Personalized photovoltaic fault prediction method based on customized prediction module and federated learning

Abstract

The invention discloses a personalized photovoltaic fault prediction method based on a customized prediction module and federal learning, and belongs to the technical field of photovoltaic power generation. According to the method, the independent customized prediction module is combined with the general trunk model, so that the unique fault characteristics of the local photovoltaic power station are captured, and the general mode of the photovoltaic fault is utilized, and the prediction precision is improved; personalized reasoning of the local model further enhances the judgment capability of local fault conditions. On the basis of federated learning, multi-source data is used for learning, a dynamic node participates in a strategy, safe sharing and efficient collaboration of data are guaranteed, and the generalization ability of the model is improved. According to the model training optimization method based on reinforcement learning, real-time and accurate fault prediction and diagnosis are realized, the operation and maintenance efficiency is greatly improved, the operation and maintenance cost is reduced, and the reliability and stability of the photovoltaic power generation system are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic power generation technology, and in particular to a personalized photovoltaic fault prediction method based on a customized prediction module and federated learning. Background Art

[0002] Globally, with growing awareness of environmental protection and concerns about the depletion of traditional fossil fuels, the demand for clean energy is rapidly growing. Photovoltaic power generation, as a crucial renewable energy source, has been widely adopted and rapidly developed worldwide due to its significant advantages, such as cleanliness and sustainability. From large-scale ground-mounted photovoltaic power plants to distributed rooftop photovoltaic power generation systems, photovoltaic facilities are mushrooming, injecting strong impetus into the transformation of energy structure and sustainable development.

[0003] However, photovoltaic power plants face numerous challenges in their actual operation. For one thing, complex and variable environmental factors, such as high temperatures, high humidity, dust, and severe diurnal temperature swings, can negatively impact the performance of photovoltaic equipment. Furthermore, equipment ages over time, leading to component wear and corrosion. These intertwined factors contribute to frequent equipment failures in photovoltaic power plants. Equipment failures not only directly reduce power generation, severely impacting photovoltaic power generation efficiency, but can also cause power outages, threatening the stability of the power system and significantly impacting both power supply and user access.

[0004] Traditional photovoltaic fault prediction methods have exposed numerous shortcomings when addressing these complex situations. Most traditional methods utilize a single model or algorithm, lacking sufficient flexibility and versatility. PV power plants in different regions vary in geographical environments, climatic conditions, equipment types, and operational management methods, resulting in unique operating characteristics and data features. Traditional methods fail to fully account for these differences, making them difficult to adapt to diverse application scenarios. This results in low prediction accuracy and an inability to meet the needs of actual production operations.

[0005] At the same time, data privacy and security issues are becoming increasingly prominent during data processing and model training. With the increasing intelligence of photovoltaic power plants, large amounts of operational data are being collected and stored. This data contains critical information about the power plant, and if leaked, it could cause significant losses to businesses and users. Traditional methods have limitations in data security and protection, making it difficult to meet the security requirements of collaborative utilization of multi-source data. This limits the efficient integration and analysis of data and hinders the development of more accurate fault prediction models.

[0006] In summary, developing accurate, personalized, and secure data processing and model training methods has become a critical issue that needs to be addressed in the photovoltaic power generation field. This is not only related to the efficient and stable operation of photovoltaic power plants, but also has important significance for promoting the sustainable development of the entire photovoltaic power generation industry. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this paper proposes a personalized photovoltaic fault prediction method based on a customized prediction module and federated learning. This method aims to address existing photovoltaic fault prediction methods, including poor model universality, insufficient data utilization, and difficulties in ensuring data privacy. Through innovative technical means, this paper achieves accurate and personalized photovoltaic fault prediction while ensuring data security and efficient collaborative utilization.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A personalized photovoltaic fault prediction method based on a customized prediction module and federated learning includes the following steps:

[0010] S1, collects multi-source data of photovoltaic power plants and performs data preprocessing;

[0011] S2, based on the preprocessed data, builds and trains a backbone model combining a deep belief network and a convolutional neural network for each PV plant to extract common features of PV faults;

[0012] S3, based on the common characteristics of photovoltaic faults and the unique data characteristics of each photovoltaic power station, building an independent customized prediction module for each photovoltaic power station to extract local unique fault characteristics;

[0013] S4, based on the real-time data quality and model training results of each PV plant, dynamically adjust the PV plants participating in federated learning and perform federated learning aggregation;

[0014] In S5, the local PV power station obtains real-time PV data and, after preprocessing, inputs it into the backbone model to extract the common features of PV faults. The common features of PV faults and the data features unique to the local PV power station are then input into the customized prediction module for fault inference calculation.

