Electric energy data intelligent acquisition terminal based on federal learning

Abstract

The invention relates to the technical field of electric meter online monitoring, in particular to an electric energy data intelligent acquisition terminal based on federal learning. The data acquisition and preprocessing module is used for acquiring electric energy data in a power grid in real time and carrying out de-noising and normalization preprocessing on the electric energy data; the model training module based on federated learning is used for training a neural network model for the preprocessed electric energy data by using local data so as to obtain a local model; the homomorphic encryption algorithm application module is used for encrypting and updating data and model parameters in a local data training process by adopting a homomorphic encryption algorithm; and the global model updating module is used for sending the neural network model subjected to multiple times of local data training and encryption updating to a central server, and optimizing and updating the neural network model through a weighted federated average aggregation algorithm. According to the invention, distributed data processing and machine learning technologies are combined to realize intelligent data analysis of electric power data, so that the intelligent level of electric power monitoring and fault early warning is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of on-line monitoring of electric meters, and particularly to an intelligent acquisition terminal for electric energy data based on federated learning. Background Art

[0002] With the rapid development of smart grid and Internet of Things technologies, the intelligent management of power systems has become an important development direction in the modern power industry. As a core component of the smart grid, smart meters are widely used in the collection, transmission, and monitoring of power data. In traditional power monitoring, data collection and processing of electric meters are mainly carried out through manual meter reading, regular inspections, etc. However, with the growth of electricity demand and the increase in the complexity of power grid management, traditional data collection methods are difficult to meet the requirements of modern power systems for real-time, accuracy, and security.

[0003] To address this challenge, smart meters are widely used, which can automatically collect and upload power data, support remote monitoring and data analysis. Smart meters are usually equipped with high-precision sensors, such as current, voltage, power, and temperature sensors, which can real-time monitor various operation indicators of the power grid and transmit the data to the background system for processing and analysis through wireless communication methods. Although this technology improves the efficiency of data collection, there are still many problems, especially in data privacy protection and distributed computing capabilities.

[0004] Current power monitoring systems usually rely on a centralized data processing mode, and all meter data needs to be uploaded to the central server for aggregation and analysis. However, with the increase in the number of devices, the centralized mode faces huge bandwidth pressure and computing burden, and there are certain security risks. Especially for users' electricity consumption data, centralized data storage and processing may bring the risk of data leakage. Therefore, how to improve the intelligence and real-time of power monitoring while ensuring data privacy has become an urgent technical problem to be solved.

[0005] Federated Learning, as an emerging distributed machine learning technology, provides a new idea for solving this problem. Federated Learning trains the model locally on the terminal device, avoiding uploading the original data to the cloud or server, thus effectively protecting user privacy. Each terminal device collaboratively learns and jointly optimizes the global model through local computing and local data processing without sharing the original data. In this way, not only the data leakage problem is avoided, but also the pressure of centralized data storage and computing is reduced.

[0006] The application of federated learning in smart grids enables power data acquisition terminals to achieve intelligent data analysis and fault detection without relying on large-scale data centers. Through this technology, smart meters can perform tasks such as local load forecasting and anomaly detection, while uploading the training results to the central server for aggregation and update. This distributed learning method not only improves the response speed and intelligence level of the power monitoring system but also greatly reduces the system's requirements for bandwidth and storage. Although federated learning provides a new solution for power data intelligent acquisition terminals, there are still some challenges in efficiently deploying and applying federated learning technology in power monitoring systems. For example, how to balance the computing load and data transmission efficiency due to limited local computing resources; how to ensure data security and privacy protection during the federated learning process; and how to cope with dynamic changes in the power grid environment, etc., are the current research focuses.

[0007] Therefore, how to combine federated learning technology to build an efficient, secure, and intelligent power data acquisition terminal has become one of the key directions for the development of power monitoring systems. Summary of the Invention

[0008] The purpose of the present invention is to combine the advantages of distributed data processing and machine learning to achieve intelligent data analysis in power monitoring systems without the need to centralize all data to the cloud or a central server, thereby reducing the pressure of data transmission and enhancing data privacy protection, so as to improve the intelligence level of power monitoring and fault warning, and a power data intelligent acquisition terminal based on federated learning is proposed.

