Power grid transient stability evaluation method and system based on deep federated transfer learning

Abstract

The invention provides a power grid transient stability evaluation method and system based on deep federated transfer learning, and the method comprises the steps: collecting the historical data of a power grid, including PMU measurement sequences in a symmetric fault scene and an asymmetric fault scene, and carrying out the data preprocessing of the historical data of the power grid; pre-training the FLACON model in the source domain power grid by using the historical data of the power grid to obtain pre-training model parameters; the pre-training model parameters serve as global initial parameters and are distributed to clients of different regional power grid nodes; the multiple clients update the local model by using the local data to obtain local model parameters; the client uploads the local model parameters to the central server, and the central server weights and aggregates the global parameters according to the data feature matching degree of the local model parameters; and collecting real-time measurement data, reasoning the real-time measurement data by using the FLACON model, and outputting a stability prediction result. The technical problems of security risk of key data and easy delay of data transmission in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid stability assessment, and in particular to a method and system for power grid transient stability assessment based on deep federated transfer learning. Background Art

[0002] Grid stability assessment is a core component of power system operation and planning. Its purpose is to analyze the system's ability to maintain stable voltage, frequency, and power angle under normal or fault conditions, ensuring power supply continuity and security. With the increasing penetration of renewable energy, diversified load demands, and frequent extreme weather events, the dynamic characteristics of the power grid are becoming increasingly complex, and traditional assessment methods based on static models are no longer sufficient. Modern assessments, incorporating advanced technologies such as time-domain simulation and artificial intelligence, quantify the system's ability to resist disturbances, identify weak links, and provide a basis for scheduling decisions and equipment upgrades. These assessments are the key technical foundation supporting the low-carbon, flexible, and reliable operation of new power systems.

[0003] Traditional deep learning relies on centralized data storage and training, requiring distributed PMU (Phasor Measurement Unit) data (such as node voltage, power, frequency and other sensitive information) to be transmitted to a central control center. This poses security risks such as eavesdropping, tampering or abuse, violating power grid data security regulations. The real-time operating data of regional power grids contains sensitive information such as user load patterns and key node status. Centralized transmission may allow attackers to locate weak links. Large-scale power grids (such as transmission and distribution grids with thousands of nodes) have huge amounts of PMU data (hundreds of MB per second). Centralized transmission requires a high-bandwidth communication network, which is prone to delays (for example, millisecond-level delays may miss the optimal control window after a fault). The central server needs to handle the storage, cleaning and feature extraction of massive amounts of data, which has a high computational load and is difficult to meet the second-level real-time requirements of transient stability assessment (the assessment must be completed within 1-2 seconds after a fault). Summary of the Invention

[0004] The purpose of the present invention is to overcome the above-mentioned technical deficiencies and provide a method and system for evaluating the transient stability of power grids based on deep federated transfer learning, so as to solve the technical problems in the prior art of security risks of key data and easy delays in data transmission.

[0005] To achieve the above technical objectives, in a first aspect, the technical solution of the present invention provides a method for evaluating power grid transient stability based on deep federated transfer learning, comprising:

[0006] Collecting historical power grid data, including PMU measurement sequences under symmetrical fault and asymmetrical fault scenarios, and performing data preprocessing on the historical power grid data;

[0007] Pre-training a FLACON model in a source domain power grid using the power grid historical data to obtain pre-trained model parameters;

[0008] Distribute the pre-trained model parameters as global initial parameters to clients of power grid nodes in different regions;

[0009] Multiple clients use local data to update local models and obtain local model parameters;

[0010] The client uploads the local model parameters to the central server, and the central server weights and aggregates the global parameters according to the data feature matching degree of the local model parameters;

[0011] Real-time measurement data is collected, the FLACON model is used to infer the real-time measurement data, and a stability prediction result is output.

[0012] Compared with the prior art, the beneficial effects of the present invention include:

[0013] The regional power grids (clients) in this application only need to share local model parameters instead of original data to avoid the leakage of sensitive power data and meet the strict data security requirements of the power industry. Reduce the risk of centralized storage: Traditional centralized training requires all data to be aggregated to a central server, which is vulnerable to network attacks or data leaks, while the federated learning model disperses the risk. By pre-training the FLACON model on the source domain power grid, the model can learn general stability characteristics (such as the dynamic response law of faults), and then migrate to different regional power grids (target domains) to reduce the impact of insufficient data in the target domain. Different regional power grids may have different data distributions due to differences in topology, load characteristics, etc. Transfer learning can alleviate the "data island" problem and enhance model adaptability. Each client only needs local data to participate in training, and there is no need for large-scale centralized labeling of data, which is especially suitable for remote areas or newly built power grid nodes. The client can update the local model in real time (such as new fault scenario data), and continuously optimize the global model through the federal aggregation mechanism to adapt to changes in power grid operation.