[0015] Furthermore, collecting multi-source data from photovoltaic power stations includes: deploying distributed intelligent sensor networks at key locations in photovoltaic power stations (such as photovoltaic modules, junction boxes, inverters, etc.); these sensors are equipped with adaptive collection frequency adjustment functions, which can adjust the data collection frequency in real time according to factors such as light intensity, temperature, and equipment operating status; when the light intensity changes in a short period of time and exceeds the set threshold, or when the equipment current and voltage fluctuate abnormally, the collection frequency is automatically increased to ensure that key data is not missed.

[0016] Furthermore, the data preprocessing includes:

[0017] Wavelet transform method is used to decompose and reconstruct the original data;

[0018] Normalize the reconstructed signal data;

[0019] The principal components of the data were extracted using principal component analysis.

[0020] Furthermore, during the data preprocessing process, the original data is decomposed and reconstructed using the wavelet transform method, which specifically includes:

[0021] First, the collected raw data is subjected to discrete wavelet transform (DWT), and multi-resolution analysis is performed through a low-pass filter h(n) and a high-pass filter g(n). The decomposition formula is:

[0022]

[0023]

[0024] Among them, x(n) is the original signal, cA j (k) is the approximate coefficient of the jth layer (low frequency part), cD j (k) is the detail coefficient (high frequency part) of the jth layer, j∈[1,J], J is the maximum number of decomposition layers; and They are conjugate mirror filters, h(n) and g(n) satisfy g(n)=(-1) n h(1-n);

[0025] Then the soft threshold function is used to process the detail coefficients of the high-frequency part to remove noise. The formula is:

[0026]

[0027] in, is the detail coefficient after threshold processing; sgn() is the sign function; λ j is the threshold of the jth layer, which is adaptively selected according to the data characteristics and noise level;

[0028] Finally, the inverse discrete wavelet transform (IDWT) is performed to reconstruct the signal to obtain purer data. The reconstruction formula is:

[0029]

[0030] Where x′(n) is the reconstructed signal.

[0031] Furthermore, in the data preprocessing process, the core formula of the principal component analysis (PCA) method is to calculate the covariance matrix of the data matrix x′(n):

[0032]

[0033] Then solve the eigenvalues and eigenvectors of the covariance matrix C, select the eigenvectors corresponding to the first z largest eigenvalues to form the transformation matrix U, then the principal component Y = x′(n)U, Y is the data after dimensionality reduction, which removes redundant information, further explores the potential value of the data, and improves the efficiency of subsequent analysis.

[0034] Furthermore, in S2, the backbone model adopts a hybrid model architecture that combines a deep belief network (DBN) with a convolutional neural network (CNN). Through this hybrid architecture, the backbone model can fully capture the common characteristics of photovoltaic fault data in time series and spatial distribution. Specifically, DBN is first used to learn the time-varying characteristics:

[0035]

[0036] The DBN is composed of multiple stacked restricted Boltzmann machines (RBMs). RBMs can discover high-order statistical features in data through unsupervised learning, providing more abstract and representative feature expressions for subsequent models. The energy function of RBM is:

[0037]

[0038] Where v is the visible layer neuron state vector; h is the hidden layer neuron state vector; θ = {w ij ,a i ,b j} is the model parameter; w ij is the weight connecting the i-th neuron in the visible layer and the j-th neuron in the hidden layer; a i is the bias of the i-th neuron in the visible layer, b j is the bias of the jth neuron in the hidden layer; m and n represent the number of neurons in the visible layer and hidden layer respectively;

[0039] The convolutional layer of the CNN is responsible for extracting the local spatial features of the data. The calculation formula of the convolutional layer is:

[0040]

[0041] in, is the output of the lth convolutional layer, σ is the activation function, is the weight of the convolution kernel in layer l, is the neuron input at the corresponding position in the l-1 layer, b lis the bias of the lth layer, P and Q are the sizes of the convolution kernel;

[0042] The stochastic gradient descent (SGD) algorithm is used to update the model parameters during backbone model training. The parameter update formula is:

[0043] θ t+1 =θ t -η▽J(θ t )

[0044] Among them, θ t is the model parameter of the tth iteration, η is the learning rate, ▽J(θ t ) is the loss function J with respect to the parameter θ t gradient;

[0045] At the same time, in order to prevent overfitting, the L2 regularization term is introduced, and the new loss function is:

[0046]

[0047] Where λ is the regularization coefficient;

[0048] Through continuous iterative training, the backbone model can accurately extract the common features of photovoltaic faults.

