Federated learning line loss calculation system and method for distribution network

Abstract

The present invention relates to the field of power systems, and in particular to a federated learning line loss calculation system and method for distribution networks, comprising: a substation edge learning system using an improved LSTM-GCN hybrid network to establish a model, and combining it with a privacy-preserving federated learning algorithm to obtain a substation line loss model; a local data center aggregates local historical data and real-time data, establishes feature engineering for line losses, and performs cluster classification; based on state estimation and topology identification, the line loss rate of each substation is obtained based on the substation line loss model; a blockchain system establishes a consensus and mutual trust mechanism between the power grid business layer and the power grid data layer to ensure the legitimacy and validity of data collection across the entire network, and provides functional support for consensus interaction, data exchange, state information interaction, and model interaction; a cloud service center receives the calculation information and clustering center of the substation edge learning system, and uses line loss rate feature engineering and cluster analysis to conduct a comprehensive analysis with substation line loss rate data to effectively protect data privacy.

Description

Technical Field

[0001] The present invention relates to the field of power systems, and in particular to a federated learning line loss calculation system and method for distribution networks, which are applied to the field of distribution network line loss management and optimization. Background Art

[0002] The line loss rate of a distribution network is a key indicator of grid efficiency. Accurately calculating this rate is crucial for energy conservation, emission reduction, and cost control. Traditional line loss calculation methods rely primarily on centralized computing architectures, requiring the collection of power data from each substation to a central server for processing. This presents challenges such as data security risks, high communication pressure, and low computing efficiency. Especially with the increasing emphasis on user data privacy, traditional centralized computing architectures struggle to meet the data security and privacy requirements of modern distribution networks.

[0003] Furthermore, distribution network equipment is widely distributed, and data collection often faces issues such as missing data and frequent topology changes, posing challenges to line loss calculation. Existing technologies struggle to effectively address these issues, resulting in low line loss calculation accuracy and an inability to provide accurate decision support for grid management.

[0004] With the development of artificial intelligence (AI) technology, the application of deep learning in power systems is increasing. However, traditional deep learning models require centralized data training, making it difficult to address data privacy issues. Federated learning, an emerging distributed machine learning paradigm, allows multiple parties to jointly train models while protecting data privacy, providing a new technical approach for distribution network line loss calculation. However, existing federated learning solutions, when applied to distribution network line loss calculation, still face challenges such as low model accuracy, low training efficiency, and imperfect security mechanisms.

[0005] Therefore, there is an urgent need for a distribution network line loss calculation system and method that can protect data privacy, improve computing efficiency, and ensure computing accuracy. Summary of the Invention

[0006] The purpose of the present invention is to provide a federated learning line loss calculation system and method for distribution networks, so as to solve the problems existing in the prior art such as insufficient data privacy protection, low computing efficiency, and low accuracy.

[0007] The present invention proposes a federated learning line loss calculation system for distribution networks, comprising:

[0008] The substation edge learning system is used to obtain substation voltage, current, power, and power factor data. It uses an improved LSTM-GCN hybrid network to build a model and combines it with a privacy-preserving federated learning algorithm to obtain a substation line loss model.

[0009] A local data center is connected to the substation edge learning system for aggregating local historical data and real-time data to the local data center, aggregating the substation data to establish feature engineering of line loss and perform cluster classification, and obtaining the line loss rate of each substation based on the substation line loss model according to state estimation and topology identification;

[0010] The blockchain system is connected to the substation edge learning system and the local data center to establish a consensus and mutual trust mechanism between the power grid business layer and the power grid data layer, ensuring the legitimacy and validity of data collection across the entire network, while providing functional support for consensus interaction, data exchange, status information interaction, and model interaction;

[0011] A cloud service center is in communication with the local data center and the blockchain system, and is configured to receive the calculation information and cluster center of the substation edge learning system, and then use line loss rate feature engineering and cluster analysis to conduct a comprehensive analysis with the substation line loss rate data to obtain line loss assessment decision recommendations for each substation;

[0012] Among them, the calculation information of the substation edge learning system includes substation line loss model information and gradient information.

[0013] Preferably, the blockchain system includes:

[0014] Communication protocol sublayer, used to realize information interaction;

[0015] System security protection sublayer, used for security protection;

[0016] The edge computing sublayer is used to realize computing information interaction.

[0017] Preferably, the area edge learning system includes:

[0018] A communication protection module, used to encrypt the transmission data before transmission and transmit it to the cloud service center, and to locally maintain the decrypted data after decryption;

[0019] Local data processing module, used to provide services for collecting and updating power user measurement data, distribution and transformation topology information, and historical model parameters;

[0020] An encryption communication module, used to encrypt data collected from the local data processing module and decrypt information transmitted by the cloud service center;

[0021] The local computing module is used to complete the upload of edge computing node line loss models and federated learning parameters, the synchronous transmission of user meter measurement data, topology information, model updates, and the integrated training of local data.

[0022] Preferably, the local data center includes:

[0023] Feature engineering module, used to build feature sets for line loss calculation based on aggregated data;

[0024] A cluster analysis module, used for clustering and classifying the feature set;

[0025] The state estimation module is used to calculate the line loss rate of each substation based on the feature set and topology identification information in combination with the substation line loss model.

[0026] Preferably, the cloud service center includes:

[0027] A model aggregation module, configured to receive and aggregate the model information and gradient information uploaded by the edge learning system in the substation area;

[0028] a gradient evaluation module, configured to perform quality evaluation on the gradient information and perform adaptive weight allocation based on the evaluation result;

[0029] A cluster analysis module, configured to perform cluster analysis based on the gradient information to identify similar station area characteristics;

[0030] The decision support module is used to generate decision recommendations for line loss assessment in substations based on cluster analysis results and line loss rate data.

[0031] Preferably, the improved LSTM-GCN hybrid network includes:

[0032] LSTM time series module, used to process the time series variation characteristics of the substation voltage, current, power, and power factor;

[0033] GCN topology module, used to process the topological characteristics of the connection relationship between distribution network equipment;

[0034] A feature fusion layer, used to fuse the temporal change feature with the topological feature to form a comprehensive representation;

[0035] The line loss prediction layer is used to calculate the line loss prediction value based on the comprehensive representation.