[0009] The technical solution of the present invention: A power data intelligent acquisition terminal based on federated learning, comprising:

[0010] A data acquisition and preprocessing module that real-time collects power data in the power grid and performs denoising and normalization preprocessing on the power data;

[0011] A model training module based on federated learning that uses local data to train a neural network model on the preprocessed power data to obtain a local model;

[0012] A homomorphic encryption algorithm application module connected to the model training module based on federated learning, which uses the homomorphic encryption algorithm to encrypt and update the data and model parameters during the local data training process;

[0013] A global model update module, including a weighted federated averaging unit, a dynamic learning rate adjustment unit, and an aggregation and synchronization unit for the global model, which sends the neural network model that has undergone multiple local data trainings and encrypted updates to the central server, and optimizes and updates the neural network model through the weighted federated averaging aggregation algorithm.

[0014] Optionally, in the data acquisition and preprocessing module, the acquired electrical energy data x t =[[v t ,[[i t ,[[p t ,[[f t represents the voltage (v t ), current (i t ), power (p t ), and frequency (f t ) data corresponding to time t. After denoising and normalization preprocessing, each data point conforms to the following standardization formula:

[0015]

[0016] where μ x is the mean of the electrical quantity data x t , and σ x is the standard deviation of the electrical quantity data x t . is the standardized data.

[0017] Optionally, the neural network model includes an input layer, a hidden layer, and an output layer connected in sequence from input to output. The output of the neural network model is the prediction result:

[0018]

[0019] where f 0 , f 1 , and f 2 are the activation functions of each layer, and X is the input data; the weights and biases of each layer are a 0 , a 1 , a 2 , respectively, and the calculation is expressed as matrix operations;

[0020] The intermediate calculation result t i of each layer is expressed as:

[0021] t i =f i (α i t i-1 +b i )

[0022] where t i is the calculation result of the i-th layer, α i and b i are the weight and bias of this layer respectively, and f i () is the activation function.

[0023] Optionally, in the homomorphic encryption algorithm application module, the gradient update of the neural network model is protected by homomorphic encryption. For each local neural network, the objective function J(θ) is defined as:

[0024]

[0025] where h θ (x i ) is the prediction of the model for the input sample x i , y i is the actual label, θ i are the model parameters, and λ is the regularization coefficient.

[0026] Optionally, the loss function and gradient calculation of the homomorphic encryption algorithm are encrypted. Let E[·] represent the homomorphic encryption operation and N represent the number of samples. Then the encrypted loss function of each local model is:

[0027]

[0028] where the gradient calculation is completed by the backpropagation algorithm and protected by homomorphic encryption at the same time. The gradient update formula is:

[0029]

[0030] where η is the learning rate, is the gradient of the loss function with respect to the weight θ j .

[0031] Optionally, the gradient update formula for each local model is:

[0032]

[0033] where θ j are the model parameters, λ is the regularization coefficient, is the differential symbol, l j represents the loss function of the j-th local model, and E[L] represents the global loss function in the encrypted state.

[0034] Optionally, the weighted federated averaging aggregation algorithm adjusts the weights of each local model according to the local data volume of each subsidiary or data center. Let the weights w j of each local model and the corresponding model parameters θj, the update formula of the global model in the neural network model is:

[0035]

[0036] where |D j | represents the local data volume of the j-th subsidiary or data center, |D irepresents the local data volume of the \(i\)-th subsidiary or data center, represents the parameters of the local model, \(\theta\) (k+1) is the updated global model parameters.

[0037] Optionally, the dynamic learning rate adjustment unit dynamically adjusts the learning rate \(\eta\) according to the loss change during the training process. The update formula for dynamic learning rate adjustment is as follows:

[0038]

[0039] where \(\eta\) (k+1) is the current learning rate, LossChange represents the change in loss in the current iteration, and \(\alpha\) is the adjustment factor.