[0049] Furthermore, in S3, the customized prediction module construction process includes:

[0050] Based on the unique historical data and operating characteristics of each PV plant, the customized prediction module adopts a multi-layer perceptron (MLP) architecture. MLP can perform highly nonlinear transformations on input data to adapt to the complex failure modes of different PV plants. The universal photovoltaic failure characteristics output by the backbone model and the data characteristics specific to the local PV plant are used as inputs to the MLP to further explore the local specific failure modes and data characteristics. For an MLP with L layers, the calculation formula is:

[0051] h l =σ(W l h l-1 +b l )

[0052] Among them, h l is the output of layer l, σ is the activation function, W l is the weight matrix of layer l, h l-1 is the output of layer l-1, b l is the bias vector of the lth layer;

[0053] The customized prediction module's parameters are adjusted and optimized using historical fault data and real-time operating data from each PV plant. The parameter gradients are calculated using a backpropagation algorithm and updated using gradient descent to minimize the error between the predicted results and the actual fault conditions. This allows the customized prediction module to closely adapt to the actual operating conditions of the local PV plant, improving the accuracy of fault prediction.

[0054] Furthermore, the specific process of S4 is as follows:

[0055] Data quality assessment uses root mean square error (RMSE) and data integrity indicators. Based on the RMSE and data integrity assessment results, a participation weight q is assigned to each PV power station. o ;RMSE calculation formula is:

[0056]

[0057] Among them, y i is the true value, is the predicted value, n is the number of samples;

[0058] During the federated learning aggregation process, only the parameters of the backbone model are aggregated. The parameters of the customized prediction module are retained locally and do not participate in the aggregation to maintain its adaptability to the local PV power station. The improved federated averaging algorithm aggregation formula is:

[0059]

[0060] in, are the backbone model parameters of the o-th PV plant after the t-th round of local training; are the global backbone model parameters after the t+1th round of aggregation.

[0061] Furthermore, in S5, after federated learning, the real-time operating data acquired by the local PV power plant is preprocessed and feature extracted, and then input into the backbone model to extract common features of PV faults. These common features are then combined with data features specific to the local power plant and input into the customized prediction module. The customized prediction module performs inference calculations on the input features based on the trained parameters.

[0062] For a customized prediction module with a failure probability output, its output P fault The input feature vector F is obtained by performing a series of matrix operations and activation function processing, which can be expressed as:

[0063] P fault =f(W final ·σ(W l-1 ·(…σ(W1·F+b1)…)+bl-1 )+b final )

[0064] Where W and b are the weight matrix and bias vector of each layer respectively; f is the final output activation function, which is used to map the output to the fault probability range.

[0065] Furthermore, the method also includes dynamic model updating, and further optimization using reinforcement learning to cope with knowledge drift caused by factors such as possible future environmental changes;

[0066] The model's prediction accuracy, recall rate and other indicators are used as reward signals, and the model's training parameters and strategies are automatically adjusted through the reinforcement learning algorithm. The Q-learning algorithm is used, and the Q value update formula is:

[0067] Q(s,a)←Q(s,a)+α[r+γmax a′ Q(s′,a′)-Q(s,a)]

[0068] Where s is the current state, a is the current action, Q(s,a) is the Q value of taking action a in state s, α is the learning rate, r is the reward, γ is the discount factor, s′ is the next state, and a′ is the next action;

[0069] The training parameters of the backbone model and customized prediction module, such as learning rate and weight, are adjusted according to the Q value. When the loss function on the validation set decreases by less than the set threshold in multiple consecutive iterations, the model update is triggered. Each photovoltaic power station re-collects new data within a certain period of time, performs preprocessing and feature extraction, and then re-participates in federated learning to update the global backbone model. The customized prediction module is independently trained and optimized based on local new data.

