An intelligent waveform measurement and perception method and system for comprehensive analysis of all elements at the grid side

Abstract

The present invention discloses an intelligent waveform measurement and perception method and system for full-element analysis of the power grid end-edge. The method of the present invention includes locally training a federated network model using real-time synchronized waveform data collected by local waveform measurement devices on end-side devices, and on the edge-side server, performing multi-device collaborative optimization on the model parameters locally trained by all end-side devices and sending the model parameters to the end-side devices; each end-side device updates the local federated network model according to the sent model parameters and performs local waveform fault perception, and on the edge-side server, waveform fault perception is performed on the power grid parameters uploaded by all end-side devices to obtain a global waveform fault perception result. The present invention aims to solve the problems of the existing power grid monitoring system in identifying and responding to power grid disturbances under the conditions of high latency, low accuracy, and lack of effective cooperation, and improve the real-time perception capabilities of fault monitoring, alarm, and prediction, etc. of the new power system.

Description

Technical Field

[0001] The present invention relates to the technical field of smart grids, and particularly to an intelligent waveform measurement and perception method and system for comprehensive element analysis at the grid edge and the grid end. Background Art

[0002] With the increasing proportion of new energy, the power grid faces an increasingly complex operating environment and non-linear dynamic disturbances, which makes the traditional power grid monitoring technology gradually unable to meet the current power grid's requirements for safe, stable and efficient operation. The existing power grid monitoring methods have the following problems: high latency, low accuracy, and limited edge-end collaboration capabilities. Traditional power grid monitoring methods have a high latency in processing and responding to power grid disturbances and cannot meet the real-time response requirements for rapid disturbances; in a complex power system, the existing systems have problems with insufficient accuracy in detecting and locating high-frequency disturbances. Especially in the case of new energy access, the accurate monitoring of power grid parameters is particularly important; the existing power grid monitoring systems mainly rely on centralized computing and control, lacking efficient collaboration between edge-side and end-side devices, resulting in low data utilization rate and unable to achieve efficient and accurate situation awareness. Therefore, there is an urgent need for a new power grid monitoring technology that can collect comprehensive element data and perform local calculations at the end side, and improve the recognition and response capabilities of power grid disturbances through edge-end collaboration. Summary of the Invention

[0003] The technical problem to be solved by the present invention: In view of the above problems of the prior art, an intelligent waveform measurement and perception method and system for comprehensive element analysis at the grid edge and the grid end are provided. The present invention aims to solve the problems of the existing power grid monitoring system in recognizing and responding to power grid disturbances under the conditions of high latency, low accuracy, and lack of effective collaboration, and improve the real-time situation awareness capabilities such as fault monitoring, alarm, and prediction of the new power system.

[0004] To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0005] An intelligent waveform measurement and perception method for comprehensive element analysis at the grid edge and the grid end, comprising the following steps:

[0006] S1. Use the real-time synchronized waveform data collected by the local waveform measurement device on the end-side device to perform local training of the federated network model, and upload the model parameters obtained from the local training to the edge-side server. The input data of the federated network model is the power grid parameters extracted from the real-time synchronized waveform data, and the output result is the local situation awareness result;

[0007] S2. Receive, on the end-side device, the model parameters sent after multi-device collaborative optimization of the model parameters obtained from the local training of all end-side devices on the edge-side server.

[0008] S3. On the edge device, update the local federated network model according to the model parameters sent by the edge server, and use the federated network model after updating the model parameters to perform local waveform fault perception on the real-time synchronized waveform data collected by the local waveform measurement device, and upload the grid parameters and the local situation perception results to the edge server, so that the edge server can perform global waveform fault perception on the grid parameters uploaded by all edge devices using the edge perception model to obtain the global waveform fault perception results.

[0009] Optionally, the function expression for multi-device collaborative optimization in step S2 is:

[0010] ,

[0011] ,

[0012] Among them, is the model parameter after multi-device collaborative optimization, is the number of edge devices, is the total number of data samples, is the edge device The model parameter before multi-device collaborative optimization of, is the edge device The number of samples of the real-time synchronized waveform data of.