[0036] Preferably, the privacy-preserving federated learning algorithm includes:

[0037] Gradient encryption module, used to encrypt locally calculated model gradients;

[0038] A differential privacy module, used to add calibration noise to the model gradient to achieve differential privacy protection;

[0039] A gradient compression module, used to compress the encrypted gradient to reduce the amount of transmitted data;

[0040] The secure aggregation module is used to aggregate multi-party gradient information while protecting individual data privacy.

[0041] Preferably, the adaptive weight allocation is implemented based on the following process:

[0042] Initialize the federated learning environment and assign weights to each client. This allows each client in each region to become a central client in subsequent federated learning. Central clients have higher weights than ordinary clients to address the potential degradation of model gradient values.

[0043] The central client sends a training request, and each client calculates the training information separately;

[0044] When a certain training time is met, the central client collects the training information of each client, combines the training information on the central client, and redistributes the weights;

[0045] Each client updates its own model weights according to the weights assigned by the central client and issues the next training request.

[0046] Preferably, the line loss calculation process of the state estimation module includes:

[0047] Use virtual measurement nodes based on digital twins to supplement missing current and voltage data;

[0048] Topology identification based on state estimation;

[0049] Based on the feature engineering of current voltage, current, active power or reactive power, the state estimation of the substation model and topology identification is performed;

[0050] Use Kalman filtering technology to correct the parameters of state estimation;

[0051] The topological information of the power equipment is obtained by extracting the characteristic quantities collected by the power electronic equipment, and the line loss rate of the substation area is constructed through the clustered substation area model;

[0052] The real-time line loss rate of each substation is obtained based on state estimation, topology information and substation line loss rate.

[0053] The specific steps of the federated learning line loss calculation method for distribution networks are as follows:

[0054] Obtain data of each substation in the distribution network;

[0055] Establishing a distributed intelligent line loss calculation network based on blockchain, wherein the distributed intelligent line loss calculation network is implemented on the blockchain system and includes a station edge learning system and a local data center;

[0056] Model training of the regional edge learning system based on the federated learning algorithm;

[0057] Clustering the trained model based on the adaptive weight clustering algorithm;

[0058] The real-time line loss rate of each substation is calculated based on the clustering-based substation line loss model and state estimation;

[0059] The line loss rate of a single substation is analyzed based on the decision support system.

[0060] This paper introduces advanced technologies such as federated learning, blockchain, and an improved LSTM-GCN hybrid network to build an edge-chain-cloud collaborative computing framework, enabling efficient and accurate line loss calculation while protecting data privacy. Furthermore, the paper incorporates an adaptive weighted clustering algorithm, digital twin technology, and a multi-source data fusion mechanism, improving model training efficiency and the adaptability of line loss calculation.

[0061] The present invention has the following beneficial effects:

[0062] 1. Effectively protect data privacy: Through federated learning and blockchain technology, a secure computing model is implemented in which data does not leave the domain and models are collaboratively optimized, fundamentally solving the problem of data privacy protection.

[0063] 2. Improve computing efficiency: Adopt a distributed computing architecture to disperse computing tasks to edge nodes in each area, reducing the computing pressure at the center; introduce gradient compression and adaptive weight clustering algorithms to reduce communication overhead and improve training efficiency.

[0064] 3. Improved computational accuracy: The improved LSTM-GCN hybrid network can simultaneously process time series data and topological relationships, improving the model's expressiveness; digital twin technology effectively solves the problem of missing data and improves computational integrity.

[0065] 4. Enhanced system stability: Multi-layer architecture design and fault-tolerant mechanism enable the system to cope with abnormal situations such as data loss and network fluctuations, and maintain stable operation.

[0066] 5. Support anomaly detection: The power theft detection mechanism based on multi-source data fusion improves the accuracy of identifying power theft behavior and reduces line loss rate.

[0067] 6. Provide decision support: Based on line loss feature analysis and clustering, provide targeted line loss assessment decision recommendations for each substation to support refined power grid management. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is the overall architecture diagram of the federated learning line loss calculation system for distribution network of the present invention;

[0069] Figure 2 1 is a hierarchical structure diagram of the federated learning line loss calculation system for distribution network of the present invention;

[0070] Figure 3 is a flow chart of the method for calculating line loss through federated learning according to the present invention;

[0071] Figure 4 is a flow chart of the adaptive weighted clustering algorithm of the present invention;

[0072] Figure 5 It is a structural diagram of the improved LSTM-GCN hybrid network of the present invention;

[0073] Figure 6 It is a flow chart of the line loss calculation process of the present invention. DETAILED DESCRIPTION

[0074] Please refer to the attached Figure 1-6 , the specific implementation of the present invention is further described in detail below with reference to the accompanying drawings.

[0075] like Figure 1 As shown, the federated learning line loss calculation system for distribution network provided by the present invention includes a substation edge learning system 1, a local data center 2, a blockchain system 3 and a cloud service center 4.

[0076] The substation edge learning system 1 is deployed at each substation's distribution transformer location to obtain substation voltage, current, power, and power factor data. It uses an improved LSTM-GCN hybrid network to build a model, combined with a privacy-preserving federated learning algorithm to obtain a substation line loss model. In one embodiment of the present invention, using a certain city's distribution network as an example, the substation edge learning system 1 uses industrial-grade edge computing devices deployed in 110 typical substations. Each substation is equipped with a computing device with a 4-8 core processor and 8-16GB of memory, ensuring stable operation even during peak power consumption periods (such as the peak air conditioning load period in summer).

[0077] Local Data Center 2, deployed at the regional power company, aggregates local historical and real-time data to the local data center. The local data center aggregates the substation data to establish feature engineering for line loss and performs cluster classification. Based on state estimation and topology identification, the substation model derives the line loss rate for each substation. Preferably, Local Data Center 2 utilizes a distributed architecture. In a specific deployment at a provincial power company, this was implemented using a cluster of 10 high-performance servers, supporting daily data processing of 15 TB, capable of handling the large-scale data analysis requirements of power grid operations.

