Cross-region photovoltaic power generation power prediction method based on federated learning
By employing personalized federated learning, learnable mask aggregation, and nuclear norm constraints, combined with Grassman manifold optimization, the problems of data privacy protection, personalized adaptation, and computational efficiency in cross-regional photovoltaic power generation forecasting are solved, achieving efficient cross-regional photovoltaic power generation forecasting.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF INFORMATION SCI & TECH
- Filing Date
- 2025-04-14
- Publication Date
- 2026-05-15
AI Technical Summary
Cross-regional photovoltaic power generation forecasting faces challenges such as data privacy protection, difficulty in adapting to personalized models, high computational load, and high optimization difficulty. Existing federated learning algorithms cannot effectively solve these problems.
By employing a personalized federated learning algorithm, combined with learnable mask aggregation, nuclear norm constraints, and Grassman manifold optimization, cross-regional photovoltaic power generation prediction is achieved by optimizing model parameters on the Grassman manifold, using learnable masks for parameter aggregation, and introducing nuclear norms to control computational complexity.
It achieves precise adaptation to personalized needs while ensuring data privacy, reduces computational load and optimization difficulty, improves model prediction performance and computational efficiency, and solves multiple challenges in cross-regional photovoltaic power generation prediction.
Smart Images
Figure CN120372682B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of photovoltaic power generation prediction, and specifically relates to a method for predicting cross-regional photovoltaic power generation. Background Technology
[0002] In the field of cross-regional photovoltaic power generation forecasting, the protection of data privacy and the construction of personalized forecasting models are crucial. On the one hand, when a company deploys photovoltaic equipment in multiple countries around the world, different countries, based on their own data privacy protection regulations, typically require that data generated domestically be stored on local servers. In this case, the company cannot directly access the raw data, which increases the difficulty of data integration and utilization to some extent. On the other hand, there are significant differences in sunlight conditions across different regions. If a single model is used for power generation forecasting, it is difficult to accurately adapt to the different conditions in each region. Therefore, this invention innovatively introduces a personalized federated learning algorithm. This algorithm can effectively integrate data information from different participants while ensuring the privacy and security of all participating parties' data, and accurately design personalized models for different participants, thereby achieving the dual goals of cross-regional photovoltaic data privacy protection and accurate power generation forecasting. However, if only existing personalized federated learning algorithms are used, the following problems will arise:
[0003] 1) The aggregation strategy for photovoltaic prediction models in different regions is too simplistic, leading to a decline in model performance. Specifically, traditional federated learning aggregation algorithms often use a simple weighted summation of the parameters of the entire model, which is clearly detrimental to the performance of federated learning models. To address this, this invention uses a learnable mask for aggregation, assigning different aggregation coefficients to each model parameter to achieve more accurate model parameter aggregation.
[0004] 2) While using learnable masks to improve model performance, the computational requirements for photovoltaic prediction models in different regions increase dramatically. Therefore, this invention needs to consider the balance between computational complexity and accuracy. To this end, this invention introduces a nuclear norm to constrain the computational complexity of the learnable mask. By constraining the rank of the mask matrix, the computational complexity can be significantly reduced while ensuring that model performance is almost unaffected.
[0005] 3) Optimizing photovoltaic prediction models for different regions simultaneously with both model parameters and masks is too difficult. To address this, this invention assumes that the model parameters reside on a Grassmann manifold, meaning that complex parameters to be optimized can be represented using simpler manifold parameters. In this case, the optimization problem of model parameters is transformed into a manifold parameter optimization problem, reducing the number of parameters to be optimized and thus lowering the optimization difficulty.
[0006] The technical fields involved in this invention are briefly described below to better understand the core content and innovative value of this invention. Specifically, the technical fields include federated learning, personalized federated learning, learnable mask aggregation, nuclear norm, and Grassman manifold.
[0007] Federated learning is a distributed machine learning approach involving multiple participants, allowing them to collaboratively train a deep learning model without sharing data. In a federated learning framework, data remains local; only model updates are sent to the server for aggregation, thus protecting the participants' data privacy. This approach is particularly suitable for domains with high data privacy requirements, such as cross-regional photovoltaic power generation forecasting.
