A method and system for voltage tracing in distribution networks based on adaptive feature extraction and MAML

Abstract

This invention relates to the field of novel power system technology, and discloses a distribution network voltage tracing method and system based on adaptive feature extraction and MAML. First, key operational data such as voltage amplitude, waveform, active and reactive power, load, and distributed generation output of each node in the distribution network are collected. Then, the voltage signal is deconstructed in the time-frequency domain using adaptive feature mode decomposition to extract key disturbance features in voltage exceedance events. Next, a deep feature fusion network and MAML meta-learning algorithm are employed to address the small sample size problem in voltage exceedance scenarios, quickly adapting to new tasks and improving the accuracy of voltage tracing. Finally, the trained model is deployed to the actual monitoring and analysis platform of the distribution network for real-time voltage tracing and monitoring analysis. This invention can effectively improve the accuracy and real-time performance of distribution network voltage management, ensure the stable operation of the power grid, and provide intelligent control support.

Description

Technical Field

[0001] This invention relates to the field of novel power system technology, specifically to a distribution network voltage tracing method and system based on adaptive feature extraction and MAML (Model-Agnostic Meta-Learning). Background Technology

[0002] With the widespread integration of distributed energy resources into modern power distribution networks, especially the high penetration of renewable energy sources such as photovoltaics and wind power, the voltage stability issues faced by distribution networks are becoming increasingly complex. Voltage fluctuations, over-limit events, and reverse power flow phenomena have become pressing technical challenges that need to be addressed in the operation of distribution networks. These over-limit events not only affect the stability and security of the power system, but may also lead to equipment damage, energy waste, and even trigger wider power outages.

[0003] Traditional voltage control methods primarily rely on static control strategies and manual dispatching, and are typically unable to respond in real time to voltage exceedance issues, especially in multi-source coordinated dispatching and complex load environments. These traditional methods struggle to adapt to the nonlinear and time-varying characteristics of voltage changes and lack fine-grained control over complex power grid environments. Therefore, there is an urgent need for a voltage source tracing and optimization method based on modern intelligent algorithms, capable of more accurately identifying voltage anomalies and achieving adaptive regulation to ensure the efficient operation and safe stability of the distribution network.

[0004] Against this backdrop, smart grid technologies, especially intelligent control technologies based on deep learning and reinforcement learning, have become important means to improve the efficiency of distribution network voltage management. Advanced methods such as adaptive feature extraction, time-frequency analysis, and reinforcement learning can significantly improve the accuracy and response speed of voltage tracing and control strategies. However, despite the progress made in voltage control, many challenges remain when facing complex operating conditions and small sample data in distribution networks, particularly in the application of rapid tracing and dynamic adjustment of voltage exceedance events, which is still immature. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a voltage tracing method for distribution networks based on adaptive feature extraction and meta-learning. Through a deep feature fusion network and meta-learning mechanism, key features are accurately extracted from voltage signals, and the method quickly adapts to new voltage limit exceedance events, thereby achieving efficient voltage tracing and control. This not only effectively overcomes the limitations of traditional voltage control methods but also improves the stability and adaptability of distribution networks under complex operating conditions. The technical solution is as follows:

[0006] The voltage tracing method for distribution networks based on adaptive feature extraction and MAML includes the following steps:

[0007] Step 1: Collect key operational data including voltage amplitude, waveform, active and reactive power, load and distributed power output of each node in the distribution network, or use simulation methods to generate multi-scenario voltage samples as auxiliary data for subsequent neural network model training.

[0008] Step 2: The voltage signal is deconstructed in the time and frequency domain using adaptive feature mode decomposition, and the voltage signal features are extracted as key disturbance features hidden in the voltage over-limit event, so as to achieve high-precision over-limit feature modeling.

[0009] Step 3: Establish a deep feature fusion network and a training mechanism driven by meta-learning to solve the small sample problem in voltage limit scenarios; specifically, use GhostNet network for deep feature extraction and combine it with MAML meta-learning algorithm for training, so that the network can quickly adapt to new tasks with small sample data.