[0040] Optionally, the global model aggregation and synchronization unit is used to synchronize multiple local data training tasks in stages during the training process. When aggregating the global model, real-time synchronization of global parameters is achieved through a distributed computing framework. The final parameters of the global model are synchronously updated through the following formula:

[0041]

[0042] Compared with the prior art, the present application includes at least one of the following beneficial technical effects:

[0043] 1. Data privacy protection driven by federated learning: The present invention adopts a federated learning framework to ensure that while the power data acquisition terminal independently trains the model locally, it does not need to upload the data to the central server, thus effectively protecting the privacy and security of user data. Federated learning trains on local data and updates the encrypted model, so that the sensitive data of each terminal is not leaked, but is fused with the global model through encrypted gradient updates. This privacy protection mechanism improves the security of power data and avoids potential risks brought by large-scale centralized data storage.

[0044] 2. Efficient distributed model training and update: Utilizing the centralized training feature of federated learning, each power data acquisition terminal independently trains the neural network model locally, and only uploads the encrypted gradient or model update, thus reducing the burden on the central server and improving the efficiency of the training process. In addition, through the weighted federated average aggregation algorithm, weighted processing of model updates is performed according to the local data volume of each subsidiary or data center, making the global model more accurate, especially effectively optimizing the training effect in the case of uneven data volume.

[0045] 3. Security Assurance Combining Homomorphic Encryption and Federated Learning: The present invention combines homomorphic encryption technology and a federated learning model, which not only ensures the privacy protection of data during the training process but also enables data to be processed and updated in an encrypted state. Even in a distributed environment, the gradients and model parameters transmitted by all parties through homomorphic encryption cannot be accessed or tampered with by unauthorized third parties. This federated learning framework based on homomorphic encryption provides an efficient and secure solution for power data acquisition terminals, meeting the growing demands of modern power systems for data privacy and security. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a system block diagram of the present invention;

[0047] Figure 2 is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] The technical solutions of the present disclosure will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present disclosure. The components of the embodiments of the present disclosure described and shown herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the claimed present disclosure, but merely represents selected embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present disclosure without creative efforts fall within the scope of protection of the present disclosure.

[0049] As Figure 1 shown, an intelligent power data acquisition terminal based on federated learning of the present invention mainly consists of a data acquisition and preprocessing module, a model training module based on federated learning, a homomorphic encryption algorithm application module, and a global model update module. Among them, the data acquisition and preprocessing module collects power data in the power grid in real time and performs denoising and normalization preprocessing on the power data; the model training module based on federated learning uses local data to train a neural network model for the preprocessed power data to obtain a local model. The homomorphic encryption algorithm application module uses the homomorphic encryption algorithm to encrypt and update the data and model parameters during the local data training process. The communication security is ensured through the TLS and AES encryption protocols to prevent data leakage. During the federated learning process, in combination with homomorphic encryption technology, perturbation processing and encryption are performed on the uploaded model updates to further ensure data privacy and avoid revealing users' electricity consumption behaviors. In this way, both the data security is improved and the privacy protection during the model training process is ensured.

[0050] The global model update module, including a weighted federated averaging unit, a dynamic learning rate adjustment unit, and an aggregation and synchronization unit for the global model, sends the neural network model that has undergone multiple local data trainings and encrypted updates to the central server, and optimizes and updates the neural network model through the weighted federated averaging aggregation algorithm. The update of the machine learning model adopts the federated learning mechanism. Through local training (such as LSTM, Isolation Forest, etc.), the device uploads the model parameters (weights and gradients) to the server. The server uses the weighted federated averaging (FedAvg-W) aggregation algorithm to aggregate the model updates of each terminal, optimizes the global model, and sends the updated model to the terminal to ensure the high efficiency, intelligence, and security of the power monitoring system. The present invention innovatively applies federated learning to the power data acquisition terminal. Each electric meter device uses LSTM for load forecasting or autoencoders for fault detection locally, and only uploads the model parameters (such as weights) instead of the original data. The weighted federated averaging aggregation algorithm is used to aggregate the model parameters of each terminal, update the global model and send it to the terminal, reducing the data transmission pressure while improving the intelligence and computing efficiency of the system.