[0008] Personalized federated learning aims to address the problem of traditional federated learning failing to capture the local data distribution of different participants. By generating personalized models for each participant, it can adapt to their unique data distribution and needs. This approach is particularly suitable for scenarios with uneven data distribution, such as personalized cross-regional photovoltaic power generation forecasting.
[0009] Learnable mask aggregation is an advanced aggregation method in federated learning. Specifically, it replaces the traditional simple weighted summation method by assigning a learnable mask to each model parameter. This method can more flexibly and accurately aggregate the model parameters of each participant, thereby improving the performance of the aggregated model.
[0010] The nuclear norm is considered a convex relaxation technique for matrix rank constraints in matrix optimization. By constraining the nuclear norm of a matrix, its rank can be controlled to some extent, while transforming a complex non-convex optimization problem into a relatively simple convex one, thus effectively reducing the complexity of the optimization process. This method balances computational feasibility and efficiency when dealing with matrix-related problems.
[0011] A Grassman manifold is a mathematical structure that describes all specific-dimensional subspaces in a vector space. For example, in three-dimensional space, a Grassman manifold can be the set of all lines or the set of all planes. By assuming that the model parameters lie on a Grassman manifold, the optimization variables can be transformed from model parameters to manifold parameters. This reduces the number of parameters to be optimized, thus lowering the optimization difficulty. Summary of the Invention
[0012] The purpose of this invention is to provide a cross-regional photovoltaic power generation prediction method based on federated learning, so as to solve the three core problems of data privacy protection, personalized adaptation, and balance between computational efficiency and model accuracy in cross-regional photovoltaic prediction, and to overcome the problem of high-dimensional parameter space optimization.
[0013] To achieve the above objectives, the present invention adopts the following technical solution:
[0014] A federated learning-based method for predicting cross-regional photovoltaic power generation includes the following steps:
[0015] Step 1: Each region participating in photovoltaic power generation prediction corresponds to one participant, and a server is deployed for cross-regional parameter aggregation; each participant prepares local data containing historical photovoltaic data and meteorological data, where the historical photovoltaic data includes photovoltaic power time series, and preprocesses the local data; the server distributes the initial photovoltaic prediction global model to each participant;
[0016] Step 2: After each participant receives the initialized photovoltaic prediction model, it is used as a local model, and local data is used to optimize the local model parameters. Each participant projects the local model parameters onto the Grassmann manifold, first optimizing the local model... The model parameters of each layer are decomposed using QR decomposition. Then, the orthogonal matrix Q is fixed to obtain the orthogonal matrix Q and the upper triangular matrix R. The upper triangular matrix R is used as the manifold parameters to be optimized for the i-th participant in the t-th round of training. ;
[0017] Step 3: The photovoltaic prediction models of each participant are locally masked.
[0018] Step 4: Each participant optimizes the learnable mask in photovoltaic prediction. ;
[0019] Step 5: Each participant submits the photovoltaic model parameters. The model is sent to the server, which detects abnormal models and then calculates the average of the parameters of the remaining models as the new global model.
[0020] Step 6: Repeat steps 2 through 5 until... It converges or loops to a specified number of times.
[0021] Furthermore, in step 1, the local data prepared by each participant is as follows: the local data of the i-th participant is... and , The input data for photovoltaic (PV) forecasting includes historical PV data and meteorological data; the historical PV data includes three hours of historical data, with each five-minute data point corresponding to a time point; the meteorological data includes temperature and solar radiation intensity. The photovoltaic power generation is to be predicted for the next half hour.
[0022] Furthermore, in step 1, the preprocessing steps for the local data are as follows: the data is filled with missing values based on K nearest neighbors, where the number of neighbors is 1. After filling the missing values, the local data is normalized by mean-variance.
[0023] Furthermore, in step 1, the server will use the same initial global photovoltaic prediction model. Distributed to each participant, where t represents the round number and g indicates that the model is a global model; Global photovoltaic prediction model. It is a time-series prediction model based on convolutional neural networks and long short-term memory networks. Historical photovoltaic data is processed through three one-dimensional convolutional layers to obtain historical photovoltaic features, and meteorological data is processed through a fully connected network layer to obtain meteorological features. Then, the historical photovoltaic features and meteorological features are concatenated and input into the long short-term memory network. Finally, the time-series prediction results of this network are input into a single-layer fully connected neural network to obtain the specific prediction results.