[0010] Step 4: Combine the auxiliary data and voltage signal features to complete the data preprocessing and splicing, and construct the training dataset for the neural network; then use the deep feature fusion network and meta-learning driven training mechanism from Step 3 to train the model and optimize the voltage over-limit tracing model; deploy the trained model to the actual distribution network monitoring and analysis operation platform to perform voltage tracing and monitoring analysis in real time.

[0011] The distribution network voltage tracing system based on adaptive feature extraction and MAML includes a data acquisition module, a feature extraction and time-frequency analysis module, a deep feature fusion network module, a meta-learning training module, and a model deployment module.

[0012] The data acquisition module is used to collect data on the operating status of the distribution network. The types of data collected include voltage amplitude, waveform, active and reactive power, load data, and distributed power output data of each node in the distribution network, which serve as auxiliary data for training the neural network model.

[0013] The feature extraction and time-frequency analysis module is responsible for processing and extracting features from the acquired voltage signal. By using the adaptive eigenmode decomposition method, the original voltage signal is deconstructed in the time-frequency domain to extract the intrinsic characteristics of voltage fluctuations and obtain voltage time-frequency feature data.

[0014] The deep feature fusion network module: based on the spliced ​​voltage time-frequency feature data and auxiliary data, it uses the deep neural network GhostNet to perform feature fusion and voltage source tracing; the neural network extracts the most discriminative features from the voltage signal through convolutional layers and multi-level feature extraction mechanisms;

[0015] The meta-learning training module enhances the adaptive capability of neural networks with small sample data through the MAML meta-learning algorithm.

[0016] The model deployment module is used to deploy the trained voltage tracing model to the actual operation platform of the distribution network for real-time prediction and monitoring.

[0017] The beneficial effects of this invention are:

[0018] This invention proposes an efficient and accurate voltage source tracing method for distribution networks by combining adaptive feature extraction and meta-learning mechanisms. This method can quickly identify the root causes of voltage exceedance events and optimize voltage regulation strategies even in complex power grid environments and with small sample data. Compared with traditional methods, this invention improves the accuracy and real-time response capability of voltage source tracing, effectively enhancing the stability and adaptive adjustment capability of distribution networks, and has broad application prospects and significant engineering value. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the distribution network voltage tracing method based on adaptive feature extraction and meta-learning of the present invention.

[0020] Figure 2 This is a schematic diagram of the distribution network voltage tracing system based on adaptive feature extraction and MAML of the present invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0022] This invention addresses the challenges faced by current power distribution networks in dealing with high penetration of distributed energy resources, voltage fluctuations, and over-limit events. Traditional voltage control methods often suffer from drawbacks such as slow response, insufficient accuracy, and poor adaptability. Therefore, this invention aims to achieve efficient voltage source tracing and control by introducing deep learning and meta-learning techniques, combined with the extraction of time-frequency features of voltage signals, thereby further improving the operational stability and security of power distribution networks.

[0023] Specific objectives include:

[0024] (1) Improve the accuracy of voltage source tracing: By using adaptive feature extraction methods (such as adaptive feature mode decomposition) to accurately deconstruct the voltage signal in the time and frequency domain, the key disturbance features in the voltage over-limit event can be identified, providing highly expressive input for voltage source tracing.

[0025] (2) Solving the problem of small sample data: By combining deep feature fusion networks with meta-learning training mechanisms (such as MAML), the neural network can quickly adapt to new tasks under small sample data, thereby improving the accuracy and efficiency of voltage tracing. Especially in real-time control, it can quickly respond to complex and changing voltage events.

[0026] (3) Realize the adaptive and real-time monitoring of the voltage traceability model: Combine time and frequency characteristics with other auxiliary data, use neural networks for training and optimization, deploy the trained model to the actual operation platform of the distribution network, realize real-time voltage traceability and monitoring, ensure the stable operation of the distribution network, and be able to adaptively adjust the voltage control strategy.

[0027] This invention provides a high-efficiency, accurate, and real-time voltage traceability and control scheme for power distribution networks, addressing the shortcomings of traditional voltage control methods in dynamic power grid environments and improving the response speed and accuracy of smart grids when facing complex problems such as voltage fluctuations and exceeding limits.