[0014] According to some embodiments of the present invention, the FLACON model includes: a convolutional layer, a Transformer block and a fully connected layer.

[0015] According to some embodiments of the present invention, the pre-trained model parameters include: convolutional neural network (CNN) parameters, multi-head attention (MHA) mechanism parameters and fully connected layer parameters.

[0016] According to some embodiments of the present invention, the global parameter aggregation formula of the central server is:

[0017]

[0018] Among them, Θ new is the global model parameter after aggregation, W k is the data feature matching weight of the k-th client, Wj Aggregation weight of the jth client: based on DTW distance or data volume ratio, Θ k are the local model parameters of the k-th client.

[0019] According to some embodiments of the present invention, collecting real-time measurement data and using the FLACON model to reason about the real-time measurement data include the following steps:

[0020] The FLACON model generates stability probabilities through the following process

[0021]

[0022] Input: Normalized PMU measurement sequence

[0023] Output: The Sigmoid activation function of the last fully connected layer is compressed to the range [0,1]. The formula is:

[0024]

[0025] Where z is the linear output (not activated value) of the fully connected layer.

[0026] According to some embodiments of the present invention, the steps further include:

[0027] The stability state is divided by the segmented threshold method based on the stability probability, and the formula is:

[0028]

[0029] In a second aspect, the technical solution of the present invention provides a power grid transient stability assessment system based on deep federated transfer learning, comprising:

[0030] Client module: Used to collect real-time measurement data from grid nodes, including voltage amplitude (V), voltage phase angle (θ), frequency (F), active power (P), and reactive power (Q), and generate historical data sets containing symmetrical and asymmetrical fault scenarios; preprocess historical data, including normalization, sliding window segmentation, and time series generation; receive global model parameters issued by the central server, update the local model, and perform local training;

[0031] Central server module: Initializes global model parameters. The global model is the FLACON model that integrates 1D convolutional neural network (CNN) and multi-head attention mechanism (MHA), including convolutional layers, Transformer blocks, and fully connected layers. It receives local model parameters or gradients uploaded by each client, performs weighted aggregation based on the dynamic matching features of each client, and generates updated global model parameters. It distributes global model parameters to clients and coordinates the federated learning process.

[0032] Transfer learning module: Pre-train the FLACON model in the source domain power grid to extract cross-domain common features; adaptively fine-tune the model parameters of the target domain power grid through inductive transfer learning;

[0033] Online evaluation module: Utilizes real-time measurement data sequences, performs forward propagation through the FLACON model, and outputs the transient stability probability of the power grid.

[0034] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, wherein the abstract drawing is identical to one of the drawings in the specification:

[0036] Figure 1 A flowchart of a power grid transient stability assessment method based on deep federated transfer learning provided by one embodiment of the present invention;

[0037] Figure 2 A structural diagram of a power grid transient stability system based on deep federated transfer learning provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0039] It should be noted that although the system diagrams illustrate functional module divisions and the flowcharts illustrate logical sequences, in certain circumstances, the steps shown or described may be performed in a sequence that differs from the module divisions in the system or the sequence in the flowcharts. The terms "first," "second," and so on, in the specification, claims, and drawings, are used to distinguish similar items and are not necessarily used to describe a specific sequence or precedence.

[0040] Reference Figure 1 , Figure 1 A flowchart of a method for evaluating power grid transient stability based on deep federated transfer learning is provided in accordance with one embodiment of the present invention. The method for evaluating power grid transient stability based on deep federated transfer learning includes but is not limited to the following steps:

[0041] Step S110 , collecting historical power grid data, including PMU measurement sequences under symmetrical fault and asymmetrical fault scenarios, and performing data preprocessing on the historical power grid data;

[0042] Step S120, using historical power grid data to pre-train the FLACON model in the source domain power grid to obtain pre-trained model parameters;

[0043] Step S130: Distribute the pre-trained model parameters as global initial parameters to clients at power grid nodes in different regions. Multiple clients use local data to update local models to obtain local model parameters.