[0024] Furthermore, the number of channels in the one-dimensional convolutional layer is 16, 32, and 64, the number of filters is 90, 20, and 50, the pooling size is 2, the convolutional kernel size is 3, the stride is 1, and the activation function is ReLU; the output dimension of the fully connected network layer is 64; the long short-term memory network contains 100 units, the dropout rate is set to 0.1, and the activation function is ReLU.
[0025] Furthermore, in step 2, QR decomposition is performed on all optimizable parameters in the network, including convolutional layers and fully connected layers. When the convolutional layer is a high-dimensional tensor structure, the img2col algorithm should be used to convert it into a matrix.
[0026] Furthermore, in step 2, for The optimization objective is as shown in formula (1):
[0027] (1)
[0028]
[0029] The local data of the i-th participant is and , Input data for photovoltaic forecasting, The corresponding photovoltaic power generation to be predicted; the initial global model received by the participants is The loss of local data on this model is Furthermore, the parameters fixed in the constraints of the Grassmann manifold are... ;
[0030] In formula (1) The update method is shown in formula (2).
[0031] (2)
[0032] in, The learning rate, representing the optimization of manifold parameters, is used in some embodiments of the present invention. Take 0.01; Indicates loss to The gradient, in some embodiments of the present invention It is calculated using the automatic differentiation mechanism of the PyTorch framework.
[0033] Furthermore, in step 3, each participant first sets the Grassman manifold parameters. Left multiplication Convert back to conventional photovoltaic model parameters Then the parameters of the conventional photovoltaic model were compared with those of the global model. Perform mask aggregation, as shown in formula (3);
[0034] (3)
[0035] in, This represents the learnable mask for the i-th participant during the t-th round of training, where 1 represents a full-1 tensor of the same dimension as the mask. This represents element-wise multiplication. In some embodiments of the invention, the learnable mask is initialized as a tensor of all 1 / 2 and is progressively optimized during the training process.
[0036] Furthermore, in step 4, the nuclear norm is used to constrain... The rank of the rank is given by the specific optimization objective as shown in formula (4).
[0037] (4)
[0038] in, express nuclear norm number, The correction factor, representing the importance of the nuclear norm, is set to 0.01 in the optimization of cross-regional photovoltaic power generation prediction. This represents the loss of the model on local data after aggregating learnable masks;
[0039] In formula (4) for The update method is shown in formula (5):
[0040] (5)
[0041] in, The learning rate is represented in some embodiments of the present invention. Take 0.01, express right The gradient.
[0042] Furthermore, in step 5, the server first calculates the mean and standard deviation of all participant models; when more than 20% of the model weights fall outside the three standard deviation range, the model is considered an abnormal model and is discarded during aggregation; for the remaining normal models, the server takes the average of the parameters of these models as the new global model, as shown in formula (6): where I is the number of participants;
[0043] (6)
[0044] Where I represents the number of participants.
[0045] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects.
[0046] This invention protects the data privacy of participants in different regions. Through a federated learning algorithm, it ensures that data does not leave the local system when predicting cross-regional power generation; instead, it achieves model-level information fusion through masked aggregation. This approach can aggregate information from different regions while complying with data privacy regulations in various countries, thereby improving model prediction performance.
[0047] This invention can accurately adapt to personalized needs. Because variables such as lighting conditions vary significantly across different regions, a personalized federated learning algorithm is needed to build a unique model for each participant. In this way, the invention can effectively capture and utilize personalized information, overcoming the limitation of traditional federated learning algorithms where the global model is difficult to adapt to personalized needs.
[0048] This invention improves the performance of aggregation models through learnable masks. Specifically, by assigning different aggregation coefficients to each parameter of the photovoltaic prediction model in different regions, more accurate model parameter aggregation can be achieved. Compared with the simple weighted summation method in traditional federated learning aggregation algorithms, the learnable mask method of this invention can better adapt to the characteristics of different parameters, thereby improving the overall performance of the model.
[0049] This invention achieves a balance between computational efficiency and model accuracy. Specifically, for the learnable masks held by photovoltaic prediction models in different regions, this invention introduces a nuclear norm to constrain their computational complexity. By constraining the rank of the mask, this invention can significantly reduce computational complexity while ensuring that model performance is almost unaffected. This method achieves an effective balance between computational efficiency and model accuracy, thereby improving the practicality and feasibility of the method.