[0013] Optionally, the federated network model includes a feature extractor w 0 deployed on the edge device, s a collaborative sensor w 1 and a feature processor w 0 deployed on the edge server. The feature extractor w 0 is used to extract features from the grid parameters h 0 The feature processor w 1 is used to further extract the features h 0 from the features h 1 The collaborative sensor w s is used to perform perception prediction on the features h 1 to obtain the predicted waveform measurement perception results.

[0014] Optionally, before step S1, it further includes splitting and deploying the feature extractor w 0 and the feature processor w 1 :

[0015] S101. Use knowledge distillation to reduce the number of model parameters of the feature extraction network model composed of a multi-layer neural network;

[0016] S102, Define the model splitting ratio :

[0017] ,

[0018] Among them, is the number of parameters of the feature extractor w 0 , and is the total number of parameters of the feature extractor w 0 and the feature processor w 1 . Determine the optimal model splitting ratio according to the computing power and memory limit of the edge device :

[0019] ,

[0020] Among them, and are the maximum and minimum model splitting ratios determined according to the memory limit respectively, and are the computing power of the edge device and the edge server respectively, and there is:

[0021] ,

[0022] ,

[0023] Among them, represents the infimum, is the model splitting ratio corresponding to the parameter storage size of the edge device, is the maximum available memory of the edge device, and are the number of parameters of the first layer and the last layer of the feature extraction network model respectively;

[0024] S103, According to the optimal model splitting ratio split the feature extraction model into the feature extractor w 0 and the feature processor w 1 , and deploy the feature extractor w 0 on the edge device, and deploy the feature processor w 1 on the edge device.

[0025] Optionally, when locally training the federated network model on the edge device using real-time synchronized waveform data in step S1, the objective function used for local training is:

[0026] ,

[0027] Among them, is the edge device kThe objective function, is the model parameter of the edge device k ; is the number of samples of the real-time synchronized waveform data of the edge device k ; is the loss function used in local training, and the federated network model is based on the i th input signal and the model parameter to obtain the waveform measurement perception result; is the i th input signal corresponding ideal waveform measurement perception result.

[0028] Optionally, an edge sensor w e is also deployed on the edge device. The edge sensor w e is used to perform perception prediction on the feature h 0 when a communication failure occurs between the edge device and the edge server to obtain the output result of the federated network model; the local training also includes training the auxiliary model composed of the feature extractor w 0 and the edge sensor w e to update the model parameters of the feature extractor w 0 and the edge sensor w e , and the function expression of the loss function used in training the auxiliary model is:

[0029] ,

[0030] where is the number of samples, is the waveform measurement perception result predicted by the edge sensor w e for the i-th sample, and is the true label value of the waveform measurement perception result of the i-th sample.

[0031] Optionally, the edge perception model includes a feature processor w 2 and an edge sensor w P . The feature processor w 2 is used to mine the feature h 1 from the feature h P2 . The edge sensor w P is used to predict and generate the edge perception value h P2 based on the feature y pas the global waveform measurement and perception result.

[0032] In addition, the present invention also provides an intelligent waveform measurement and perception system for comprehensive element analysis at the grid edge, including an edge server and a plurality of end devices respectively connected to the edge server, and the end devices are programmed or configured to execute the intelligent waveform measurement and perception method for comprehensive element analysis at the grid edge.

[0033] In addition, the present invention also provides a computer-readable storage medium, in which a computer program or instruction is stored, and the computer program or instruction is programmed or configured to execute the intelligent waveform measurement and perception method for comprehensive element analysis at the grid edge through a processor.

[0034] In addition, the present invention also provides a computer program product, including a computer program or instruction, and the computer program or instruction is programmed or configured to execute the intelligent waveform measurement and perception method for comprehensive element analysis at the grid edge through a processor.