[0051] Please refer to Figure 2 , the specific operation process of an intelligent power data acquisition terminal based on federated learning of the present invention is as follows:

[0052] 1. Acquisition and preprocessing of multi-sensor data of electric meters

[0053] The power data acquisition terminal collects various data in the power grid in real time through a group of sensors (such as voltage sensors, current sensors, temperature sensors), including voltage, current, power, frequency, etc. The collected data x t =[v t , i t , p t , f t represents the voltage (v t ), current (i t ), power (p t ), and frequency (f t ) data corresponding to the moment t. To ensure the high quality of the data, the present invention also includes denoising processing and normalization of the data, so that each data point conforms to the following normalization formula:

[0054]

[0055] Among them, μ x is the mean value of the power data x t , σ x is the standard deviation of the power data, is the normalized data.

[0056] 2. Model training based on federated learning

[0057] In the present invention, each power data acquisition terminal (such as an electric meter) uses local data to train a neural network model. The network adopts a traditional three-layer structure (input layer, hidden layer, output layer), and the output of the model is the prediction result:

[0058]

[0059] where f 0 , f 1 , and f 2 are the activation functions of each layer respectively, and X is the input data. The weights and biases of each layer are a 0 , a 1 , a 2 respectively, and the calculation can be expressed as matrix operations. The intermediate calculation result t i of each layer is given by the following formula:

[0060] t i = f i (α i t i-1 + b i )

[0061] where t i is the calculation result of the i-th layer, α i and b i are the weight and bias of this layer respectively, and f i () is the activation function.

[0062] 3. Application of Homomorphic Encryption Algorithm

[0063] To protect the privacy and security of training data, the present invention adopts homomorphic encryption technology when each power data acquisition terminal conducts local training. Each subsidiary company and the network-level data center use homomorphic encryption to encrypt the data and model parameters during their training processes to ensure that even if the data is transmitted, it will not be leaked.

[0064] Specifically, the gradient update of the neural network model is protected through homomorphic encryption. For each local neural network, the objective function J(θ) is defined as:

[0065]

[0066] where h θ (x i ) is the prediction of the model for the input sample x i , y i is the actual label, θ i is the parameter of the local model, and λ is the regularization coefficient. To ensure privacy, all model update processes are carried out in an encrypted state.

[0067] 3.1 Homomorphic Encryption Loss Function

[0068] In the homomorphic encryption algorithm, the loss function and gradient calculation are encrypted. Let E[·] represent the homomorphic encryption operation and N represent the number of samples. Then the encrypted loss function of each local model is as follows:

[0069]

[0070] This means that all operations in the loss function, including predictions and labels, need to be homomorphically encrypted to ensure that data is not leaked during transmission.

[0071] 3.2 Gradient Update and Homomorphic Encryption

[0072] In the present invention, the gradient is calculated through the backpropagation algorithm and protected using homomorphic encryption. The gradient update formula is as follows:

[0073]

[0074] where η is the learning rate, is the gradient of the loss function with respect to the weight θ j . In the homomorphic encryption environment, all gradient calculations are performed in the encrypted state, ensuring data privacy.

[0075] In the homomorphic encryption environment, the loss function and gradient calculation of the local model are encrypted. For each subsidiary N and network-level data center C, their loss functions are respectively expressed as:

[0076]

[0077] where E[h j (x i )] and E[h C (x i )] are the encrypted prediction results of the subsidiary and the network-level data center respectively.

[0078] Homomorphic encryption enables gradient calculation without decryption, ensuring privacy protection among all participating parties (subsidiaries and network-level data centers). The gradient update formula for each local model is:

[0079]

[0080] where θ j is the model parameter, λ is the regularization coefficient, and is the differential symbol, l jLet \(L_j\) denote the loss function of the \(j\)-th local model, and \(E[L]\) denote the global loss function in the encrypted state. The gradient of the loss function with respect to each parameter is calculated, and all operations are performed in the encrypted state.