[0050] This invention reduces the optimization difficulty and improves optimization efficiency. Specifically, the parameters of photovoltaic prediction models for different regions lie on the Grassmann manifold, transforming the optimization problem of model parameters into a manifold optimization problem. By using simple manifold parameters to represent complex parameters to be optimized, this invention significantly reduces the optimization difficulty. This method makes the optimization process more efficient and stable, improving the optimization effect and convergence speed of the model. Attached Figure Description
[0051] Figure 1 This is a flowchart of the cross-regional photovoltaic power generation prediction method based on federated learning of the present invention;
[0052] Figure 2 This is a model architecture diagram of the present invention. Detailed Implementation
[0053] The invention will now be further explained with reference to the accompanying drawings.
[0054] This invention takes the construction of cross-regional photovoltaic power generation prediction as an example, and analyzes the photovoltaic power plant dataset of Bridgestone Changzhou as an example. This dataset contains data from January to July 2018, with one data point every 5 minutes. Specific features include power generation, temperature, and irradiance. Using the method of this invention, a mean square error of 0.1129 can be achieved in time series prediction. It is worth noting that this invention focuses on solving the problems of cross-regional photovoltaic data privacy protection, improving the accuracy of power generation prediction, and effectively controlling the computational load. In terms of data privacy protection, this invention adopts federated learning technology to ensure that data is stored locally and does not leak, thereby effectively maintaining the data security of different regions. Regarding the accuracy of power generation prediction, this invention introduces learnable mask aggregation for photovoltaic prediction models of different regions, thereby improving the accuracy of prediction results. In terms of computational load control, this invention innovatively uses the nuclear norm to constrain the rank of the learnable mask, significantly reducing the computational load and improving computational efficiency with almost no loss of model performance. In addition, this invention focuses on solving the problem of excessive optimization difficulty when simultaneously optimizing the mask and model parameters. To address this, this invention proposes that the photovoltaic prediction model parameters for different regions lie on the Grassmann manifold. Based on this, the invention reduces the number of parameters to be optimized, effectively lowering the complexity and difficulty of the optimization process, thereby achieving efficient optimization of the model parameters. The technical solution of this invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0055] Step 1: In the initialization phase, each region participating in photovoltaic power generation prediction corresponds to a participant. Furthermore, a server needs to be deployed for cross-regional parameter aggregation. Specifically, The input data for photovoltaic (PV) forecasting specifically includes historical PV power time series data and meteorological data. The historical PV power time series data contains three hours of historical data, with each half-hour period corresponding to a specific time point. The meteorological data includes temperature and solar radiation intensity. Before inputting the data into the model, missing values need to be filled using K-nearest neighbors, where the number of neighbors can be set to 1. After missing value filling, the local data is normalized using mean-variance. This represents the photovoltaic power generation to be predicted for the next half hour. Subsequently, the server will initialize the same global model. Distributed to each participant, where t represents the round number and g indicates that the model is a global model. Specifically, a global photovoltaic prediction model. This is a time-series prediction model based on convolutional neural networks and long short-term memory networks. The specific network structure is shown in the attached figure. Figure 2 As shown. For the convolutional neural network, it contains three one-dimensional convolutional layers with 16, 32, and 64 channels respectively, a kernel size of 3, a stride of 1, and 90, 20, and 50 filters respectively, while using a pooling size of 2. ReLU is used as the activation function. For the long short-term memory network, it contains 100 units with a dropout rate of 0.1 and ReLU as the activation function. Historical photovoltaic data is processed through the convolutional neural network to obtain historical photovoltaic features, while meteorological data is processed through a fully connected network layer to obtain meteorological features. The output dimension of the fully connected network layer is 64. Subsequently, the historical photovoltaic features and meteorological features are concatenated and input into the long short-term memory network. The temporal prediction result of this network is then input into a single-layer fully connected neural network to obtain the specific prediction result. Furthermore, the batch size of this model is set to 750.