[0078] Blockchain system 3 is used to establish a consensus and mutual trust mechanism between the power grid business layer and the power grid data layer, ensuring the legitimacy and validity of data collection across the entire network, while providing functional support for consensus interaction, data exchange, state information interaction, and model interaction. During the implementation of the present invention, blockchain system 3 adopts a consortium chain structure, with each substation edge node and the local data center serving as on-chain nodes, jointly maintaining a distributed ledger. Taking the actual application of a regional power grid as an example, when the number of nodes is 21, the system adopts an improved PBFT (Practical Byzantine Fault Tolerance) consensus algorithm, with a fault tolerance rate of f = (n-1) / 3 = 6.67. This means that the system can still operate normally even if up to 6 nodes fail, ensuring the reliability of data interaction.

[0079] Cloud Service Center 4, deployed on the power company's central server, receives computational information and cluster centers from the substation edge learning system trained through federated learning. It then uses line loss rate feature engineering and cluster analysis combined with substation line loss rate data to generate line loss assessment and decision-making recommendations for each substation. The computational information from the substation edge learning system includes substation line loss model information and gradient information. In an actual deployment case at a power grid company, Cloud Service Center 4, by integrating with the existing power grid management system, enables visualization of line loss data and anomaly warnings, helping operations and maintenance personnel promptly identify and address high line loss issues.

[0080] like Figure 2 As shown, the hierarchical structure of the system of the present invention consists of a basic layer, a support layer, a business layer, and an application layer, forming a complete edge-chain-cloud collaborative computing framework.

[0081] The foundational layer, deployed in the power edge gateway, is responsible for network communication, data communication interaction, and model interaction. The support layer, the interaction layer for underlying grid equipment, is responsible for information exchange and status monitoring. The business layer, including local data centers and substation edge learning systems, is responsible for edge-side model training and state estimation interaction. The application layer, located in the cloud service center, provides decision support and a user interface.

[0082] The structure and function of the blockchain system, such as Figure 1 As shown, the blockchain system 3 includes a communication protocol sublayer 31, a system security protection sublayer 32 and an edge computing sublayer 33.

[0083] The communication protocol sublayer 31 enables information exchange and supports multiple communication protocols, including UDP, TCP, and MQTT, ensuring reliable and real-time data transmission between nodes. In actual distribution network applications, when substation equipment is connected via different communication methods (such as wired fiber, 4G / 5G wireless, and power line carrier), the communication protocol sublayer 31 can automatically adapt to different protocols and convert them into a unified format, greatly improving system compatibility.

[0084] System Security Sublayer 32 is used for security protection, providing identity authentication, access control, and intrusion detection. In a power company application, System Security Sublayer 32 uses a PKI (public key infrastructure)-based identity authentication mechanism, combined with a role-based access control strategy, to effectively prevent unauthorized access and ensure the safe operation of critical grid infrastructure.

[0085] The edge computing sublayer 33 enables computational information exchange, supporting confidential computing and secure multi-party computing, enabling blockchain nodes to collaborate while protecting data privacy. Furthermore, the edge computing sublayer 33 supports smart contract execution. In actual deployment, three types of smart contracts have been developed: model verification contracts (for verifying the validity of model updates), permission management contracts (for controlling access rights of various parties), and data exchange contracts (for regulating the data exchange process). These contracts execute automatically without human intervention, significantly improving system efficiency.

[0086] The structure and function of the area edge learning system. The area edge learning system 1 includes a communication protection module 11, a local data processing module 12, an encryption communication module 13 and a local calculation module 14.

[0087] Communication protection module 11 is used to encrypt data before transmission to the cloud service center and to maintain the decrypted data locally after decryption. In the actual application scenario of distribution network line loss calculation, communication protection module 11 uses the AES-256 encryption algorithm to encrypt sensitive data (such as customer meter readings and substation load data), and the key is exchanged using a secure key exchange protocol. For example, in one distribution network implementation, when the system detects electricity usage data in high-value commercial areas, it automatically increases the encryption strength to ensure data transmission security.

[0088] The local data processing module 12 is used to provide services for collecting and updating power user measurement data, distribution transformer topology information, and historical model parameters. In actual deployment, the local data processing module 12 supports access to multiple data sources, including smart meters, distribution transformer monitoring equipment, and line sensors. Taking a residential substation as an example, the system collects data such as three-phase voltage (sampling accuracy of 0.1V per phase voltage), three-phase current (sampling accuracy of 0.01A per phase current), active power (sampling accuracy of 1W), and power factor (sampling accuracy of 0.001). The sampling frequency is dynamically adjusted according to the application scenario, generally once per minute. When the load changes sharply (such as during peak hours in the morning and evening), it can be increased to once every 10 seconds to ensure data timeliness.

[0089] The encryption communication module 13 is used to encrypt data collected from the local data processing module and decrypt information transmitted by the cloud service center. In actual applications of the power distribution network, the encryption communication module 13 works in conjunction with the communication protection module 11 to form a complete end-to-end encrypted communication link. For example, when the system transmits sensitive data containing characteristics of user electricity usage patterns in a substation, a double encryption mechanism is used. Even if the communication link is eavesdropped, the attacker cannot obtain valid information.

[0090] The local computing module 14 is used to upload the line loss model and federated learning parameters to the edge computing node, synchronize the measurement data of user meters in the substation area, topology information, model updates, and integrate the training of local data. In actual distribution network line loss calculation applications, the local computing module 14 uses a lightweight deep learning framework optimized for the limited resources of edge devices. For example, on an edge device deployed in a certain substation, through model quantization and pruning technology, a model that originally required 4GB of memory was optimized to only 800MB of memory, significantly reducing hardware costs.

[0091] The structure and function of the local data center, the local data center 2 includes a data aggregation module 21, a feature engineering module 22, a cluster analysis module 23 and a state estimation module 24.

[0092] Data Aggregation Module 21 receives and aggregates local historical and real-time data from each substation's edge learning system. In a real-world application at a provincial power company, Data Aggregation Module 21 employed a distributed data storage architecture supporting petabyte-level data storage. When deployed in an urban distribution network encompassing 300 substations, the system processed 15TB of data daily, enabling hierarchical data management and efficient retrieval, with an average query response time of less than 200ms.