[0081] 4. Global Model Aggregation and Update

[0082] 4.1 Weighted Federated Averaging

[0083] After multiple local trainings and encrypted gradient updates, each power data acquisition terminal (subsidiary and network-level data center) submits its encrypted model update to the central server. The server aggregates through the FedAvg algorithm. In the traditional FedAvg method, all local model updates are treated equally. However, different local data volumes and data qualities may lead to unbalanced effects. Therefore, to optimize the training effect of the global model, the present invention proposes a weighted federated averaging (FedAvg-W) aggregation algorithm, which adjusts the weight of each local model according to the local data volume of each subsidiary or data center. Given the weight \(w_j\) of each local model j and the corresponding model parameters \(\theta_j\), the update formula for the global model is:

[0084]

[0085] where \(|D_j|\) j represents the local data volume of the \(j\)-th subsidiary or data center, \(|D_i|\) i represents the local data volume of the \(i\)-th subsidiary or data center, \(\theta_j\) represents the parameters of the local model, and \(\theta\) (k+1) is the updated global model parameter. This formula represents weighted averaging, giving higher weights to the model updates of subsidiaries or data centers with larger local data volumes.

[0086] 4.2 Dynamic Learning Rate Adjustment

[0087] To further improve the training effect of the global model, the present invention combines a dynamic learning rate adjustment method. By dynamically adjusting the learning rate \(\eta\) according to the loss change during training to avoid overfitting and accelerate convergence. The update formula for dynamic learning rate adjustment is as follows:

[0088]

[0089] where \(\eta\) (k+1) is the current learning rate, LossChange represents the change in loss in the current iteration, and \(\alpha\) is the adjustment factor (usually set to 0.1 - 0.5), which is optimized according to experiments. This strategy slows down the oscillation during training by reducing the learning rate and avoids premature convergence.

[0090] 4.3 Aggregation and Synchronization of the Global Model

[0091] To improve the training efficiency and system stability, a distributed synchronization optimization method is adopted, that is, multiple local training tasks can be carried out in parallel and staged synchronization is performed during the training process. When aggregating the global model, real-time synchronization of global parameters is achieved through an efficient distributed computing framework (such as MPI, Horovod or ParameterServer) to ensure the stability and accuracy of training. The final parameters of the global model can be synchronously updated through the following formula:

[0092]

[0093] This formula represents that after each model update, the training models of each subsidiary company and the model of the network-level data center will merge and update the global model.

[0094] In addition, it should be further noted that in the power dispatching center of a certain city, an intelligent power data acquisition terminal system based on federated learning is deployed on multiple substations and important power equipment to achieve power load forecasting and fault diagnosis. The system uses distributed intelligent electricity meters equipped with high-precision voltage, current, power and frequency sensors to collect key data of the power grid in real time. Through the training of local neural network models, the system can perform real-time analysis and prediction on the power usage of each meter area. After data preprocessing, the system will perform data normalization and denoising to improve the data quality and training accuracy.

[0095] The system adopts a federated learning framework. On the premise of ensuring user data privacy, through distributed local training and global model aggregation, the accuracy of power load forecasting is greatly improved. Each meter terminal independently trains a neural network model and uses encryption technology to protect the model training process to ensure the security of data during transmission. By weighted aggregating the models of each meter, the system can comprehensively consider the grid load characteristics of different regions, optimize the global prediction model, and thus provide more accurate load forecasting data for the power dispatching center.

[0096] In addition, the fault diagnosis function of the system has also been significantly improved. Based on real-time data collection and intelligent analysis, the system can timely detect abnormal situations in the power grid, such as current overload, voltage fluctuation or equipment overheating, etc., and help operation and maintenance personnel take necessary maintenance measures through early warnings. Through the collaborative training of federated learning, the system can continuously optimize the fault diagnosis model, improve the sensitivity and accuracy of fault detection, and thus reduce the fault occurrence rate of the power system.

[0097] Overall, the intelligent power data acquisition terminal system based on federated learning effectively improves the accuracy of power load forecasting, enhances the precision of fault diagnosis, and ensures data privacy protection. Through the distributed training and encryption protection of federated learning, the system achieves intelligent power grid management, helping power companies optimize resource allocation and maintenance strategies while ensuring data security, and significantly improving the stability and operation efficiency of the power grid.