[0035] Compared with the prior art, the present invention can mainly achieve the following beneficial effects: The method of the present invention includes locally training a federated network model using real-time synchronous waveform data collected by local waveform measurement devices on end devices, and on the edge server, co-optimizing the model parameters obtained from the local training of all end devices and distributing the model parameters to the end devices; each end device updates the local federated network model according to the distributed model parameters and performs local waveform fault perception, and on the edge server, waveform fault perception is performed on the grid parameters uploaded by all end devices to obtain the global waveform fault perception result. The method of the present invention can solve the problems of the existing power grid monitoring system in identifying and responding to power grid disturbances under the conditions of high latency, low accuracy, and lack of effective cooperation by constructing an efficient edge-end collaborative architecture, perform comprehensive element data collection and local calculation at the end side, improve the ability to identify and respond to power grid disturbances through edge-end collaboration, enhance the real-time performance and accuracy of power grid disturbance monitoring in the new power system, optimize the collaborative efficiency between end devices and edge devices, and at the same time solve the problems of data transmission delay and low utilization efficiency of computing resources through the deployment of a lightweight federated network, and effectively solve the difficult problems of rapid detection and high-precision identification of dynamic disturbances in the new power system while ensuring the privacy and security of user data, and can be used for situation awareness such as monitoring, alarming, and prediction of various faults in the new power system. Description of the Drawings

[0036] Figure 1 It is a schematic diagram of the basic process of the method of the embodiment of the present invention.

[0037] Figure 2 It is a schematic diagram of a working process of an end device and an edge server in the embodiment of the present invention.

[0038] Figure 3 This is a schematic diagram of a framework structure for the end - side device and its waveform measurement device in an embodiment of the present invention.

[0039] Figure 4 This is a schematic layout diagram of the federated network model and the edge - side perception model in an embodiment of the present invention.

[0040] Figure 5 This is a schematic diagram of the system structure in an embodiment of the present invention. Detailed implementation manners

[0041] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be further described in detail below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0042] As Figure 1 shown, the intelligent waveform measurement and perception method for the full - element analysis of the power grid end - edge in this embodiment includes the following steps:

[0043] S1. Use the real - time synchronous waveform data collected by the local waveform measurement device on the end - side device to perform local training of the federated network model, and upload the model parameters obtained from the local training to the edge - side server. The input data of the federated network model is the power grid parameters extracted from the real - time synchronous waveform data, and the output result is the local situation awareness result.

[0044] S2. Receive, on the end - side device, the model parameters sent down after multi - device collaborative optimization of the model parameters locally trained by all end - side devices on the edge - side server.

[0045] S3. Update the local federated network model on the end - side device according to the model parameters sent down by the edge - side server, and use the federated network model with updated model parameters to perform local waveform fault perception on the real - time synchronous waveform data collected by the local waveform measurement device, and upload the power grid parameters and the local situation awareness result to the edge - side server, so that the edge - side server can perform global waveform fault perception on the power grid parameters uploaded by all end - side devices using the edge - side perception model to obtain the global waveform fault perception result.

[0046] As Figure 2As shown, in this embodiment, an efficient end-edge collaboration architecture is constructed through end-side devices and edge-side servers. The end-side devices use the real-time synchronized waveform data collected by local waveform measurement devices to perform local training of the federated network model and local waveform fault perception; the edge-side servers perform multi-device collaborative optimization and use the edge-side perception model to perform global waveform fault perception to obtain the global waveform fault perception result, thus forming a power grid end-edge all-element analysis system. In one or more embodiments, the working process of the power grid end-edge all-element analysis system includes: using the end-side devices to collect power grid voltage and current signals in real time, calculating all-element electrical parameters such as phase angle, frequency, amplitude, and frequency change rate; deploying a lightweight federated network module in the end-side devices, realizing intelligent analysis of the power grid panoramic data through local situation awareness, and uploading the data to the edge-side devices after lossless compression; the edge-side devices integrate the model parameters uploaded by multiple end-side devices, adopt a synchronous aggregation strategy for model training, optimize the federated network model parameters and send them to the end-side devices, and finally realize the intelligent monitoring of power grid parameters by the end-side devices, as well as various applications such as fault location, oscillation monitoring, load prediction, and inertia estimation. It comprehensively supports power grid parameter measurement, intelligent analysis, and multi-scenario applications, improving the intelligent level and real-time performance of power grid situation awareness.

[0047] Figure 3 This is a schematic diagram of the framework structure of the end-side device and its waveform measurement device in this embodiment. In one or more embodiments, the edge side in this embodiment includes a waveform measurement device and an edge-side device, which are used for power grid end-edge all-element analysis. Among them, the waveform measurement device is implemented by a field programmable gate array (FPGA). The FPGA is used to perform high-frequency synchronous sampling on voltage and current signals, and calculate key electrical parameters such as phase angle, frequency, and amplitude in real time to ensure the accuracy and timeliness of data acquisition. Through real-time synchronous sampling, the field programmable gate array FPGA can process signals up to several hundred hertz, ensuring the accuracy and timeliness of data acquisition.