[0028] Example 1:

[0029] The first embodiment of the invention is as follows: Figure 1 As shown, this embodiment provides a distribution network voltage tracing method based on adaptive feature extraction and meta-learning, including:

[0030] Step 1: Collect key operational data such as voltage amplitude, waveform, active and reactive power, load and distributed power output of each node in the distribution network, or use simulation methods to generate voltage samples for multiple scenarios, so as to provide multi-dimensional and high-quality data support for subsequent neural network model training.

[0031] This step aims to provide high-quality data support for subsequent neural network training and voltage sourcing tasks. Specifically, it includes the following sub-steps:

[0032] Step 1.1: Determine the data type;

[0033] Based on the requirements of the distribution network voltage tracing task, the necessary data types are determined, mainly including:

[0034] Voltage amplitude and waveform data: Covers voltage change information at all key nodes, used to characterize voltage fluctuation trends;

[0035] Active and reactive power data: used to characterize the impact of source load changes on voltage and reflect the operating status of the power grid;

[0036] Distributed power generation output data: Real-time power data of distributed power sources such as photovoltaic and wind power, used to identify the source of voltage disturbances;

[0037] Load data: The time-series variation characteristics of node loads help reveal the effect of load disturbances on voltage over-limits.

[0038] Step 1.2: Design of data acquisition method;

[0039] Depending on the system application scenario, data can come from the following two sources:

[0040] Actual data acquisition: Obtain actual operating data through devices such as dispatch systems, smart meters, or PMUs (Pressure Measuring Units).

[0041] Simulation data generation: Using power system simulation platforms (such as OpenDSS, MATPOWER, DIgSILENT, etc.), a typical distribution network model containing distributed generation sources is constructed, and different load disturbance and distributed generation output scenarios are set to simulate and generate corresponding voltage response data.

[0042] Step 1.3: Data preprocessing and standardization;

[0043] The raw data obtained from collection or simulation are cleaned, denoised, missing values ​​are filled in, and time scales are aligned. Normalization or standardization methods are used to process data features to improve the numerical stability and convergence speed of neural network training.

[0044] Step 1.4: Construction of training samples;

[0045] Based on the temporal characteristics and node correlation of voltage over-limit events, a sample dataset required for neural network training is constructed. Each training sample includes multi-dimensional input features (voltage, current, power, etc.) within a time window and its corresponding target label (such as the source node number or over-limit source feature).

[0046] Through the above processing, a high-quality sample set covering various operating scenarios and disturbance modes is constructed, providing a solid foundation for subsequent voltage over-limit identification and neural network source tracing model training.

[0047] Step 2: This step aims to identify the key disturbance features hidden in voltage over-limit events. It proposes an adaptive feature mode decomposition method to deconstruct the original voltage signal in the time and frequency domain, extract its intrinsic characteristics, and achieve high-precision over-limit feature modeling, providing highly expressive feature inputs for subsequent voltage source tracing models.

[0048] To accurately identify voltage over-limit characteristics caused by factors such as distributed generation disturbances and load fluctuations in the distribution network, a feature extraction mechanism based on Adaptive Feature Mode Decomposition (AFMD) is adopted.

[0049] Suppose the single-phase voltage signal at a certain node is: ,in t Let u be the time variable, representing the node at time. t The voltage amplitude at a given time contains various components such as periodicity and non-stationary disturbances, making it difficult to extract deep features related to exceeding limits through direct analysis.

[0050] The AFMD algorithm is used to convert the original single-phase voltage signal It is decomposed into several intrinsic mode functions (IMFs) and residual signals, as shown in the following formula:

[0051] ;

[0052] in, For the first k Each modal component represents a local characteristic component in the voltage signal; K This represents the total number of modes; The residual term represents the non-periodic variation trend or long-term disturbance. This decomposition process can effectively extract the disturbance behavior of the voltage signal at different frequency ranges, improving the accuracy of subsequent feature recognition.

[0053] For each modal component Perform a Hilbert transform to obtain its instantaneous amplitude. and instantaneous frequency This reflects the local energy and frequency changes of each mode at different time scales. Further, a two-dimensional time-frequency feature matrix is ​​constructed to structurally represent the amplitude and frequency information; the voltage time-frequency feature matrix... ,as follows:

[0054] ;

[0055] In the formula, T Indicates the total number of time steps; the first k Listed as number k The amplitude changes of each mode over all time steps; each row represents the combination of instantaneous energy characteristics of different mode components at a certain moment. This feature matrix integrates the local energy information of voltage disturbances in multiple frequency modes, and can simultaneously characterize the weak disturbance characteristics and frequency domain focusing effects before and after voltage overshoot occurs.