[0070] Compared with the existing technology, the beneficial effects of the present invention are as follows: the present invention not only captures the unique fault characteristics of local photovoltaic power stations through the combination of independent customized prediction modules and a universal backbone model, but also utilizes the universal pattern of photovoltaic failures, thereby improving the prediction accuracy; the personalized reasoning of the local model further enhances the ability to judge local fault conditions. Based on federated learning, the use of multi-source data for learning and the dynamic node participation strategy ensures data security sharing and efficient collaboration, and improves the generalization ability of the model. The model training optimization method based on reinforcement learning realizes real-time and accurate fault prediction and diagnosis, greatly improves operation and maintenance efficiency, reduces operation and maintenance costs, and enhances the reliability and stability of the photovoltaic power generation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a flow chart of the personalized photovoltaic fault prediction method based on customized prediction module and federated learning of the present invention. DETAILED DESCRIPTION

[0072] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0073] like Figure 1 As shown, an embodiment of the present invention provides a personalized photovoltaic fault prediction method based on a customized prediction module and federated learning, comprising the following steps:

[0074] S1, collect multi-source data and pre-process the collected data using wavelet transform and normalization methods; specifically:

[0075] In a large photovoltaic power station, a distributed intelligent sensor network is deployed on 100 photovoltaic panels, 5 combiner boxes and 3 inverters. During the operation of a week, when the light intensity changes by more than 100W / m in a certain hour, 2 When the sensor automatically increases the collection frequency from once per minute to once every 10 seconds, a total of 50,000 valid data items are collected.

[0076] The collected raw voltage data is subjected to wavelet transform noise reduction. The decomposition layer number J=5, the db4 wavelet basis function is used, and the approximate coefficients and detail coefficients of each layer are obtained through discrete wavelet transform; the detail coefficients are thresholded, and the threshold λ j Calculated according to the following formula:

[0077]

[0078] Where σ is the noise standard deviation and n is the data length.

[0079] After reconstruction using an inverse discrete wavelet transform, the voltage data's noise was significantly reduced, and the signal-to-noise ratio (SNR) improved by 15dB. Data containing 10 features, including voltage, current, and temperature, were normalized and subjected to principal component analysis. The normalized data range was unified to [0, 1]. Principal component analysis selected the top three principal components, achieving a cumulative contribution rate of 90%, effectively reducing the data dimensionality.

[0080] S2, building a backbone model, which uses a hybrid model architecture combining a deep belief network (DBN) and a convolutional neural network (CNN) as the backbone; specifically:

[0081] The DBN consists of three RBM layers. The number of neurons in the visible layer of each RBM is determined to be 100 based on the input data dimensions, and the number of neurons in the hidden layer is 80, 60, and 40, respectively. The CNN consists of three convolutional layers with kernel sizes of 3×3, 5×5, and 3×3, respectively, with a step size of 1. Reluctant Unit (ReLU) is used as the activation function. The backbone model is trained using the stochastic gradient descent algorithm, with a learning rate of η = 0.01 and 1000 iterations.

[0082] S3 builds a customized prediction module that uses a multi-layer perceptron (MLP) architecture to perform highly nonlinear transformations on input data to adapt to the complex failure modes of different power plants. Specifically:

[0083] A customized prediction module was built for this PV power plant, using a three-layer MLP. The number of input neurons in the first layer matched the common PV fault feature dimension output by the backbone model, 50. The number of neurons in the hidden layer was 30 and 20, respectively, and the number of neurons in the output layer was 1 (representing the failure probability). The Sigmoid function was used as the activation function. Using the plant's historical fault data and real-time operating data from the past year, the backpropagation algorithm was used to calculate parameter gradients. Parameters were updated at a learning rate of 0.001, and training was performed over 500 iterations.

[0084] S4, based on the real-time data quality and model training results of each PV plant, dynamically adjusts the PV plants participating in federated learning and performs federated learning aggregation; specifically:

[0085] Assume that there are 5 photovoltaic power stations participating in federated learning. Each power station completes data collection, preprocessing, and training of local models (backbone model and customized prediction module) according to the above method. In the process of federated learning, after the local model is trained in the tth round, the parameters of the backbone model are The data quality of each power station was evaluated using RMSE and data integrity metrics. The backbone model parameters were then aggregated using an improved federated averaging algorithm, completing a round of federated learning.

[0086] S5, local personalized reasoning: Real-time operating data obtained from the local PV power station is input into the backbone model to extract common features of PV faults. The common features of PV faults are then combined with the data features unique to the local power station and input into the customized prediction module for reasoning and calculation. Specifically:

[0087] At a certain moment, the real-time collected and pre-processed photovoltaic power station operation data is input into the backbone model and customized prediction module in turn. The backbone model extracts the common features of photovoltaic faults, and the customized prediction module combines the local data features. Assuming that the output of the customized prediction module is P fault . Set the fault threshold P threshold =0.7, if P fault If the threshold is exceeded, a fault is determined to be possible. The other components connected to the faulty component are identified, and the possible range of the fault is assessed. In this embodiment, the five components closest to the component are set to be affected, and a fault warning is issued in a timely manner.