[0056] Step 2: After receiving the initialized photovoltaic prediction model, each participant trains the model using local data and optimizes the local model parameters. The structure of the local photovoltaic prediction model is identical to that of the global photovoltaic prediction model. It is worth noting that the model parameters should lie on a Grassmanifold. The reason for choosing this manifold is that deep neural networks contain a large number of redundant variables, and the actual effective parameters often reside in a subspace of the original parameter space. The Grassmanifold can describe the effective parameters in the subspace, thus reducing the difficulty of model optimization. Therefore, unlike traditional deep learning which requires optimizing all parameters, this invention only needs to optimize the parameters on the manifold. To achieve this, the model needs to be optimized first. Perform QR decomposition to obtain the orthogonal matrix Q and the upper triangular matrix R. It is worth noting that this invention requires QR decomposition of all optimizable parameters in the network, including convolutional and fully connected layers. When the convolutional layer is a high-dimensional tensor structure, the img2col algorithm should be used to convert it into a matrix. Subsequently, fix the orthogonal matrix Q, and use the upper triangular matrix R as the manifold parameter of the i-th participant in the t-th round of training. . The specific optimization objective is shown in formula (1):
[0057] (1)
[0058]
[0059] The local data of the i-th participant is and The global model received by the participants is The loss of local data on the model is This loss can be the mean squared error in photovoltaic forecasting. Furthermore, the parameters fixed in the constraints of the Grassmann manifold are... ;
[0060] For photovoltaic prediction models The update method is shown in formula (2):
[0061] (2)
[0062] in, This represents the learning rate when optimizing parameters; in a cross-regional photovoltaic power generation prediction system, it can be taken as 0.01. The mean square error L represents the loss. The gradient is calculated using the automatic differentiation mechanism of the PyTorch framework.
[0063] Step 3: The photovoltaic prediction models of each participant are locally masked. Specifically, each participant first performs mask aggregation on the Grassmann manifold parameters. Left multiplication Convert back to conventional photovoltaic model parameters Then the parameters of the conventional photovoltaic model were... With global model Mask aggregation is performed, as shown in formula (3):
[0064] (3)
[0065] in, This represents the learnable mask for the i-th participant during the t-th round of training, where 1 represents a full-1 tensor of the same dimension as the mask. This represents element-wise multiplication. It's worth noting that the learnable mask is initialized as a tensor of all 1s and 2s and is progressively optimized during training. In cross-regional photovoltaic power generation prediction systems, The initialization can use a model with the same dimension as the global model, where all elements are... The matrix.
[0066] Step 4: Each participant optimizes the learnable mask in photovoltaic prediction. In this case, the nuclear norm is needed to constrain [the situation]. The rank of is given by the specific optimization objective as shown in formula (4):
[0067] (4)
[0068] in, express The nuclear norm number, and The correction factor representing the importance of the nuclear norm is set to 0.01 in the optimization of cross-regional photovoltaic power generation prediction.
[0069] Learnable masks for photovoltaic prediction The update method is shown in formula (5):
[0070] (5)
[0071] in, The learning rate can be set to 0.01 in cross-regional photovoltaic power generation prediction optimization. express right The gradient.
[0072] Step 5: Each participant submits the photovoltaic prediction model parameters. The parameters are sent to the server, and after receiving the parameters from all participants, the server takes the average of these parameters as the new global model, as shown in formula (6):
[0073] (6)
[0074] Here, I represents the number of participants. In cross-regional photovoltaic power generation prediction systems, I is often between several hundred and several thousand, with the specific value determined by the number of regions.
[0075] Step 6: Repeat steps 2 through 5 until the global model is complete. The system can converge or loop to a specified number of iterations. In a cross-regional photovoltaic power generation prediction system, the specified number of iterations can be selected as 500.