[0044] Step S140: The client uploads the local model parameters to the central server, and the central server weights and aggregates the global parameters according to the data feature matching degree of the local model parameters;

[0045] Step S150 : collecting real-time measurement data, using the FLACON model to infer the real-time measurement data, and outputting a stability prediction result.

[0046] In one embodiment, a method for evaluating transient stability of a power grid based on deep federated transfer learning includes: collecting historical power grid data, including PMU measurement sequences under symmetrical fault and asymmetrical fault scenarios, and performing data preprocessing on the historical power grid data; using the historical power grid data to pre-train a FLACON model in a source domain power grid to obtain pre-trained model parameters; using the pre-trained model parameters as global initial parameters and distributing them to clients at power grid nodes in different regions; multiple clients use local data to update local models to obtain local model parameters; the clients upload the local model parameters to a central server, and the central server weights and aggregates the global parameters based on the data feature matching degree of the local model parameters; collecting real-time measurement data, using the FLACON model to infer the real-time measurement data, and outputting stability prediction results.

[0047] Among them, the local model training and update unit

[0048] Function: Update local models based on the federated learning framework and perform training or inference tasks;

[0049] Technical features:

[0050] Model structure: Adopts FLACON architecture, including:

[0051] 1D convolutional layer: 5 layers, 64 filters per layer (kernel size 1×4), activation function ReLU, output local feature map;

[0052] Multi-head attention layer (MHA): 8 heads, calculate query matrix Qu, key matrix K, value matrix V a The attention weight is:

[0053]

[0054] Fully connected layer: 2 layers (128-dimensional and 1-dimensional), Sigmoid activation output stability probability (0-1);

[0055] Loss function: Focal Loss, the formula is:

[0056] L focal =-α t (1-p t ) γ log(p t )

[0057] Among them, α t is the class balance factor (unstable sample weight = 0.8), γ = 2 is the focusing parameter;

[0058] Convergence judgment: If the local training meets one of the following conditions, it converges:

[0059] The number of training times reaches the preset value (such as 50 epochs);

[0060] Validation set accuracy ≥ 95% or loss ≤ 0.1.

[0061] PMUs (Phasor Measurement Units) are core devices used in power systems to monitor grid dynamics in real time. They achieve nanosecond-level time synchronization through the Global Positioning System (GPS) and collect synchronized time series data of the following five key electrical quantities, forming the PMU measurement sequence:

[0062] Voltage amplitude (V): The effective value of the node voltage (unit: pu or kV), reflecting the energy transmission capacity of the power grid; Voltage phase angle (θ): The phase angle of the node voltage (unit: degree), which is the core parameter for transient stability assessment (the power angle difference directly determines the system synchronization); Frequency (F): The system frequency (unit: Hz), which is 50 / 60Hz in normal operation and deviates during faults; Active power (P): The active power of the generator / load (unit: MW), reflecting the balance of energy supply and demand; Reactive power (Q): The reactive power of the generator / load (unit: Mvar), affecting voltage stability.

[0063] Sequence characteristics: High refresh rate: typically 10-100Hz (i.e. 10-100 samples per second), which can be increased to 500Hz during faults; Time synchronization: All node data is based on the GPS time scale (such as UTC time), with a timestamp accuracy of ≤1μs, ensuring the spatiotemporal consistency of cross-regional data; High dimensionality: A single node generates a 5-dimensional time series sequence, and a large-scale power grid (such as a 1000-node system) forms a 5000-dimensional high-dimensional sequence.

[0064] This application's federated learning architecture ensures that PMU measurement data remains locally within the client through local training and uploads only updated model parameters, mitigating the data leakage risks inherent in centralized architectures and complying with grid data security regulations. Cross-regional collaborative optimization supports independent training for grids in different regions (e.g., city centers and suburbs). A central server aggregates the global model, taking into account the characteristics of each region. This improves the model's adaptability to heterogeneous grid data and avoids system failures caused by single-node failures.

[0065] The 1D convolutional layer captures local temporal characteristics of electrical quantities such as voltage and frequency (e.g., the slope of a voltage sag within 0.1 seconds after a fault). The multi-head attention mechanism (MHA) models global dependencies across nodes (e.g., the dynamic relationship between generator power angle difference and active power). This achieves a comprehensive recognition accuracy of over 98% for symmetrical and asymmetrical faults (e.g., 3PB, SLG, and LL), significantly outperforming traditional LSTM / CNN models. The focus loss function specifically addresses data imbalance. By dynamically adjusting the loss weight for the minority class (unstable samples), the model's recall rate for unstable scenarios increases from 40% with traditional methods to 99%, reducing the risk of missed detections.