[0048] The end-side device can be implemented on platforms such as Raspberry Pi and STM32. Local analysis is performed by running the federated network to reduce upload delay, and the model performance is optimized through binarization and knowledge distillation methods. The data is uploaded to the edge-side device or server through an Ethernet or wireless communication module for regional power grid situation awareness. The intelligent waveform measurement and perception device can calculate all-element electrical parameters such as phase angle, frequency, amplitude, frequency change rate, broadband oscillation parameters, and fault occurrence time in real time, providing millisecond-level dynamic response to support power grid situation awareness and fault location. Finally, the data is transmitted to the edge-side server through Ethernet (TCP / IP) or a wireless communication module (such as Wi-Fi, 4G, 5G) for local power grid situation awareness.

[0049] In step S1 of this embodiment, when locally training the federated network model using real-time synchronized waveform data on the edge device, the objective function adopted for local training is:

[0050] ,

[0051] where, is the objective function of the edge device k , are the model parameters of the edge device k , is the number of samples of the real-time synchronized waveform data of the edge device k , is the loss function adopted during local training, is the waveform measurement perception result of the federated network model based on the i th input signal and the model parameters , is the i rd input signal corresponding ideal waveform measurement perception result. Among them, the loss function adopted during local training can adopt the required loss function according to needs, such as cross-entropy loss, etc.

[0052] As an alternative implementation, the input power grid parameters of the federated network model in this embodiment include the phase angle, which refers to the phase difference between voltage and current and is used to judge the change of power grid load and power factor; the frequency, which refers to the change of power grid frequency, and too large fluctuations in frequency indicate abnormal power grid operation; the amplitude, which refers to the amplitude of voltage and current, and voltage exceeding ±0.05 p.u. is regarded as voltage over-limit; the frequency change rate, which refers to the rate of frequency change and reflects the change trend of power grid disturbance. As an alternative implementation, the local situation awareness result of the federated network model in this embodiment can be one of short-term frequency fluctuation, load, disturbance, wide-frequency oscillation parameters, and fault information. Among them, the wide-frequency oscillation parameters refer to the frequency, damping ratio, and mode number of power grid wide-frequency oscillation; the fault information includes the fault occurrence time and fault type. The fault occurrence time refers to the start and end times of different fault events, and the fault types include faults such as new energy power source switching / high-frequency switch / line short circuit / load shedding in the distribution network under distributed new energy access. The intelligent waveform measurement perception device for comprehensive analysis of all elements at the grid edge can provide dynamic parameters of the power grid at the millisecond level through real-time calculation of the edge-side synchronous phasor measurement unit FPGA module, and these parameters will be used for subsequent situation awareness and fault location. Federated learning enables the device to perform real-time model update and adjustment according to local data through distributed learning and local training, without uploading all data to the cloud, which can reduce network bandwidth consumption and improve real-time performance.

[0053] In this embodiment, an edge-side parallel collaborative processing mechanism is implemented through the edge-side device and the edge-side server. By optimizing the distributed training and parallel computing of the device, the efficiency and real-time performance of power grid situation awareness are achieved. The edge-side device uses real-time power grid measurement data for feature extraction and local model training. The edge-side device completes the optimization and update of the global model by receiving the model parameters uploaded by the edge-side device, and ensures the parameter consistency and convergence of the global model through the synchronous aggregation strategy. This method significantly optimizes the utilization rate of distributed computing resources, has characteristics such as strong computing performance and resource friendliness, and provides strong technical support for real-time state awareness and fault location in a dynamic power grid environment.