[0093] The feature engineering module 22 is used to establish a feature set for line loss calculation based on aggregated data. In the actual application of distribution network line loss calculation, the feature engineering module 22 extracts multidimensional features from the raw data, including time characteristics (such as power consumption period and seasonal factors), load characteristics (such as load curve shape and peak-to-valley difference), and grid characteristics (such as line parameters and topology). For example, for a commercial district substation, the system extracts over 120 feature dimensions, including 24-hour load curve characteristics, weekday / weekend power consumption pattern characteristics, and temperature correlation characteristics, to construct a comprehensive feature vector, providing a foundation for subsequent line loss analysis.

[0094] Cluster analysis module 23 is used to cluster and classify feature sets. In an actual deployment case of the present invention, cluster analysis module 23 uses a modified K-means algorithm to perform cluster analysis based on the characteristics of different substation types. For example, in a certain urban power grid, the system automatically divides 120 substations into five categories by analyzing their characteristics: high-density residential areas, commercial areas, industrial areas, mixed areas, and low-density residential areas. Each substation type uses a targeted line loss calculation model to improve calculation accuracy.

[0095] State estimation module 24 is used to calculate the line loss rate for each substation based on the feature set and topology identification information, combined with the substation line loss model. In actual distribution network line loss calculations, state estimation module 24 uses weighted least squares to perform state estimation and combines it with Kalman filtering technology to improve the accuracy and robustness of the estimation results. For example, during the line loss calculation process for a substation in an industrial area, the system detected large voltage fluctuations. Using Kalman filtering technology, the state parameters were dynamically adjusted, reducing the line loss calculation error from the original 5.2% to 1.8%, significantly improving the calculation accuracy.

[0096] The structure and functions of the cloud service center, the cloud service center 4 includes a model aggregation module 41, a gradient evaluation module 42, a cluster analysis module 43 and a decision support module 44.

[0097] The model aggregation module 41 is used to receive and aggregate model information and gradient information uploaded by the substation edge learning system. In the application scenario of distribution network line loss calculation, the model aggregation module 41 adopts a secure aggregation mechanism to perform aggregate calculations while protecting the privacy of individual gradients. For example, in an application case of a regional power grid, when the system receives encrypted gradient information from 87 substations, it completes the aggregate calculation using secure aggregation technology. The entire process takes only 3.5 seconds, and no single node can obtain the original gradient information of other nodes.

[0098] The gradient assessment module 42 is used to assess the quality of gradient information and perform adaptive weight assignment based on the assessment results. In practice, the gradient assessment module 42 evaluates gradient quality using multiple metrics, such as gradient amplitude, directional consistency, and stability. For example, during seasonal load fluctuations in a distribution network, the system detected degraded gradient quality in some substations. By dynamically adjusting the weight coefficient (reducing the weight from 0.3 to 0.1), the system effectively avoided the negative impact of these substations on the global model and maintained the stability of model training.

[0099] Cluster analysis module 43 performs cluster analysis based on gradient information to identify similar substation characteristics. In actual deployments of line loss calculations, cluster analysis module 43 uses an adaptive weighted clustering algorithm to classify substations based on gradient characteristics. For example, in a case study at a provincial power company, the system automatically categorized 150 substations into three types: stable, fluctuating, and abnormal. These abnormal substations were then monitored closely to promptly detect abnormal behaviors such as electricity theft.

[0100] Decision support module 44 is used to generate decision recommendations for substation line loss assessment based on cluster analysis results and line loss rate data. In the actual application of distribution network management, decision support module 44 combines line loss rate time series analysis, anomaly detection, and multi-source data analysis to generate visualization reports and optimization recommendations. For example, in an application case of a regional power grid, the system detected that the line loss rate of a substation was consistently higher than that of similar substations by more than 15%. Through analysis, it was found that this was caused by unbalanced transformer load. The system automatically proposed load adjustment recommendations. After implementation, the line loss rate of the substation was reduced by 3.2 percentage points, saving approximately 45MWh of electricity annually.

[0101] Improved LSTM-GCN hybrid network structure:

[0102] The improved LSTM-GCN hybrid network of the present invention includes an LSTM timing module 51, a GCN topology module 52, a feature fusion layer 53 and a line loss prediction layer 54.

[0103] The LSTM timing module 51 processes the temporal variation characteristics of the substation's voltage, current, power, and power factor. In practical applications for calculating line losses in distribution networks, the LSTM timing module 51 employs a multi-layer bidirectional LSTM structure to simultaneously capture both forward and reverse temporal dependencies. For example, in analyzing line losses in a residential substation, this module effectively captures load variations between daytime lows and evening peaks, as well as differences in electricity consumption between weekends and weekdays, improving the expressiveness of temporal features.

[0104] The GCN topology module 52 processes the topological features of the connectivity between distribution network devices. In actual distribution network applications, the GCN topology module 52 uses a multi-layer graph convolutional architecture to map distribution network devices into graph nodes and power connectivity into graph edges. For example, in a specific industrial park distribution network application, the system maps 57 distribution devices (including transformers, circuit breakers, load switches, etc.) into graph nodes and 74 power connection lines into graph edges. Using a two-layer graph convolutional network (64 feature channels per layer), the system extracts topological features, effectively capturing the connectivity between devices and the characteristics of power flow.

[0105] The mathematical expression of the GCN layer is as follows:

[0106] ,

[0107] in: For the The node feature matrix of the layer, the dimension is the number of nodes No. Layer feature dimension; For the The node feature matrix of the layer, the dimension is the number of nodes x Layer feature dimension; To preserve the adjacency matrix of directionality and impedance information, the inverse of impedance (i.e., admittance) can be directly used as the edge weight to represent the connection relationship and electrical characteristics between nodes; For the The learnable weight matrix of the layer has the dimension Layer feature dimension x Layer feature dimension; is the activation function. In this embodiment, the ReLU function is used, which is defined as ReLU .

[0108] The feature fusion layer 53 is used to fuse time series features with topological features to form a comprehensive representation. In the actual application scenario of distribution network line loss calculation, the feature fusion layer 53 uses attention and gated fusion mechanisms to dynamically adjust feature weights based on task importance. For example, in a line loss analysis of a mixed-use area, the system automatically increases the weight of topological features (from 0.4 to 0.7) during high-load periods, while increasing the weight of time series features during low-load stable periods. This dynamic adjustment enables the model to better adapt to different operating conditions.