[0088] S6, dynamic model update, uses reinforcement learning to optimize to further cope with knowledge drift caused by factors such as possible environmental changes; specifically:

[0089] The model is optimized using the Q-learning algorithm. Prediction accuracy and recall are used as reward signals. For example, when the prediction accuracy reaches 90% and the recall reaches 85% in a certain round, a reward of r = 10 is given. The learning rate α is set to 0.1, and the discount factor γ is set to 0.9. In state s, action a is taken to adjust the learning rate. The Q-value update formula is Q(s,a)←Q(s,a)+α[r+γmax a′ Q(s′,a′)-Q(s,a)] updates the Q value, and then adjusts the training parameters of the backbone model and customized prediction module.

[0090] Finally, it should be noted that the above embodiments are intended to illustrate the technical solutions of the present invention and do not constitute any form of limitation of the present invention. Those skilled in the art should fully understand that it is entirely feasible to modify the technical solutions described in the above embodiments or to replace any or all of the technical features with equivalents. Such modifications or replacements, as long as they do not deviate from the scope of protection defined by the claims of the present invention, should be considered reasonable extensions of the present invention.

Claims

1. A personalized photovoltaic fault prediction method based on customized prediction modules and federated learning, characterized by: The following steps are involved: S1, collects multi-source data of photovoltaic power plants and performs data preprocessing; S2, based on the preprocessed data, builds and trains a backbone model combining a deep belief network and a convolutional neural network for each PV plant to extract common features of PV faults; S3, based on the common characteristics of photovoltaic faults and the unique data characteristics of each photovoltaic power station, building an independent customized prediction module for each photovoltaic power station to extract local unique fault characteristics; S4, based on the real-time data quality and model training results of each PV plant, dynamically adjust the PV plants participating in federated learning and perform federated learning aggregation; In S5, the local PV power station obtains real-time PV data and, after preprocessing, inputs it into the backbone model to extract the common features of PV faults. The common features of PV faults and the data features unique to the local PV power station are then input into the customized prediction module for fault inference calculation.

2. The personalized photovoltaic fault prediction method based on customized prediction module and federated learning according to claim 1, characterized in that: Collecting multi-source data from photovoltaic power stations includes: deploying a distributed intelligent sensor network at key locations in the photovoltaic power station. The sensors are equipped with an adaptive acquisition frequency adjustment function, which can adjust the data acquisition frequency in real time according to light intensity, temperature, and equipment operating status.

3. The personalized photovoltaic fault prediction method based on customized prediction module and federated learning according to claim 1 or 2, characterized in that: The data preprocessing includes: Wavelet transform method is used to decompose and reconstruct the original data; Normalize the reconstructed signal data; The principal components of the data were extracted using principal component analysis.

4. The personalized photovoltaic fault prediction method based on customized prediction module and federated learning according to claim 3 is characterized in that: The wavelet transform method used is used to decompose and reconstruct the original data, including: First, the collected raw data is subjected to discrete wavelet transform, and multi-resolution analysis is performed through low-pass filter h(n) and high-pass filter g(n). The decomposition formula is: Among them, x(n) is the original signal, cA j (k) is the approximation coefficient of the jth layer, cD j (k) is the detail coefficient of the jth layer, j∈[1,J], J is the maximum number of decomposition layers; and They are conjugate mirror filters; Then the soft threshold function is used to process the detail coefficients of the high-frequency part to remove noise. The formula is: in, is the detail coefficient after threshold processing; sgn() is the sign function; λ j is the threshold of the jth layer, which is adaptively selected according to the data characteristics and noise level; Finally, the inverse discrete wavelet transform is performed to reconstruct the signal to obtain purer data. The reconstruction formula is: Where x′(n) is the reconstructed signal.

5. The personalized photovoltaic fault prediction method based on customized prediction module and federated learning according to claim 4, characterized in that: The core formula of the principal component analysis method is to calculate the covariance matrix of the data matrix x′(n): Then solve the eigenvalues and eigenvectors of the covariance matrix C, select the eigenvectors corresponding to the first z largest eigenvalues to form the transformation matrix U, then the principal component Y = x′(n)U, where Y is the data after dimensionality reduction.