[0066] By pre-training common features on standard power grids (such as IEEE 39-bus), the system can rapidly adapt to new topologies (such as IEEE 68-bus) through inductive transfer learning. This method improves accuracy from 87% in the source domain to 99% with only 10% of the target domain data, significantly reducing retraining costs. The central server weights aggregation parameters based on the match between client data and the global model (such as the sum of DTW distances). Clients with high data feature similarity contribute more, improving cross-domain migration efficiency and avoiding model bias caused by "one-size-fits-all" aggregation. The model can be locally deployed on the substation edge server, reducing cloud communication latency and achieving an end-to-end response time of less than 100ms. It supports real-time linkage with stabilization and control devices (such as PMU-BAS) to trigger emergency controls such as generator and load shedding.

[0067] Data preprocessing is performed on the historical data of the power grid, including the following steps: standardizing the historical data of the power grid, generating time series samples through a sliding window, and labeling the historical data of the power grid with the stability status determined by the transient stability index. The standardized data distribution is more concentrated, avoiding gradient explosion / vanishing. Feature fairness: prevents the model from being biased towards high-amplitude features due to dimensional differences (such as the value of P is usually much larger than the change in F). The window length of 1.28 seconds covers the key dynamic processes after the fault (such as the first swing period of the generator is about 0.5-1 second), effectively capturing transient features such as voltage sag and frequency oscillation. Data enhancement: Generate more training samples through overlapping sampling (for example, 1000 seconds of original data can generate about 6200 samples), alleviating the scarcity of power grid fault data.

[0068] The pre-trained model parameters include: convolutional neural network (CNN) parameters, multi-head attention (MHA) mechanism parameters, and fully connected layer parameters.

[0069] 1. Convolutional Neural Network (CNN) Parameters

[0070] The CNN layer is responsible for extracting local spatiotemporal features of the PMU measurement sequence (such as voltage sag slope and frequency jump pattern) and includes the following trainable parameters:

[0071] 1. Convolution kernel weight (W conv ): Dimension: The convolution kernel shape of each convolution layer is 1×4×C in ×C out

[0072] 1×4: convolution kernel size in the time dimension (processing 4 consecutive time points at a time);

[0073] C in : Number of input channels (e.g. the first layer is 5, corresponding to 5 electrical quantities: V / θ / F / P / Q);

[0074] C out : The number of output channels (e.g. 64, indicating the extraction of 64 local features);

[0075] Number: 5 convolutional layers.

[0076] 2. Convolutional layer bias (b conv ): Each convolutional layer corresponds to a bias vector with a dimension of C out , the total number of parameters is 5×64=320.

[0077] 3. Activation function parameters: The ReLU activation function has no parameters and only introduces nonlinear transformations.

[0078] 2. Multi-Head Attention (MHA) Mechanism Parameters

[0079] The MHA layer is responsible for modeling global dependencies across nodes (such as the coordinated changes in generator power angle difference and active power). The parameters are divided into the following parts:

[0080] 1. Projection Matrix (Query / Key / Value Matrix)

[0081] Parameter definition:

[0082] Query matrix W q , bond matrix W k , value matrix W v , the shape is D model ×D head ×H

[0083] D model : Input feature dimension (such as 64 dimensions output by CNN);

[0084] D head : The dimension of a single attention head (e.g. 8 dimensions);

[0085] H: number of attention heads (e.g. 8);

[0086] Quantity: Each Transformer block contains 3 sets of projection matrices, with a total parameter of: 3×(D model ×D head ×H) = 3 × (64 × 8 × 8) = 12,288 (Note: Assuming each Transformer block has 8 heads, D head =8).

[0087] 2. Multi-head splicing matrix (W o )

[0088] The 8-head attention output is concatenated and projected back to the model dimension, with a shape of H×D head ×D model , the number of parameters is:

[0089] 8×8×64=4,096

[0090] 3. Feedforward Neural Network (FFN) Parameters

[0091] Each Transformer block contains two fully connected layers:

[0092] First floor: D model ×D ffn (e.g. 64×256), the number of parameters is 64×256=16,384;

[0093] Second layer: D ffn ×D model (e.g. 256×64), the number of parameters is 256×64=16,384;

[0094] Total FFN parameters: 16,384 + 16,384 = 32,768 per Transformer block, or 32,768 × 4 = 131,072 for 4 blocks.