[0109] The mathematical expression of feature fusion is as follows:

[0110] ,

[0111] in: is the fused feature representation, and its dimension is the number of nodes × fused feature dimension; It is the feature output by the LSTM timing module, and its dimension is the number of nodes × LSTM feature dimension; The feature output by the GCN topology module has a dimension of the number of nodes × GCN feature dimension (the GCN feature dimension is adjusted to the same as the LSTM feature dimension through the projection matrix); is the gating vector, with a dimension of the number of nodes × 1 and a value range of [0, 1], which controls the fusion ratio of the two features; Represents element-wise multiplication (Hadamard product). Gate vector The calculation formula is:

[0112] ,

[0113] in: is the weight matrix, with the dimension of (LSTM feature dimension + GCN feature dimension) × 1; is the bias vector, dimension is 1; For the sigmoid activation function, ensure that the gate value is in the range of [0,1]; Indicates connecting LSTM features and GCN features on the feature dimension, where the dimension is the number of nodes × (LSTM feature dimension + GCN feature dimension).

[0114] The line loss prediction layer 54 calculates line loss predictions based on the comprehensive representation. In distribution network applications, this layer uses a fully connected network structure to map fused features to a line loss rate prediction space. For example, in a residential substation application, the system maps fused features to line loss rate predictions using a three-layer fully connected network (with layer sizes of 256, 128, and 64, respectively). The final layer uses a sigmoid activation function to constrain the output to the range [0, 1], representing the line loss rate as a percentage.

[0115] The mathematical expression for line loss prediction is as follows:

[0116] ,

[0117] in: is the predicted value of line loss rate, with the dimension of the number of nodes × 1, which represents the predicted value of line loss rate of each node (corresponding to each substation in the distribution network); is the output layer weight matrix, the dimension is the fusion feature dimension × 1; is the output layer bias vector, dimension is 1; In practical applications, we can add an activation function (such as Sigmoid) to limit the output range to a reasonable range.

[0118] Privacy-preserving federated learning algorithm:

[0119] The privacy-preserving federated learning algorithm of the present invention includes a gradient encryption module, a differential privacy module, a gradient compression module, and a security aggregation module, forming a complete privacy protection mechanism.

[0120] The gradient encryption module encrypts locally calculated model gradients. In practical applications for calculating line losses in distribution networks, the gradient encryption module employs homomorphic encryption technology, enabling direct calculations on encrypted data without the need for decryption. For example, in a regional power grid application, the system employs the Paillier homomorphic encryption algorithm with a 2048-bit key. Once the local gradient is calculated by the edge device in the substation, it is immediately encrypted, ensuring that sensitive information (such as user electricity usage characteristics) is not leaked. At the same time, the central server can still perform aggregate calculations on the encrypted gradient.

[0121] The differential privacy module is used to add calibration noise to the model gradient to achieve differential privacy protection. In actual distribution network applications, the differential privacy module implements ($\epsilon$,$\delta$)-differential privacy protection by adding Laplace or Gaussian noise. For example, in a power company's line loss calculation system, for a substation containing a large amount of sensitive data of industrial and commercial users, the system is set to =0.5 provides stronger privacy protection, while for areas that mainly contain public facilities, it is set =1.5, improving model accuracy while protecting privacy.

[0122] The mathematical expression of differential privacy protection is as follows:

[0123] ,

[0124] in: The gradient after adding noise has the same dimension as the original gradient; is the original gradient; is the noise function, generating a noise vector of the same dimension as the original gradient; Sensitivity, which indicates the maximum impact of a single sample on the gradient, is determined by gradient clipping in the distribution network line loss calculation, with a typical value of 0.1-1.0; is the privacy budget parameter, which controls the privacy protection strength. Provides stronger privacy protection but may reduce model accuracy. For the Laplace mechanism, the noise obeys Distribution, that is, the probability density function is ; For Gaussian mechanism, the noise obeys distribution, where is the upper bound of the probability of privacy leakage, which is usually set to a very small value (such as ).

[0125] Compress first and then encrypt. The gradient compression module should first compress the original gradient, and then the gradient encryption module should encrypt the compressed gradient. In the actual application scenario of distribution network line loss calculation, the gradient compression module adopts the gradient compression method based on importance. For example, in the application case of a regional distribution network, the system compresses the gradient data originally with a size of 64MB, and by setting the threshold parameter =0.01 (i.e. retaining gradient components with absolute values ​​greater than 0.01), reducing the data size to 12.8MB and achieving a compression rate of 80%, significantly reducing communication overhead and being particularly effective in rural power grid areas with limited bandwidth.

[0126] The mathematical expression of gradient compression is as follows:

[0127] ,

[0128] in: is the compressed gradient, the dimension is the same as the original gradient, but most of the elements are 0; is the original gradient; is the threshold parameter, which controls the compression rate. This will result in higher compression ratios but may result in some loss of information; is an indicator function, which is 1 when the condition is met and 0 otherwise. ; Represents element-wise multiplication (Hadamard product). This formula means retaining the gradient components whose absolute values ​​are greater than the threshold $\theta$ and setting the other components to zero to achieve gradient sparsification and compression.

[0129] The secure aggregation module is used to aggregate gradient information from multiple parties while protecting individual data privacy. In the application scenario of calculating line losses in distribution networks, the secure aggregation module uses secure multi-party computing technology. For example, in an application case at a provincial power company, when 87 substations participated in federated learning, the system used a secure aggregation protocol, which allowed the central server to only obtain the aggregated gradient sum and prevented it from obtaining the gradient information of any individual substation. Even if the central server colluded with some substations (no more than half of the total), it would not be able to decipher the gradient information of other substations, thus ensuring data privacy and security from an institutional perspective.

[0130] Adaptive weight clustering algorithm:

[0131] The adaptive weighted clustering algorithm of the present invention is implemented based on the following process:

[0132] Initialize the federated learning environment and assign weights to each client so that each substation client becomes a central client in subsequent federated learning. The central client has a higher weight than the ordinary client to deal with the problem of possible degradation of the quality of the model gradient value. In the actual application of distribution network line loss calculation, the initial weight is assigned according to the quality and representativeness of the substation data. For example, in the application case of a certain city's distribution network, the system assigns initial weights to 120 substations based on factors such as the substation data completeness rate, equipment coverage rate, and historical model performance: the 10 core substations with the highest quality receive a high weight of 0.3-0.5, the 30 substations with medium quality receive a medium weight of 0.1-0.3, and the remaining 80 substations receive a low weight of 0.01-0.1, ensuring that high-quality data has a greater influence in the early stages of training.