[0054] On the edge-side server, all the data uploaded by the edge-side devices will be integrated, summarized and used for local training. On the edge-side device, multi-device collaborative optimization is performed based on the uploaded edge-side data, and the optimized model parameters are sent to the edge-side device. Through this edge-side collaborative processing mechanism, the system can maintain an efficient state awareness ability and ensure that the device always maintains the best parameter detection and fault location ability in a dynamic power grid environment. To ensure the consistency and global convergence of the edge-side model update, a synchronous aggregation strategy is adopted for the edge-side model and the edge-side model. Synchronous aggregation is a global model update method that requires all devices to complete local training before aggregating the parameters of the global model. The synchronous aggregation strategy adopted for multi-device collaborative optimization in step S2 of this embodiment has the following functional expression:

[0055] ,

[0056] ,

[0057] where, is the model parameter after multi-device collaborative optimization, is the number of edge-side devices, is the total number of data samples, is the edge-side device model parameter before multi-device collaborative optimization, is the edge-side device number of samples of real-time synchronous waveform data.

[0058] As Figure 4 shown, the federated network model in this embodiment includes a feature extractor w 0 deployed on the edge-side device, s a collaborative sensor w 1 and a feature processor w 0 deployed on the edge-side server. The feature extractor w 0 is used to extract features from power grid parameters h 0 , and the feature processor w 1 is used to process the featuresh 0 Further mined features h 1 , collaborative sensor w s For features h 1 Perform perception prediction to obtain the predicted waveform measurement perception result. As Figure 4 Shown, the end-side device in this embodiment is also deployed with an end sensor w e , the end sensor w e For when a communication failure occurs between the end-side device and the edge-side server, perform perception prediction on the features h 0 To obtain the output result of the federated network model; the local training also includes training the auxiliary model composed of the feature extractor w 0 , end sensor w e To update the model parameters of the feature extractor w 0 , end sensor w e , and the loss function used when training the auxiliary model The function expression of is:

[0059] ,

[0060] Among them, Is the number of samples, Is the waveform measurement perception result predicted by the end sensor w e For the first Sample, Is the first True label value of the waveform measurement perception result of the sample.

[0061] In this embodiment, the federated network model is divided into a main model (the feature extractor w deployed on the end-side device 0 , collaborative sensor w s And the feature processor w deployed on the edge-side server 1 ) and an auxiliary model (the feature extractor w deployed on the end-side device 0 And end sensor w e ), and run on the edge-side server and the end-side device respectively, which not only realizes end-edge collaborative training, but also significantly optimizes the utilization rate of distributed computing resources. The end-side device first uses grid data x As input, obtain the initial feature representation through the phasor calculation method h 0 = f ( w 0 , x ). Subsequently, the feature processor in the edge-side server w1 For further mining of the representation, we obtain h 1 = f ( w 1 , h 0 ). Based on this, the main model generates a predicted output y s = f (w s , h 1 ). Meanwhile, the auxiliary model directly generates local prediction perception on the edge device y e = f ( w e , h 0 ). At the same time, on the edge server, further feature perception can be performed on the parameters in the feature processor w 1 and the edge-side sensor w 2 to obtain the edge-side model perception quantity w P y p . Compared with the traditional model segmentation training method, the edge-side parallel coordination processing design has the following two major innovation points: 1) Reduced communication overhead: By decoupling the gradient calculation from w 1 , the edge server does not need to return the gradients of backpropagation to the edge device, thus significantly reducing the communication volume and saving about one-fourth of the communication overhead. 2) Parallel computing acceleration: With edge-side parallel coordination processing, the gradients of the feature extractor w 0 on the edge device and w 1 on the edge server can be calculated simultaneously, thus shortening the time required for backpropagation by nearly half.

[0062] Figure 4 As w shown, the edge-side perception model in this embodiment includes a feature processor w 2 and an edge-side sensor w P , and the feature processor w 2 is used to mine the feature h 1 h P2 P from the feature w , and the edge-side sensor h P2 is used to predict and generate the edge-side perception value y p ​As the global waveform measurement perception result, the training of the edge-side perception model is for the feature extractor w 0 , the feature processor w 1 , the feature processor w 2 and the edge-side sensor w P to train the end-to-end network model composed of them. Specifically, the required loss function and training method can be adopted according to needs, which will not be elaborated here. Among them, the end sensor w e , the collaborative sensor w s , and the edge-side sensor w P are all existing well-known network models.