[0095] 4. Layer Normalization (LayerNorm) Parameters

[0096] Each Transformer block contains 2 layers of normalization, with parameters being the mean and variance vectors, and dimension D model , the total number of parameters is 2×64×4=512 (4 Transformer blocks).

[0097] 3. Fully connected layer parameters

[0098] The fully connected layer is responsible for mapping global features into stability probabilities and contains the following parameters:

[0099] 1. Feature Compression Layer

[0100] The global maximum pooling is followed by a fully connected layer to compress the 64-dimensional features to 128 dimensions with a parameter of 64×128=8,192.

[0101] 2. Classification layer

[0102] The output layer is a fully connected layer that maps 128-dimensional features to 1-dimensional probability values ​​with a parameter of 128×1=128.

[0103] 3. Activation function parameters

[0104] The Sigmoid activation function has no parameters and is only used to compress the output to the [0,1] range.

[0105] 4. Shared Parameters in Federated Transfer Learning

[0106] In the federated learning framework, pre-trained model parameters are distributed to each client through a central server, including:

[0107] 1. Global shared parameters

[0108] CNN layer weight W conv , bias b conv ; MHA layer projection matrix W q ,W k ,W v , splicing matrix W o ; FFN layer weights and bias; fully connected layer weights and bias.

[0109] 2. Client private parameters

[0110] None, all model parameters are shared globally, and the client only stores locally optimized gradients or updated copies.

[0111] 3. Transfer learning fine-tuning parameters

[0112] When fine-tuning in the target domain, the CNN and MHA layer parameters are usually frozen, and only the fully connected layer parameters (such as the classification layer weights) are adjusted to adapt to the feature distribution differences of the new power grid topology.

[0113] Furthermore, the global parameter aggregation formula of the central server is:

[0114]

[0115] Among them, Θ new is the global model parameter after aggregation, W k is the data feature matching weight of the k-th client, W j Aggregation weight of the jth client: based on DTW distance or data volume ratio, Θ is the local model parameter of the kth client.

[0116] The FLACON model is used to reason about real-time measurement data, including the following steps:

[0117] The FLACON model generates stability probabilities through the following process

[0118]

[0119] Input: Normalized PMU measurement sequence

[0120] Output: The Sigmoid activation function of the last fully connected layer is compressed to the range [0,1]. The formula is:

[0121]

[0122] Where z is the linear output (not activated value) of the fully connected layer.

[0123] The stability state is divided by the segmented threshold method based on the stability probability, and the formula is:

[0124]

[0125] Reference Figure 2 , Figure 2 A structural diagram of a power grid transient stability system based on deep federated transfer learning provided by one embodiment of the present invention.

[0126] In one embodiment, a power grid transient stability assessment system based on deep federated transfer learning applies the power grid transient stability assessment method based on deep federated transfer learning as described above, including: a client module: used to collect real-time measurement data of power grid nodes, including voltage amplitude (V), voltage phase angle (θ), frequency (F), active power (P), reactive power (Q), and generate a historical data set containing symmetric and asymmetric fault scenarios; preprocess the historical data, including standardization, sliding window segmentation and time series generation; receive global model parameters sent by the central server, update the local model and perform local training; central server module: initialize the global model parameters, the global model is a fusion 1D volume The FLACON model, which combines a convolutional neural network (CNN) with a multi-head attention mechanism (MHA), includes convolutional layers, Transformer blocks, and fully connected layers. It receives local model parameters or gradients uploaded by each client, performs weighted aggregation based on the dynamic matching features of each client, and generates updated global model parameters. It distributes global model parameters to clients and coordinates the federated learning process. The transfer learning module pre-trains the FLACON model in the source domain power grid to extract cross-domain common features. It adaptively fine-tunes the model parameters in the target domain power grid through inductive transfer learning. The online evaluation module uses real-time measurement data sequences to perform forward propagation through the FLACON model and output the transient stability probability of the power grid.

[0127] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely located relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0128] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate and may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0129] In addition, an embodiment of the present invention also provides a computer-readable storage medium, which stores computer-executable instructions, and the computer-executable instructions are executed by a processor or controller, for example, by a processor in the above-mentioned terminal embodiment, so that the above-mentioned processor can execute the power grid transient stability assessment method based on deep federated transfer learning in the above-mentioned embodiment.