[0056] By analyzing the voltage time-frequency characteristic matrix X Statistical analysis can identify key disturbance segments in voltage over-limit events, enabling accurate extraction of voltage over-limit features. The proposed features not only possess excellent time-frequency localization capabilities but also effectively reflect the root cause disturbances of voltage over-limit events, providing high-quality input for subsequent small-sample voltage source tracing learning.

[0057] Step 3: A method combining a deep feature fusion network and a meta-learning training mechanism is proposed to address the small sample size problem in voltage overshoot scenarios. By using the GhostNet network for deep feature extraction and training it with the Model-Agnostic Meta-Learning (MAML) algorithm, the network can quickly adapt to new tasks with small sample data, thereby improving the accuracy and efficiency of voltage tracing.

[0058] The core objective of this step is to build an efficient and accurate few-sample voltage tracing model based on meta-learning and deep feature fusion networks. The fast adaptation problem in voltage tracing is addressed using a lightweight GhostNet neural network and the MAML algorithm.

[0059] In voltage source tracing tasks, the network needs to extract key spatiotemporal features from multidimensional voltage signals. These features can effectively characterize voltage exceedance events. Therefore, the GhostNet network is used for feature extraction, and the specific process consists of the following steps:

[0060] 1) Main Feature Extraction: Extracting the main feature from the input voltage time-frequency feature matrix. X Perform convolution operations to extract the main features of low-frequency and high-frequency signals:

[0061] ;

[0062] in, W 1 represents the convolution kernel weight, responsible for extracting the main frequency features of the voltage signal. Conv stands for convolution operation, which is typically used for feature extraction of images or time-series data. This step helps the network capture key temporal changes in the voltage signal.

[0063] 2) Secondary feature extraction: After obtaining the main features, further processing of the main features is performed. H 1. Simplify the process to obtain secondary features. H 2:

[0064] ;

[0065] in, W 2 represents the weights for the second-layer feature extraction. CheapOp is an optimization operation in GhostNet that simplifies the extracted features, reducing computational complexity. In this way, the network can capture deeper feature information, further enhancing its ability to represent voltage signals.

[0066] 3) Feature fusion: Combining main features H 1 and secondary features H 2. Perform fusion to obtain the final feature representation. Z :

[0067] ;

[0068] This feature vector contains multidimensional information about the voltage signal in the time and frequency domain, which can comprehensively characterize the spatiotemporal characteristics of voltage over-limit, providing a foundation for subsequent voltage source tracing.

[0069] To enhance the model's ability to adapt quickly to new tasks, this step also employs the MAML meta-learning training mechanism. This mechanism aims to enable the network to rapidly adapt to new environments when faced with small sample data, thereby enhancing its adaptability in voltage tracing tasks. The meta-learning process is as follows:

[0070] 1) Training phase: In each task In the first step, the currently shared initial model parameters are used. θ Perform a gradient update on the training set for the task to obtain the task-specific parameters corresponding to that task. :

[0071] ;

[0072] in, For learning rate, Indicates the first i Loss function on the training set for each task For the loss function with respect to the parameters θ The gradient.

[0073] 2) Verification phase: During the task On the validation set, use the updated parameters. Calculate the verification loss :

[0074] ;

[0075] in, For the task The validation set To update the model on validation samples The above prediction For predicted values ​​and true labels The losses between them.

[0076] 3) Final optimization phase: Summarize the validation losses of all tasks and calculate the meta-loss function. :

[0077] ;

[0078] in, This represents the number of tasks included in the meta-learning phase.

[0079] By minimizing the meta-loss function, the shared initial model parameters are... θ Optimize to obtain the optimal initial parameters that enable rapid adaptation. :

[0080] ;

[0081] in, These are the optimal initial parameters obtained by optimizing the meta-loss function.