[0063] The implementation of the federated network model is mainly divided into four parts: model construction and initialization, local training, global optimization, and model distribution. Model construction and initialization: Construct a lightweight federated network suitable for the computing power of the edge side, and use binary quantization and knowledge distillation methods to reduce the number of model parameters on the edge side; according to the device resource conditions on the edge side, divide the edge-side and end-side models and determine the segmentation ratio , to achieve model compression. Local training: The edge-side device uses local data to train the feature extractor and uploads the intermediate parameters to the edge-side server. Global optimization: The edge-side server receives the update results of all edge-side devices and updates the global model through synchronous aggregation. Model distribution: The edge-side server distributes the optimized model to the edge-side devices to start a new round of training. According to the computing power, memory limit, and communication conditions of the edge-side devices, the complete deep learning model needs to be split into two parts for model segmentation to achieve model initialization. The model is split into two parts: the edge-side model part and the end-side model part. Edge-side model part: Deployed on the edge-side device for feature extraction and preliminary calculation; End-side model part: Deployed on the edge-side server for inference tasks with high computational complexity.

[0064] In this embodiment, before step S1, it also includes splitting and deploying the feature extractor w 0 and the feature processor w 1 :

[0065] S101, use knowledge distillation to reduce the number of model parameters of the feature extraction network model composed of a multi-layer neural network;

[0066] S102, define the model segmentation ratio :

[0067] ,

[0068] Among them, is the number of parameters of the feature extractor w 0 , is the feature extractor w0 and the feature processor w 1 The total number of parameters of, determine the optimal model splitting ratio according to the computing power and memory limit of the edge device :

[0069] ,

[0070] Among them, and are the maximum and minimum model splitting ratios determined according to the memory limit respectively, and are the computing power of the edge device and the edge server respectively, and there are:

[0071] ,

[0072] ,

[0073] Among them, represents the infimum, is the model splitting ratio The corresponding parameter storage size of the edge device, is the maximum available memory of the edge device, and are the number of parameters of the first layer and the last layer of the feature extraction network model respectively; among them, the model splitting ratio The corresponding parameter storage size of the edge device is the product of the storage size and the number of parameters of the model parameters, is the number of layers of the feature extraction network model;

[0074] S103, according to the optimal model splitting ratio Split the feature extraction model into the feature extractor w 0 and the feature processor w 1 , and the feature extractor w 0 Deploy on the edge device, and deploy the feature processor w 1 Deploy on the edge device.

[0075] In summary, the method of this embodiment has the following characteristics: Lightweight federated network design and optimization: A lightweight federated learning module is deployed for local data analysis and situation awareness. On the device side, the collected panoramic power grid information (phase angle, frequency, amplitude, rate of change of frequency, wideband oscillation parameters, fault occurrence time, etc.) is input into the federated network module for local situation awareness. Through distributed learning and local training, federated learning enables devices to perform real-time model updates and adjustments based on local data without uploading all data to the cloud, reducing network bandwidth consumption and improving real-time performance. Edge-side collaborative processing and optimization: After the edge-side device losslessly compresses the data, it uploads the data to the edge-side device through a wireless communication network. The edge-side device aggregates the data from multiple edge-side devices and performs local model training and optimization based on this data. Through local training, the edge-side device can update the parameters of the federated network module and send these optimized models to the edge-side device to improve the power grid state monitoring and fault location capabilities of the entire system. Data security and sharing mechanism: In this system, each edge-side device can independently collect and process data and upload part of the data to the edge-side device. The edge-side device performs model training based on the uploaded data and optimizes the data transmission efficiency through lossless data compression technology. All edge-side devices in the system share the updated federated network model, but the data of each device is retained locally to ensure data privacy and security. The power grid edge-side all-element analysis method and system can dynamically respond to power grid state changes, improve the intelligent monitoring level of the power grid, and realize applications such as power grid disturbance detection, load forecasting, inertia estimation, fault location, and oscillation monitoring through all-element power grid information collection and real-time analysis.