6. The personalized photovoltaic fault prediction method based on customized prediction module and federated learning according to claim 5, characterized in that: The specific method of building and training the backbone model is: First, use the deep belief network DBN to learn the time-varying features: The DBN is composed of multiple restricted Boltzmann machines (RBMs) stacked together, and the energy function of RBM is: Where v is the visible layer neuron state vector; h is the hidden layer neuron state vector; θ = {w ij ,a i ,b j } is the model parameter; w ij is the weight connecting the i-th neuron in the visible layer and the j-th neuron in the hidden layer; a i is the bias of the i-th neuron in the visible layer, b j is the bias of the jth neuron in the hidden layer; m and n represent the number of neurons in the visible layer and hidden layer respectively; The convolutional layer of the convolutional neural network CNN is responsible for extracting the local spatial features of the data. The calculation formula of the convolutional layer is: in, is the output of the lth convolutional layer, σ is the activation function, is the weight of the convolution kernel in layer l, is the neuron input at the corresponding position in the l-1 layer, b l is the bias of the lth layer, P and Q are the sizes of the convolution kernel; During backbone model training, the stochastic gradient descent algorithm is used to update the model parameters. The parameter update formula is: Among them, θ t is the model parameter of the tth iteration, η is the learning rate, is the loss function J with respect to the parameter θ t gradient; At the same time, in order to prevent overfitting, the L2 regularization term is introduced, and the new loss function is: Among them, λ is the regularization coefficient; through continuous iterative training, the backbone model can accurately extract the common features of photovoltaic faults.

7. The personalized photovoltaic fault prediction method based on customized prediction module and federated learning according to claim 6, characterized in that: The customized prediction module construction process includes: The customized prediction module adopts a multi-layer perceptron (MLP) architecture. It uses the general features of photovoltaic faults output by the backbone model and the data features unique to the local photovoltaic power station as the input of the MLP to further explore the local specific fault modes and data features. For an MLP with L layers, the calculation formula is: h l =σ(W l h l-1 +b l ) Among them, h l is the output of layer l, σ is the activation function, W l is the weight matrix of layer l, h l-1 is the output of layer l-1, b l is the bias vector of the lth layer; The parameters of the customized prediction module are adjusted and optimized using the historical fault data and real-time operation data of each photovoltaic power station. The gradient of the parameters is calculated using the backpropagation algorithm, and the parameters are updated according to the gradient descent method to minimize the error between the prediction results and the actual fault conditions.

8. The personalized photovoltaic fault prediction method based on customized prediction module and federated learning according to claim 7, characterized in that: In S4, the data quality assessment uses the root mean square error and data integrity indicators. According to the root mean square error and data integrity assessment results, a participation weight q is assigned to each photovoltaic power station. o ; During the federated learning aggregation process, only the parameters of the backbone model are aggregated. The parameters of the customized prediction module are retained locally and do not participate in the aggregation to maintain its adaptability to the local PV power station. The improved federated averaging algorithm aggregation formula is: in, are the backbone model parameters of the o-th PV plant after the t-th round of local training; are the global backbone model parameters after the t+1th round of aggregation.

9. The personalized photovoltaic fault prediction method based on customized prediction module and federated learning according to claim 7, characterized in that: In S5, for a customized prediction module with a fault probability output, its output P fault The input feature vector F is obtained by performing a series of matrix operations and activation function processing, which is expressed as: P fault =f(W final ·σ(W l-1 (…σ(W1 F+b1)…)+b l-1 )+b final ) Where W and b are the weight matrix and bias vector of each layer respectively; f is the final output activation function, which is used to map the output to the fault probability range.

10. The personalized photovoltaic fault prediction method based on customized prediction module and federated learning according to claim 1, characterized in that: The Q-learning algorithm is used to optimize the model, and the Q value update formula is: Q(s,a)←Q(s,a)+α[r+γmax a′ Q(s′,a′)-Q(s,a)] Where s is the current state, a is the current action, Q(s,a) is the Q value of taking action a in state s, α is the learning rate, r is the reward, γ is the discount factor, s′ is the next state, and a′ is the next action; The training parameters of the backbone model and customized prediction module are adjusted according to the Q value. When the loss function on the validation set decreases by less than the set threshold in multiple consecutive iterations, the model update is triggered. Each photovoltaic power station re-collects new data within a certain period of time, performs preprocessing and feature extraction, and then re-participates in federated learning to update the global backbone model. The customized prediction module is separately trained and optimized based on local new data.

Images