[0082] Through the above training process, the network can quickly adjust its parameters with a small number of new samples, improving its adaptability to new voltage limit-breaking tasks.

[0083] Step 4: Combining the voltage signal features extracted in Steps 1 and 2 with other auxiliary data, complete the data preprocessing and concatenation to construct the training dataset for the neural network. By inputting the training data, utilize the aforementioned deep feature fusion network and meta-learning-driven training mechanism to train the model, ultimately optimizing the voltage over-limit tracing model. The trained model is deployed in a real-world distribution network environment, enabling real-time voltage tracing and monitoring, ensuring the operational stability of the distribution network.

[0084] At the beginning of this step, the multidimensional auxiliary data collected in step 1 and the voltage time-frequency features extracted in step 2 are concatenated to form a complete training dataset. The data concatenation process is as follows:

[0085] Voltage signal characteristics: using the time-frequency characteristic matrix obtained in step 2 X This matrix contains the multi-dimensional characteristics of the voltage signal in the time-frequency domain.

[0086] Other auxiliary data: The power grid data (active and reactive power) and load data collected in step 1 are used as supplementary features. These data have a direct impact on voltage source tracing.

[0087] Finally, the spliced ​​dataset The input feature set that makes up the neural network is represented as:

[0088] ;

[0089] in, It is the time-frequency characteristic matrix of the voltage signal. This refers to the active power data of the power grid. This refers to the reactive power data of the power grid. This is the load data. By stitching these multidimensional features together, comprehensive input information can be provided to the neural network.

[0090] After data preparation, the training phase of the neural network begins. During training, we use the deep feature fusion network (GhostNet) designed in the aforementioned steps, combined with the MAML meta-learning training mechanism, to ensure that the network can quickly adapt to voltage limit exceedance scenarios with small sample sizes. The specific training steps are as follows:

[0091] Initialize network parameters: Initialize the model parameters according to the network architecture. The network includes multiple convolutional layers and meta-learning modules, which are used to extract the time-frequency features of the voltage signal and perform source tracing.

[0092] Forward propagation: the concatenated dataset The input is processed by a neural network, including convolutional layers and feature fusion layers, and the output is the voltage over-limit prediction result.

[0093] Calculate the loss function: Compare the true labels of voltage exceedances with the model predictions to calculate the mean squared error (MSE), specifically in the form of:

[0094] ;

[0095] in, It is the actual label. It is the result of network prediction. N 1 represents the sample size.

[0096] Backpropagation and parameter update: The network parameters are backpropagated using gradient descent to update the network weights and minimize the loss function.

[0097] ;

[0098] in, For loss function Relative to parameters θ gradient, These are the updated parameters.

[0099] After the neural network training is completed, the model will be deployed to the actual operating environment of the power distribution network for real-time voltage over-limit tracing.

[0100] Example 2:

[0101] Reference Figure 2 As an embodiment of the present invention, a distribution network voltage tracing system based on adaptive feature extraction and meta-learning is provided, including a data acquisition module, a feature extraction and time-frequency analysis module, a deep feature fusion network module, a meta-learning training module, and a model deployment module.

[0102] The data acquisition module is used to collect data on the operating status of the distribution network. The types of data it collects include, but are not limited to, voltage amplitude, waveform, active and reactive power, load data, and distributed generation output data at each node of the distribution network. This data provides the foundation for subsequent feature extraction and neural network model training. Through sensors and data interfaces, the module ensures the acquisition of real-time and historical operating data of the distribution network, guaranteeing data diversity and high quality, and providing reliable input for voltage tracing.

[0103] The feature extraction and time-frequency analysis module is responsible for processing and extracting features from the acquired voltage signal. By using methods such as adaptive eigenmode decomposition (AFMD), the original voltage signal is deconstructed in the time and frequency domain to extract the intrinsic characteristics (such as instantaneous amplitude and frequency) of the voltage fluctuation.

[0104] The deep feature fusion network module is used to perform feature fusion and voltage source tracing based on the spliced ​​time-frequency feature data using the GhostNet deep neural network. This network, through convolutional layers and a multi-level feature extraction mechanism, can extract the most discriminative features from the voltage signal.