[0076] As Figure 5As shown in the figure, this embodiment also provides an intelligent waveform measurement and perception system for full-element analysis of the grid edge, including an edge server and a plurality of terminal devices respectively connected to the edge server. The terminal devices are programmed or configured to execute the intelligent waveform measurement and perception method for full-element analysis of the grid edge. Among them, the terminal devices: The terminal devices are designed as intelligent waveform measurement and perception devices, which include a terminal synchronous phasor measurement unit (FPGA) and terminal devices (high-performance edge computing modules such as Raspberry Pi and microcontroller STM32 series), used to collect information such as voltage and current of the power grid in real time, calculate full-element electrical parameters such as phase angle, frequency, and amplitude, and upload the processed data to the edge device through a wireless network. The lightweight federated network module built into the terminal device can perform local situation awareness and model update. Edge server: The edge device is responsible for receiving data from multiple terminal devices, integrating and summarizing the data, and carrying out the training of the local model. The edge device sends the optimized lightweight federated network parameters to the terminal device to ensure the coordinated operation of the entire system. In the edge server, the data uploaded by all terminal devices will be integrated, summarized, and used for local training. On the edge device, based on the uploaded terminal data, the edge system optimizes the model through multiple rounds of training and sends the updated lightweight federated network parameters to the terminal device. Through this edge-to-edge collaborative processing mechanism, the system can maintain high-efficiency situation awareness capabilities and ensure that the device always maintains the best parameter detection and fault location capabilities in a dynamic power grid environment. The terminal device deploys a feature extraction module to collect signals such as voltage and current of the power grid in real time and complete feature extraction. The terminal device uses local data for local training and updates the parameters of the terminal model, with an update frequency of once a day or once a week. The terminal device uploads the extracted intermediate parameters and partial model update parameters to the edge device. On the edge device, a classification or perception module is deployed to receive the parameters uploaded by the terminal device and complete the remaining calculation tasks. At the same time, it aggregates the model updates from multiple terminal devices, optimizes the global model, and sends the updated model parameters through a synchronization strategy. The intelligent waveform measurement and perception system for full-element analysis of the grid edge in this embodiment uses synchronous information at the terminal to collect the panoramic information of the grid voltage and current; calculates full-element electrical parameters including phase angle, frequency, amplitude, and rate of change of frequency locally; deploys a lightweight federated network in the device data management module; inputs the panoramic information into the lightweight federated network module and performs local situation awareness; the terminal device compresses the data losslessly and uploads it to the edge device or the edge server; the edge device aggregates all terminal information and then performs local training, updates the parameters of the lightweight federated network module, and sends them to the terminal device for new applications such as grid parameter detection, fault location, oscillation monitoring, load forecasting, and inertia estimation. The purpose of the present invention is to improve the real-time situation awareness capabilities such as disturbance monitoring, alarm, and prediction of the new power system, and solve the problems of high latency and low accuracy in identifying new power grid disturbances in the new power system.

[0077] In addition, this embodiment also provides a computer-readable storage medium storing a computer program or instructions, which are programmed or configured to execute the intelligent waveform measurement and perception method for full-element analysis at the grid edge by a processor.

[0078] In addition, this embodiment also provides a computer program product including a computer program or instructions, which are programmed or configured to execute the intelligent waveform measurement and perception method for full-element analysis at the grid edge by a processor.

[0079] Those skilled in the art should understand that the technical solutions provided by the embodiments of the present invention can be in the form of a method, a system, or a computer program product. Therefore, the present invention can be implemented in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can be implemented in the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program code. The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for realizing the functions specified in one Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks. These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means realizes the functions specified in one Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks. These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for realizing the functions specified in one Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0080] The above are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the concept of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements should also be regarded as within the protection scope of the present invention.

Claims

1. An intelligent waveform measurement and perception method for full-factor analysis at the power grid end, characterized in that: The steps include: S1, on the end-side device, local training of the federated network model is performed using the real-time synchronous waveform data collected by the local waveform measurement device, and the model parameters obtained by the local training are uploaded to the edge-side server. The input data of the federated network model is the power grid parameters extracted based on the real-time synchronous waveform data, and the output result is the local situation awareness result; S2, receiving, on the end-side device, model parameters obtained by local training of all end-side devices on the edge-side server after multi-device collaborative optimization, the model parameters sent down; S3, the end-side device updates the local federated network model according to the model parameters sent by the edge server, and uses the federated network model with updated model parameters to perform local waveform fault perception on the real-time synchronized waveform data collected by the local waveform measurement device, and uploads the power grid parameters and local situation perception results to the edge server, so that the edge server uses the edge perception model to perform global waveform fault perception on the power grid parameters uploaded by all end-side devices to obtain global waveform fault perception results.