[0133] A central client issues a training request, and each client calculates the training information separately. In actual distribution network applications, a training request includes parameters such as the current global model parameters, training rounds, and learning rate. For example, in a line loss calculation system for a regional power grid, the central client sets a batch size of 64, a learning rate of 0.001, and 10 training rounds, and sends it to each substation client. After receiving the request, the client trains the model on the local data, calculates the gradient information, and transmits it encrypted. The entire process is completed entirely locally, and sensitive data does not leave the substation.

[0134] When a certain training time is met, the central client aggregates the training information from each client, combines it, and redistributes the weights. In practical applications of distribution network line loss calculation, the training time can be set to a fixed number of rounds or to a preset loss threshold. For example, in a power company application, the system sets the training rounds to 10 rounds or stops when the loss function improves by less than 0.1%. The central client then receives the encrypted gradients uploaded by each substation for secure aggregation and weight redistribution.

[0135] Each client updates its own model weights based on the weights assigned by the central client and issues the next training request. In actual application scenarios, clients update local model parameters according to the assigned weights based on the global update information they receive. For example, in an application with a cluster of substations in a residential area, after each substation receives the updated weights, it fuses the global model with the local model based on the weight ratio. For example, if the weight of a substation is 0.2, the fusion formula is: new model = 0.2 × global model + 0.8 × local model. Training then continues based on the updated model, forming an iterative optimization process.

[0136] During weighted clustering, the gradient values ​​are used as cluster centers, and cluster centers are clustered using an adaptive weighted clustering algorithm. In the actual application of distribution network line loss calculation, the clustering algorithm uses an adaptive threshold method based on Euclidean distance. For example, in an application case where a regional power grid contains 150 substations, the system calculates the Euclidean distance between the gradients of each substation and sets the distance threshold to 30% of the gradient mean (the actual value is 0.25). When the gradient distance between two substations is less than this threshold, they are classified into the same category. Ultimately, the 150 substations are automatically divided into five clusters, providing a basis for differentiated training strategies.

[0137] The mathematical expression of gradient quality evaluation is as follows:

[0138]

[0139] in: Gradient quality score, scalar value, usually ranging between [-1,1]; is the normalized gradient L2 norm, which maps the original L2 norm to the [0,1] interval; is the cosine similarity between the gradient and the global average gradient, and its value range is [-1,1]; is the normalized gradient variance, mapping the original variance to the [0,1] interval; is the weight coefficient, controlling the importance of the three factors, and .

[0140] The normalization function can use Min-Max normalization:

[0141] ,

[0142] Or after Z-score normalization, apply the Sigmoid function to map the value to the (0,1) interval:

[0143] ,

[0144] in is the sample mean, is the sample standard deviation.

[0145] In the application case of a regional power grid, it was found that Setting it too high will cause the large gradient area to dominate the training, and If the setting is too high, useful difference information will be ignored. Through experiments, it is determined that the above weight ratio can achieve a good balance in various distribution network scenarios.

[0146] Based on the gradient quality, the adaptive weight calculation formula is as follows:

[0147] ,

[0148] in: For Taiwan The update weight is a scalar value, and the sum of all the area weights is 1; For Taiwan The denominator is the sum of the gradient quality scores of all substations, ensuring that the weight sum is 1. In the actual distribution network line loss calculation application, if some substations have If it is a negative value, set it to a very small positive number (such as 0.01) to prevent the negative weight from causing the model to diverge.

[0149] The specific steps of the federated learning line loss calculation method for distribution network of the present invention are as follows:

[0150] Step 1: Obtain data from each substation in the distribution network. Voltage, current, power, and power factor data for each substation in the distribution network are obtained through the grid business layer and the grid data layer. In actual distribution network applications, data acquisition equipment is deployed at all levels of distribution transformers and key nodes, forming a multi-source, heterogeneous data acquisition network. For example, in a certain urban distribution network application, the system collects data through multiple devices, including smart meters (collection frequency 15 minutes), distribution transformer monitoring equipment (collection frequency 1 minute), and key line sensors (collection frequency 10 seconds), forming a complete data acquisition chain.

[0151] Step 2: Establish a blockchain-based distributed intelligent line loss calculation network. This distributed intelligent line loss calculation network is implemented on the blockchain system and includes a regional edge learning system and a local data center. In the actual application of distribution network line loss calculation, the distributed intelligent line loss calculation network adopts a consortium chain structure. For example, in a provincial power company application, the system built a consortium chain network consisting of 21 nodes (including 10 core regional power company nodes and 11 large distribution station nodes). The improved PBFT consensus mechanism ensures the credibility of data and calculations, and the system fault tolerance rate reaches 6.67%. Even if 6 nodes fail, the entire system can still operate normally.

[0152] Step 3: Model training for the substation edge learning system is performed based on a federated learning algorithm. In the actual scenario of calculating line losses in a distribution network, model training utilizes asynchronous federated learning. For example, in an application where a regional power grid includes different substation types, the system allows different substations to participate in training at different rates: commercial substations equipped with high-performance computing equipment can participate in global model updates more frequently (e.g., once an hour), while remote rural substations with limited computing resources can participate less frequently (e.g., once a day). This flexible mechanism significantly improves the system's adaptability and scalability.

[0153] Step 4: Cluster the trained model using an adaptive weighted clustering algorithm. In actual distribution network line loss calculations, clustering results are used to guide differentiated training strategies. For example, in a city grid with 120 substations, the system uses adaptive clustering to categorize the substations into five types: high-density residential areas (42), commercial areas (25), industrial areas (18), mixed areas (20), and low-density residential areas (15). For the commercial area cluster, the system increases the learning weight for weekday / weekend power consumption patterns; for the industrial area cluster, the system strengthens the ability to identify the start-up and shutdown characteristics of high-power equipment. This differentiated strategy enables the system to optimize for the characteristics of different substation types.