[0105] The meta-learning training module improves the adaptive ability of neural networks with small sample data through the MAML meta-learning algorithm.

[0106] The model deployment module is responsible for deploying the trained voltage tracing model to the actual operation platform of the distribution network for real-time prediction and monitoring.

Claims

1. A power distribution network voltage tracing method based on adaptive feature extraction and MAML, characterized in that, Comprise the following steps: Step 1: Collect key operation data including voltage amplitude, waveform, active and reactive power, load and distributed power output of each node in the power distribution network, or generate multiple scenario voltage samples by simulation means as auxiliary data for subsequent neural network model training; Step 2: Adopt adaptive feature modal decomposition based voltage signal time-frequency domain decomposition to extract voltage signal features as key disturbance features implied in voltage excursion events, and realize high-precision excursion feature modeling; Step 3: Establish a deep feature fusion network and a meta-learning driven training mechanism to solve the small sample problem under voltage excursion scenarios; Specifically, through the GhostNet network for deep feature extraction, and combining the MAML meta-learning algorithm for training, the network can quickly adapt to new tasks under small sample data; Step 4: Combine the auxiliary data and voltage signal features to complete data preprocessing and splicing, and build a neural network training dataset; Then use the deep feature fusion network and the meta-learning driven training mechanism of step 3 to train the model, and complete the optimization of the voltage excursion tracing model; The trained model is deployed to the actual power distribution network monitoring and analysis platform for real-time voltage tracing and monitoring analysis.

2. The power grid voltage tracing method based on adaptive feature extraction and MAML according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1: Data type determination; According to the voltage tracing task requirements of the power distribution network, determine the required data types, which specifically include: Voltage amplitude and waveform data: covering voltage change information of all key nodes, used to describe voltage fluctuation trend; Active and reactive power data: used to represent the influence of source and load changes on voltage, reflecting the operation state of the power grid; Distributed power output data: including real-time power data of photovoltaic and wind power distributed power sources, used to identify voltage disturbance sources; Load data: time series variation characteristics of node load, used to reveal the effect of load disturbance on voltage excursion; Step 1.2: Data acquisition method design; According to the system application scenario, the data comes from the following two ways: Actual data acquisition: obtain actual operation data through dispatching system, smart meter or PMU device; Simulation data generation: use power system simulation platform to build a typical power distribution network model containing distributed power sources, set different load disturbance and distributed power output scenarios, and simulate corresponding voltage response data; Step 1.3: Data preprocessing and standardization; Clean, denoise, fill in missing values, align the operation of unified time scale, and process data features by normalization or standardization method to improve the numerical stability and convergence speed of neural network training; Step 1.4: Training sample construction; According to the time series characteristics and node correlation of voltage excursion events, construct the sample dataset required for neural network training; Each training sample includes multi-dimensional input features and corresponding target labels within a time window.

3. The power grid voltage tracing method based on adaptive feature extraction and MAML according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1: Let the single-phase voltage signal of a certain node be where t is the time variable, u denotes the voltage amplitude of the node at time t ; the original single-phase voltage signal is decomposed into several intrinsic mode functions and a residual signal by using an adaptive empirical mode decomposition algorithm, and the formula is as follows: ； wherein, is the number of the first modal component; k is the number of the first modal component; K is the number of the total modal components; is the residual term, representing non-periodic variation trends or long-term disturbances; Step 2.2: For each modal component Performing Hilbert transform to get its instantaneous amplitude and instantaneous frequency , reflecting the local energy and frequency variation of each mode at different time scales; further constructing a two-dimensional time-frequency feature matrix to structure the amplitude and frequency information, the voltage time-frequency feature matrix as follows: ； wherein T denotes the total number of time steps; the k column lists the amplitude variation of the k modal at all time steps; each row represents the instantaneous energy feature combination of different modal components at a certain time instant.