2. The intelligent waveform measurement and perception method for full-factor analysis of power grid end edges according to claim 1 is characterized in that: The function expression for multi-device collaborative optimization in step S2 is: ， ， in, The model parameters after multi-device collaborative optimization, is the number of devices on the end side, is the total number of data samples, For end-side devices Model parameters before multi-device collaborative optimization, For end-side devices The number of samples of real-time synchronous waveform data.

3. The intelligent waveform measurement and perception method for full-factor analysis of power grid end edges according to claim 1 is characterized in that: The federated network model includes a feature extractor w0 and a collaborative sensor w deployed on the end-side device. s The feature processor w1 deployed on the edge server and the feature extractor w0 are used to extract features of power grid parameters. h 0, feature processor w1 is used to process the feature h 0Further discovered features h 1. Collaborative Perceptron w s For features h 1 Perform perception prediction to obtain the predicted waveform measurement perception results.

4. The intelligent waveform measurement and perception method for full-factor analysis of power grid end edges according to claim 3 is characterized in that: Step S1 also includes the segmentation deployment feature extractor w0 and feature processor w1: S101, using knowledge distillation to reduce the number of model parameters of a feature extraction network model composed of a multi-layer neural network; S102, define model segmentation ratio : ， in, is the number of parameters of feature extractor w0, is the total number of parameters of the feature extractor w0 and the feature processor w1. The optimal model segmentation ratio is determined based on the computing power and memory limitations of the end device. : , in, and are the maximum and minimum model split ratios determined according to memory limitations, and They are the computing power of the client device and the edge server respectively, and they are: ， ， in, represents the infimum, is the model segmentation ratio The corresponding parameter storage size of the end-side device, The maximum value of available memory on the client device. and are the number of parameters of the first and last layers of the feature extraction network model respectively; S103, based on the optimal model segmentation ratio The feature extraction model is divided into a feature extractor w0 and a feature processor w1, and the feature extractor w0 is deployed on the end-side device, and the feature processor w1 is deployed on the end-side device.

5. The intelligent waveform measurement and perception method for full-factor analysis of power grid end edges according to claim 3 is characterized in that: When local training of the federated network model is performed on the end-side device using real-time synchronized waveform data in step S1, the objective function used in the local training is: ， in, For end-side devices k The objective function is For end-side devices k The model parameters, For end-side devices k The number of samples of real-time synchronous waveform data, is the loss function used in local training, The federated network model is based on i Input signal and model parameters The waveform measurement perception results, For the i Input signal The corresponding ideal waveform measurement perception results.

6. The intelligent waveform measurement and perception method for full-factor analysis of power grid end edges according to claim 5 is characterized in that: The terminal side device is also deployed with a terminal sensor w e , the end sensor w e Used to detect feature errors when communication failure occurs between the client device and the edge server. h 0 performs perception prediction to obtain the output result of the federated network model; the local training also includes the feature extractor w0, the end sensor w e The auxiliary model is trained to update the feature extractor w0 and the end perceptron w e The model parameters and the loss function used when training the auxiliary model The function expression is: ， in, is the sample size, The end sensor w e The waveform measurement perception result predicted for the i-th sample, is the true label value of the waveform measurement perception result of the i-th sample.

7. The intelligent waveform measurement and perception method for full-factor analysis of power grid end edges according to claim 1 is characterized in that: The side perception model includes a feature processor w2 and a side perception w P The feature processor w2 is used to h 1. Mining out features h P2 , side sensor w P Used according to characteristics h P2 Predicting and generating side perception values y p As the global waveform measurement perception result.

8. An intelligent waveform measurement and perception system for full-factor analysis of power grid edge, comprising an edge server and a plurality of edge devices respectively connected to the edge server, characterized in that: The end-side device is programmed or configured to execute the intelligent waveform measurement and perception method for full-factor analysis of the power grid end-side as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program or instruction stored therein, characterized in that: The computer program or instruction is programmed or configured to execute the intelligent waveform measurement and perception method for full-factor analysis at the power grid end as described in any one of claims 1 to 7 through a processor.

10. A computer program product comprising a computer program or instructions, characterized in that The computer program or instruction is programmed or configured to execute the intelligent waveform measurement and perception method for full-factor analysis at the power grid end as described in any one of claims 1 to 7 through a processor.

Images