[0154] Step 5: Calculate the real-time line loss rate of each substation based on the clustered substation line loss model and state estimation. In the actual application of distribution network line loss calculation, the line loss calculation process includes the following sub-steps:

[0155] Step 5.1: Use virtual measurement nodes based on digital twins to supplement missing current and voltage data. Missing data is a common problem in real-world distribution network applications. For example, in a suburban distribution network application, an average of 7.5% of data points were missing due to sensor failures and communication interruptions. The system constructed a digital twin model based on historical data and physical laws, calculating the current and voltage data of the missing nodes in real time. This achieved a 97.3% compensation rate, effectively resolving the issue of incomplete data.

[0156] Step 5.2: Perform topology identification based on state estimation. In distribution network applications, topology identification is crucial for line loss calculation. For example, after a regional distribution network underwent line reconstruction, the actual topology differed from the original records. The system automatically identified the new topology by analyzing the measurement data, with an accuracy rate of 95.6%, avoiding line loss calculation errors caused by topology errors.

[0157] Step 5.3: Based on the feature engineering of current voltage, current, and active or reactive power, perform state estimation on the substation model and topology identification. State estimation is a key step in the calculation of actual distribution network line losses. For example, in an industrial substation application, the system uses weighted least squares to estimate the grid state. Considering the large load fluctuations caused by the start-up and shutdown of industrial equipment, the system assigns different weights to the measured data in different time periods: the data weight during stable operation is 1.0, while the data weight during the start-up and shutdown transition period is reduced to 0.6, significantly improving the accuracy of state estimation.

[0158] Step 5.4: Use Kalman filtering to correct the state estimation parameters. In practical applications of distribution network line loss calculation, Kalman filtering effectively handles measurement noise and system dynamics. For example, in a commercial district application, when air conditioning load fluctuated significantly within a short period of time (such as during the midday warming period), the system dynamically adjusted the state estimation parameters using an extended Kalman filter, reducing the state estimation error from 5.2% to 1.8%, improving the accuracy and stability of line loss calculations.

[0159] The mathematical expression of the extended Kalman filter is as follows:

[0160] Prediction steps:

[0161] ,

[0162] ,

[0163] Update steps:

[0164] ,

[0165] ,

[0166] ,

[0167] in: is the prior state estimate at time k. The distribution network line loss calculation includes state variables such as node voltage and line current. The dimension is the number of state variables. is the state transfer function, which describes how the state variables evolve over time and is usually based on the power flow equation in the distribution network; For control input, in the distribution network, it can be a control action such as a switch state change; is the prior estimation error covariance matrix, with the dimension of the number of state variables x the number of state variables, which represents the uncertainty of the prior estimation; is the Jacobian matrix of the state transfer matrix, that is right The partial derivative matrix of , with dimension of number of state variables x number of state variables; is the process noise covariance matrix, with the dimension of the number of state variables × the number of state variables, which represents the uncertainty of the system model. In actual distribution network applications, it is usually set to a diagonal matrix based on experience, with diagonal elements between 10 to 10 between; is the Kalman gain matrix, with the dimension of the number of state variables x the number of observation variables, which controls the influence of observation information on state update; is the Jacobian matrix of the observation matrix, that is right The partial derivative matrix of , the dimension is the number of observation variables x the number of state variables; is the observation noise covariance matrix, with dimensions of number of observed variables x number of observed variables, which represents the measurement uncertainty. In actual distribution network applications, it is usually set according to the sensor accuracy. The standard deviation of voltage measurement is usually 0.1-0.5V, and the standard deviation of current measurement is usually 0.01-0.05A; is the observation value vector at time k, which contains actual measurement values ​​such as voltage and current, and its dimension is the number of observation variables; Mapping state variables to observation space for observation functions; is the posterior state estimate at time k, which integrates the prediction and observation information, and its dimension is the number of state variables; is the posterior estimation error covariance matrix, with the dimension of the number of state variables x the number of state variables, which represents the uncertainty of the posterior estimation; is the identity matrix with dimensions of number of state variables x number of state variables. In the practical application of distribution network line loss calculation, state variables usually include the voltage amplitude and phase angle of each node, and observation variables include measured voltage, current, active power, and reactive power.

[0168] Step 5.5: Extract the topological information of the power equipment using the features collected by the power electronic equipment, and construct the substation line loss rate using the clustered substation model. In actual distribution network applications, topological information is a key input for line loss calculation. For example, in an application case following a city power grid renovation, the system automatically extracts the latest topological relationships using intelligent distribution terminals and line switch status signals. Combined with the clustered substation line loss model, it constructs a highly adaptable line loss calculation model, successfully addressing the challenge of frequent topology changes.

[0169] Step 5.6: Calculate the real-time line loss rate for each substation based on the state estimation, topology information, and substation line loss rate. In the actual application of distribution network line loss calculation, the real-time line loss rate calculation takes into account the influence of multiple factors. For example, in a mixed-area substation application, the system combines state estimation results, the latest topology information, and load characteristics to calculate the line loss rate in real time. This also takes into account the effects of ambient temperature (line loss increases by approximately 3.9% for every 10°C increase) and load imbalance, making the line loss calculation more accurate and reducing the calculation error from the original ±4.5% to ±1.2%.

[0170] The calculation formula of line loss rate is as follows:

[0171] ,

[0172] Where: LossRate is the line loss rate, the unit is percentage (%); is the input power, i.e. the total power output by the distribution transformer in the substation, in kilowatts (kW); is the output power, that is, the total power consumption at the user end, in kilowatts (kW); the difference between the two is divided by the input power, and then multiplied by 100% to get the percentage of line loss rate. In the actual application of distribution network line loss calculation, It is directly obtained through the measurement equipment on the distribution transformer side. It is calculated by summarizing the meter readings of all users. The calculation period is usually hourly or daily. The monthly line loss rate is obtained by taking the weighted average of the daily average values.

[0173] Step 6: Analyze the line loss rate of a single substation based on the decision support system. In the practical application of distribution network line loss management, the decision support system provides comprehensive line loss analysis capabilities. For example, in a regional power grid application, the system compared the line loss rates of similar substations (for example, comparing a commercial district substation with 25 other commercial district substations) and found that the line loss rate of this substation was consistently 3.7 percentage points higher than the average. Further analysis revealed that this was caused by a severe imbalance in the three-phase load (an imbalance of 17.8%). The system automatically generated three-phase load adjustment recommendations. After implementation, the line loss rate of this substation was reduced by 2.4 percentage points, saving approximately 36MWh of electricity annually.