[0098] The above specific embodiments are merely several alternative embodiments of the present invention. Based on the technical solution of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. An intelligent power data collection terminal based on federated learning, characterized in that: include: A data acquisition and preprocessing module collects electric energy data in the power grid in real time and performs denoising and normalization preprocessing on the electric energy data; A model training module based on federated learning, training a neural network model using local data on the preprocessed electric energy data to obtain a local model; A homomorphic encryption algorithm application module is connected to the model training module based on federated learning, and uses a homomorphic encryption algorithm to encrypt and update data and model parameters in the local data training process; The global model update module includes a weighted federated averaging unit, a dynamic learning rate adjustment unit, and a global model aggregation and synchronization unit. The neural network model that has undergone multiple local data training and encryption updates is sent to a central server, and the neural network model is optimized and updated through a weighted federated averaging aggregation algorithm.

2. According to claim 1, the intelligent power data collection terminal based on federated learning is characterized in that: In the data collection and preprocessing module, the collected electric energy data x t =[v t ,i t ,p t ,f t ] represents the voltage corresponding to time t (v t ), current (i t ), power (p t ) and frequency (f t ) data, after denoising and normalization preprocessing, each data point conforms to the following standardized formula: Among them, μ x The power data x t The mean value, σ x The power data x t The standard deviation of is the standardized data.

3. According to the federated learning-based intelligent power data collection terminal of claim 1, it is characterized in that: The neural network model includes an input layer, a hidden layer and an output layer connected in sequence from input to output. The output of the neural network model To predict the results: Among them, f0, f1 and f2 are the activation functions of each layer respectively, X is the input data; the weight and bias of each layer are a0, a1, a2 respectively, and the calculation is expressed as a matrix operation; The intermediate calculation results of each layer t i It is expressed as: t i =f i (α i t i-1 +b i ) Among them, t i is the calculation result of the i-th layer, α i and b i are the weight and bias of the layer respectively, f i () is the activation function.

4. The intelligent power data collection terminal based on federated learning according to claim 1 is characterized in that: In the homomorphic encryption algorithm application module, the gradient update of the neural network model is protected by homomorphic encryption. For each local neural network, the objective function J(θ) is defined as: Among them, h θ (x i ) is the model's response to the input sample x i The prediction of i is the actual label, θ i is the model parameter and λ is the regularization coefficient.

5. The intelligent power data collection terminal based on federated learning according to claim 4 is characterized in that: The loss function and gradient calculation of the homomorphic encryption algorithm are encrypted. Let E[·] represent the homomorphic encryption operation and N represent the number of samples. Then the encrypted loss function of each local model is: The gradient calculation is completed through the back-propagation algorithm and protected by homomorphic encryption. The gradient update formula is: Where η is the learning rate, is the loss function with respect to weight θ j gradient.

6. The intelligent power data collection terminal based on federated learning according to claim 5 is characterized in that: The gradient update formula for each local model is: Among them, θ j is the model parameter, λ is the regularization coefficient, is the differential symbol, l j represents the loss function of the j-th local model, and E[L] represents the global loss function in the encrypted state.

7. The intelligent power data collection terminal based on federated learning according to claim 6 is characterized in that: The weighted federated average aggregation algorithm adjusts the weight of each local model according to the amount of local data of each subsidiary or data center. Suppose the weight of each local model is w j and the corresponding model parameters θ j , the update formula of the global model in the neural network model is: Among them, |D j | represents the local data volume of the jth subsidiary or data center, |D i | represents the amount of local data of the ith subsidiary or data center, represents the parameters of the local model, θ (k+1) are the updated global model parameters.

8. The intelligent power data collection terminal based on federated learning according to claim 7 is characterized in that: The dynamic learning rate adjustment unit dynamically adjusts the learning rate η according to the loss change during the training process. The update formula of the dynamic learning rate adjustment is as follows: Among them, η (k+1) is the current learning rate, LossChange represents the change in loss in the current iteration, and α is the adjustment factor.

9. The intelligent power data collection terminal based on federated learning according to claim 8, characterized in that: The aggregation and synchronization unit of the global model is used to synchronize multiple local data training tasks in stages during the training process. When the global model is aggregated, the real-time synchronization of global parameters is achieved through the distributed computing framework. The final parameters of the global model are synchronized and updated using the following formula:

Images