[0130] Those skilled in the art will appreciate that all or some of the steps and systems in the method disclosed above can be implemented as software, firmware, hardware, and appropriate combinations thereof. Some physical components or all physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, and the computer-readable medium can include computer storage media (or non-transitory media) and communication media (or temporary media). As known to those skilled in the art, the term computer storage media is included in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data) and is volatile and non-volatile, removable, and non-removable. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory, or other memory technology, CD-ROM, digital versatile disks (DVD), or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage, or other magnetic storage devices, or can be used to store desired information and any other medium that can be accessed by a computer. Furthermore, as is well known to those skilled in the art, communication media typically embodies computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0131] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the above implementation. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of the present invention.

[0132] The specific embodiments of the present invention described above do not limit the scope of protection of the present invention. Any other corresponding changes and modifications made based on the technical concept of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A method for evaluating power grid transient stability based on deep federated transfer learning, characterized in that: Including steps: Collecting historical power grid data, including PMU measurement sequences under symmetrical fault and asymmetrical fault scenarios, and performing data preprocessing on the historical power grid data; Pre-training a FLACON model in a source domain power grid using the power grid historical data to obtain pre-trained model parameters; Distribute the pre-trained model parameters as global initial parameters to clients of power grid nodes in different regions; Multiple clients use local data to update local models and obtain local model parameters; The client uploads the local model parameters to the central server, and the central server weights and aggregates the global parameters according to the data feature matching degree of the local model parameters; Real-time measurement data is collected, the FLACON model is used to infer the real-time measurement data, and a stability prediction result is output.

2. The method for evaluating power grid transient stability based on deep federated transfer learning according to claim 1, characterized in that: The FLACON model includes: convolutional layers, Transformer blocks and fully connected layers.

3. The method for evaluating power grid transient stability based on deep federated transfer learning according to claim 1, wherein: The historical data of the power grid is preprocessed, comprising the steps of: The historical data of the power grid are standardized, and time series samples are generated through a sliding window. The historical data of the power grid are labeled, and the label is the stability state determined by the transient stability index.

4. The method for evaluating power grid transient stability based on deep federated transfer learning according to claim 2, wherein: The pre-training model parameters include: convolutional neural network (CNN) parameters, multi-head attention (MHA) mechanism parameters and fully connected layer parameters.

5. The method for evaluating power grid transient stability based on deep federated transfer learning according to claim 4 is characterized in that: The global parameter aggregation formula of the central server is: Among them, Θ new is the global model parameter after aggregation, W k is the data feature matching weight of the k-th client, W j Aggregation weight of the jth client: based on DTW distance or data volume ratio, Θ k are the local model parameters of the k-th client.

6. The method for evaluating power grid transient stability based on deep federated transfer learning according to claim 5, characterized in that: Collecting real-time measurement data and using the FLACON model to infer the real-time measurement data includes the following steps: The FLACON model generates stability probabilities through the following process Input: Normalized PMU measurement sequence That is, real-time measurement data; Output: The Sigmoid activation function of the last fully connected layer is compressed to the range [0,1]. The formula is: Where z is the linear output (not activated value) of the fully connected layer.

7. The method for evaluating power grid transient stability based on deep federated transfer learning according to claim 6, characterized in that: Also includes the steps: The stability state is divided by the segmented threshold method based on the stability probability, and the formula is:

8. A power grid transient stability assessment system based on deep federated transfer learning, applying the power grid transient stability assessment method based on deep federated transfer learning according to any one of claims 1 to 7, characterized in that: include: Client module: Used to collect real-time measurement data from grid nodes, including voltage amplitude (V), voltage phase angle (θ), frequency (F), active power (P), and reactive power (Q), and generate historical data sets containing symmetrical and asymmetrical fault scenarios; preprocess historical data, including normalization, sliding window segmentation, and time series generation; receive global model parameters issued by the central server, update the local model, and perform local training; Central server module: Initializes global model parameters. The global model is the FLACON model that integrates 1D convolutional neural network (CNN) and multi-head attention mechanism (MHA), including convolutional layers, Transformer blocks, and fully connected layers. It receives local model parameters or gradients uploaded by each client, performs weighted aggregation based on the dynamic matching features of each client, and generates updated global model parameters. It distributes global model parameters to clients and coordinates the federated learning process. Transfer learning module: pre-trains the FLACON model in the source domain power grid to extract cross-domain common features; Adaptively fine-tune the model parameters of the target domain power grid through inductive transfer learning; Online evaluation module: Utilizes real-time measurement data sequences, performs forward propagation through the FLACON model, and outputs the transient stability probability of the power grid.

Images