[0174] In practical applications of this invention, the decision support system also integrates a power theft detection mechanism based on multi-source data fusion. For example, in a rural power grid application, the system identified three abnormal line loss points by integrating power consumption data, equipment operating status, historical power usage patterns, and geographic information. On-site inspections confirmed that power theft occurred in two of these locations. The detection accuracy reached 93.5%, significantly higher than the 72% accuracy of traditional methods, providing strong support for power line loss management.

[0175] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A federated learning line loss calculation system for distribution networks, characterized by: The system includes: The substation edge learning system is used to obtain substation voltage, current, power, and power factor data. It uses an improved LSTM-GCN hybrid network to build a model and combines it with a privacy-preserving federated learning algorithm to obtain a substation line loss model. A local data center is connected to the substation edge learning system for aggregating local historical data and real-time data to the local data center, aggregating the substation data to establish feature engineering of line loss and perform cluster classification, and obtaining the line loss rate of each substation based on the substation line loss model according to state estimation and topology identification; The blockchain system is connected to the substation edge learning system and the local data center to establish a consensus and mutual trust mechanism between the power grid business layer and the power grid data layer, ensuring the legitimacy and validity of data collection across the entire network, while providing functional support for consensus interaction, data exchange, status information interaction, and model interaction; A cloud service center is in communication with the local data center and the blockchain system, and is configured to receive the calculation information and cluster center of the substation edge learning system, and then use line loss rate feature engineering and cluster analysis to conduct a comprehensive analysis with the substation line loss rate data to obtain line loss assessment decision recommendations for each substation; The calculation information of the substation edge learning system includes substation line loss model information and gradient information; The local data center includes: Data aggregation module, used to receive and aggregate local historical data and real-time data from edge learning systems in each area; Feature engineering module, used to build feature sets for line loss calculation based on aggregated data; A cluster analysis module, used for clustering and classifying the feature set; A state estimation module, configured to calculate the line loss rate of each substation area based on the feature set and topology identification information in combination with the substation area line loss model; The line loss calculation process of the state estimation module includes: Use virtual measurement nodes based on digital twins to supplement missing current and voltage data; Topology identification based on state estimation; Based on the feature engineering of current voltage, current, active power or reactive power, the state estimation of the substation model and topology identification is performed; Use Kalman filtering technology to correct the parameters of state estimation; The topological information of the power equipment is obtained by extracting the characteristic quantities collected by the power electronic equipment, and the line loss rate of the substation area is constructed through the clustered substation area model; The real-time line loss rate of each substation is obtained based on state estimation, topology information and substation line loss rate.

2. The federated learning line loss calculation system for distribution network according to claim 1, characterized in that: The blockchain system includes: Communication protocol sublayer, used to realize information interaction; System security protection sublayer, used for security protection; The edge computing sublayer is used to realize computing information interaction.

3. The federated learning line loss calculation system for distribution network according to claim 1, characterized in that: The edge learning system includes: A communication protection module, used to encrypt the transmission data before transmission and transmit it to the cloud service center, and to locally maintain the decrypted data after decryption; Local data processing module, used to provide services for collecting and updating power user measurement data, distribution and transformation topology information, and historical model parameters; An encryption communication module, used to encrypt data collected from the local data processing module and decrypt information transmitted by the cloud service center; The local computing module is used to complete the upload of edge computing node line loss models and federated learning parameters, the synchronous transmission of user meter measurement data, topology information, model updates, and the integrated training of local data.

4. The federated learning line loss calculation system for distribution network according to claim 1, characterized in that: The cloud service center includes: A model aggregation module, configured to receive and aggregate the model information and gradient information uploaded by the edge learning system in the substation area; a gradient evaluation module, configured to perform quality evaluation on the gradient information and perform adaptive weight allocation based on the evaluation result; A cluster analysis module, configured to perform cluster analysis based on the gradient information to identify similar station area characteristics; The decision support module is used to generate decision recommendations for line loss assessment in substations based on cluster analysis results and line loss rate data.

5. The federated learning line loss calculation system for distribution network according to claim 1, characterized in that: The improved LSTM-GCN hybrid network includes: LSTM time series module, used to process the time series variation characteristics of the substation voltage, current, power, and power factor; GCN topology module, used to process the topological characteristics of the connection relationship between distribution network equipment; A feature fusion layer, used to fuse the temporal change feature with the topological feature to form a comprehensive representation; The line loss prediction layer is used to calculate the line loss prediction value based on the comprehensive representation.

6. The federated learning line loss calculation system for distribution network according to claim 1, characterized in that: The privacy-preserving federated learning algorithm includes: Gradient encryption module, used to encrypt locally calculated model gradients; A differential privacy module, used to add calibration noise to the model gradient to achieve differential privacy protection; Gradient compression module, used to compress encrypted gradients to reduce the amount of transmitted data; The secure aggregation module is used to aggregate multi-party gradient information while protecting individual data privacy.

7. The federated learning line loss calculation system for distribution network according to claim 4, characterized in that: The adaptive weight allocation is achieved based on the following process: Initialize the federated learning environment and assign weights to each client. This allows each client in each region to become a central client in subsequent federated learning. Central clients have higher weights than ordinary clients to address the issue of degradation in the quality of model gradient values. The central client sends a training request, and each client calculates the training information separately; When the training time is met, the central client collects the training information of each client, combines the training information in the central client, and redistributes the weights; Each client updates its own model weights according to the weights assigned by the central client and issues the next training request.

8. A method for calculating line loss through federated learning for distribution networks, applied to a system for calculating line loss through federated learning for distribution networks as claimed in any one of claims 1 to 7, characterized in that: The specific steps of this method are as follows: Obtain data of each substation in the distribution network; Establishing a distributed intelligent line loss calculation network based on blockchain, wherein the distributed intelligent line loss calculation network is implemented on the blockchain system and includes a station edge learning system and a local data center; Model training of the regional edge learning system based on the federated learning algorithm; Clustering the trained model based on the adaptive weight clustering algorithm; The real-time line loss rate of each substation is calculated based on the clustering-based substation line loss model and state estimation; The line loss rate of a single substation is analyzed based on the decision support system.

Images