[0076] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A cross-regional photovoltaic power generation prediction method based on federated learning, characterized in that: Includes the following steps: Step 1: Each region participating in photovoltaic power generation prediction corresponds to one participant, and a server is deployed for cross-regional parameter aggregation; each participant prepares local data containing historical photovoltaic data and meteorological data, where the historical photovoltaic data includes photovoltaic power time series, and preprocesses the local data; the server distributes the initial global photovoltaic prediction model to each participant. , where t represents the number of rounds, and g indicates that the model is a global model; Step 2: After each participant receives the initialized photovoltaic prediction model, it is used as a local model, and local data is used to optimize the local model parameters. Each participant projects the local model parameters onto the Grassman manifold. First, QR decomposition is performed on the model parameters of each layer of the local model to obtain the orthogonal matrix Q and the upper triangular matrix R. Then, the orthogonal matrix Q is fixed, and the upper triangular matrix R is used as the manifold parameters to be optimized for the i-th participant in the t-th round of training. ; For manifold parameters The optimization objective is as shown in formula (1): （1） The local data of the i-th participant is and , Input data for photovoltaic forecasting, The corresponding photovoltaic power generation to be predicted; the initial global model received by the participants is Local data is used to initialize the global model. The loss is Furthermore, the parameters fixed in the constraints of the Grassmann manifold are... ; In formula (1) The update method is shown in formula (2): （2） in, This represents the learning rate when optimizing manifold parameters; Indicates loss manifold parameters The gradient; Step 3: The photovoltaic prediction models of each participant are locally masked and aggregated. Specifically, each participant first sets the Grassman manifold parameters... Left multiplication Convert back to conventional photovoltaic model parameters Then the parameters of the conventional photovoltaic model were compared with those of the global model. Perform mask aggregation; Step 4: Each participant optimizes the learnable mask in photovoltaic prediction. ; Use nuclear norm to constrain The rank of is given by the specific optimization objective as shown in formula (4): （4） in, express nuclear norm number, The correction factor representing the importance of the nuclear norm number. This represents the loss of the model on local data after aggregating learnable masks; In formula (4) for learnable masks The update method is shown in formula (5): （5） in Indicates the learning rate. express For learnable masks The gradient; Step 5: Each participant submits the photovoltaic model parameters. The model is sent to the server, which detects abnormal models and then calculates the average of the parameters of the remaining models as the new global model. Step 6: Repeat steps 2 through 5 until... It converges or loops to a specified number of times.
2. The method according to claim 1, characterized in that: In step 1, the local data prepared by each participant is as follows: the local data of the i-th participant is... and , The input data for photovoltaic (PV) forecasting includes historical PV data and meteorological data; the historical PV data includes three hours of historical data, with each five-minute data point corresponding to a time point; the meteorological data includes temperature and solar radiation intensity. The photovoltaic power generation is to be predicted for the next half hour.
3. The method according to claim 1, characterized in that: In step 1, the preprocessing steps for the local data are as follows: fill the data with missing values based on K nearest neighbors, where the number of neighbors is 1. After filling the missing values, normalize the local data by mean-variance.
4. The method according to claim 1, characterized in that: In step 1, the global photovoltaic prediction model It is a time-series prediction model based on convolutional neural networks and long short-term memory networks. Historical photovoltaic data is processed through three one-dimensional convolutional layers to obtain historical photovoltaic features, and meteorological data is processed through a fully connected network layer to obtain meteorological features. Then, the historical photovoltaic features and meteorological features are concatenated and input into the long short-term memory network. Finally, the time-series prediction results of this network are input into a single-layer fully connected neural network to obtain the specific prediction results.
5. The method according to claim 4, characterized in that: The number of channels in the one-dimensional convolutional layer is 16, 32, and 64, the number of filters is 90, 20, and 50, the pooling size is 2, the convolutional kernel size is 3, the stride is 1, and the activation function is ReLU; the output dimension of the fully connected network layer is 64; the long short-term memory network contains 100 units, the dropout rate is set to 0.1, and the activation function is ReLU.
6. The method according to claim 1, characterized in that: In step 2, QR decomposition is performed on all optimizable parameters in the network, including convolutional layers and fully connected layers. When the convolutional layer is a high-dimensional tensor structure, the img2col algorithm should be used to convert it into a matrix.
7. The method according to claim 1, characterized in that: In step 3, the parameters of the conventional photovoltaic model are compared with those of the global model. Mask aggregation is performed, as shown in formula (3): （3） in, This represents the learnable mask for the i-th participant during the t-th round of training, where 1 represents a full-1 tensor of the same dimension as the mask. This represents element-wise multiplication; the learnable mask is initialized to a tensor of all 1 / 2 and is gradually optimized during training.
8. The method according to claim 1, characterized in that: In step 5, the server first calculates the mean and standard deviation of all participant models; when more than 20% of the model weights fall outside the three standard deviation range, the model is considered an abnormal model and is discarded during aggregation; for the remaining normal models, the server takes the average of the parameters of these models as the new global model, as shown in formula (6): where I is the number of participants; （6） Where I represents the number of participants.