4. The power grid voltage tracing method based on adaptive feature extraction and MAML according to claim 3, characterized in that, Step 3 specifically includes: Step 3.1: Feature extraction using GhostNet network; Main feature extraction: on the input voltage time-frequency feature matrix X Convolution operation is performed to extract the main features of low and high frequency signals H 1: ； wherein, W 1is the convolution kernel weight, responsible for extracting the main frequency characteristics of the voltage signal; Conv represents the convolution operation; Secondary feature extraction: Simplify the primary feature H 1 to get secondary feature H 2: ； wherein, W 2 is the second layer feature extraction weight; CheapOp is an optimization operation in GhostNet, which is used to simplify the extracted features and reduce the computational complexity; Feature fusion: fuse the primary feature H 1 and the secondary feature H 2 to obtain the final feature representation Z : ； Step 3.2: Use MAML meta-learning training mechanism to improve the model's rapid adaptation ability to new tasks, the meta-learning process is as follows: Training phase: In each task , first use the current shared initial model parameters θ to perform one gradient update on the training set of the task, obtaining the task-specific parameters corresponding to the current task: ； wherein, is a learning rate, denotes the loss function on the i th task training set, is the gradient of the loss function with respect to the parameters θ ; Validation phase: On the validation set of the task , the updated parameters are used to compute the validation loss : ； wherein, is the task validation set, is the prediction of the updated model on the validation samples is the loss between the predicted values and the true labels is the loss between the predicted values and the true labels ​ Final optimization phase: aggregate validation loss of all tasks, compute meta-loss function : ； wherein, is the number of tasks contained in the meta-learning phase; by minimizing the meta-loss function, the shared initial model parameters θ are optimized to obtain optimal initial parameters with fast adaptation capability : ； wherein, is the optimal initial parameter obtained by optimizing the meta-loss function.

5. The power grid voltage tracing method based on adaptive feature extraction and MAML according to claim 4, characterized in that, Step 4 specifically includes: Step 4.1: Concatenate the auxiliary data and the voltage signal features, the voltage signal features are the time-frequency feature matrix obtained in step 2 X , the auxiliary data are the grid power data and the load data collected in step 1; the concatenated dataset is used as the input feature set of the neural network, denoted as: ； wherein is a time-frequency feature matrix of the voltage signal, is active power data of the power grid, is reactive power data of the power grid, is load data; Step 4.2: Use a deep feature fusion network and combine MAML meta-learning training mechanism to train the model, the specific process is as follows: Step 4.2.1: Initialize network parameters; Initialize the parameters of the model according to the network architecture, the network includes multiple convolutional layers and meta-learning modules, which are used to extract time-frequency features of voltage signals and trace the source; Step 4.2.2: Forward propagation; The concatenated dataset The input neural network outputs the voltage out-of-limit prediction result after processing by the convolutional layer and the feature fusion layer. Step 4.2.3: Calculate the loss function; The real labels using voltage out-of-limits are compared with the model prediction results, and the mean square error is calculated as a loss function In particular, it is: ； wherein, is the actual label, is the network prediction, N 1 is the number of samples; Step 4.2.4: Back propagation and parameter update; The network parameters are updated by gradient descent method, and the network weights are updated to minimize the loss function, as shown in the following formula: ； where, is the learning rate, is the loss function is the parameter, θ is the gradient, is the updated parameter.

6. The power grid voltage tracing system based on adaptive feature extraction and MAML, characterized in that, It includes data acquisition module, feature extraction and time-frequency analysis module, deep feature fusion network module, meta-learning training module and model deployment module. The data acquisition module is used to collect the data of the power distribution network operation state; The collected data types include voltage amplitude, waveform, active and reactive power, load data and distributed power output data of each node of the power distribution network, which are used as auxiliary data for neural network model training. The feature extraction and time-frequency analysis module is responsible for processing and extracting features from the collected voltage signals. By using the adaptive feature modal decomposition method, the original voltage signal is decomposed in the time-frequency domain, and the intrinsic characteristics of the voltage fluctuation are extracted to obtain the voltage time-frequency feature data. The deep feature fusion network module is based on the spliced voltage time-frequency feature data and auxiliary data, and uses the deep neural network GhostNet for feature fusion and voltage tracing; The neural network extracts the most discriminative features from the voltage signal through convolutional layers and multi-level feature extraction mechanisms. The meta-learning training module improves the adaptive ability of the neural network under small sample data through the MAML meta-learning algorithm. The model deployment module is used to deploy the trained voltage tracing model to the actual operation platform of the power distribution network for real-time prediction and monitoring.

Images