage resistance prediction and risk assessment method according to claim 1, characterized in that, In step S5, the parameter update of the double-task model adopts an incremental learning manner: When new wind disaster related data is obtained, the existing parameters of the model are fine-tuned based on the new data to optimize the prediction ability of the model.

10. The self-supervised learning based photovoltaic panel wind damage resistance prediction and risk assessment method according to claim 1, characterized in that, In step S5, the adaptation to cross-regional demand adopts a transfer learning manner: based on the wind disaster feature data of the target region, the parameters of the double-task model optimized by training are fine-tuned to adapt the model to the photovoltaic panel wind damage resistance prediction and risk assessment scene of the target region.