An edge computing power grid operation data acquisition method and system

Abstract

The application relates to the technical field of data collection, and discloses a power grid operation data collection method and system based on edge computing, which comprises an edge computing data collection terminal and a central server; the edge computing data collection terminal comprises a sensor module, a local computing module and a communication module, the sensor module is used for collecting operation parameter data of nodes; the local computing module pre-processes the operation parameter data and identifies event information; the communication module reports the event information to the central server in real time; the central server receives the event information and generates an operation adjustment strategy; the edge computing data collection terminal dynamically adjusts a control instruction according to the operation adjustment strategy; and the central server writes all received data and event information into a time sequence database. Through construction of a distributed data collection system, the application improves the effective reduction of bandwidth load, and realizes intelligent regulation and control of a node-level sampling period, a reporting frequency and AI model parameters.

Description

Technical Field

[0001] This invention relates to the field of data acquisition technology, and more specifically, to a method and system for acquiring power grid operation data using edge computing. Background Technology

[0002] With the continuous development of smart grids and new power systems, power grid operation monitoring technology is gradually evolving towards distributed and intelligent approaches. Traditional centralized data acquisition methods are no longer sufficient to meet the high-frequency, low-latency monitoring requirements for key operating parameters such as voltage, current, frequency, and circuit breaker status. Edge computing, as a key technology supporting decentralized sensing and local intelligent processing, enables rapid local response, distributed decision-making, and event-driven data uploading, thereby enhancing the real-time monitoring and adaptive adjustment capabilities of the power grid system. Currently, solutions combining big data processing frameworks (such as Kafka and ClickHouse) with distributed scheduling mechanisms have been initially applied in some scheduling systems, becoming an important foundation for the evolution of smart grid systems to higher levels.

[0003] However, existing power grid operation data acquisition methods mostly still employ a centralized architecture, which presents several problems. For example, patent CN110287228A primarily relies on periodically downloading and parsing E-files, resulting in significant delays in acquisition response and an inability to achieve rapid sensing at the second or even millisecond level. All data must be transmitted to a central node for processing, heavily relying on network and central resources, which can easily cause bandwidth pressure and processing bottlenecks in large-scale deployment scenarios. Event recognition logic is usually based on static comparison of field values, lacking intelligent discrimination capabilities and making it difficult to accurately identify complex abnormal behaviors. Furthermore, there is a lack of operational strategy feedback and edge execution mechanisms, failing to form a complete closed-loop system of "data acquisition—event recognition—strategy generation—local control." These limitations restrict further improvements in the response timeliness, data efficiency, and intelligence level of power grid monitoring systems.

[0004] Therefore, it is necessary to design an edge computing-based method and system for acquiring power grid operation data to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a method and system for acquiring power grid operation data using edge computing, aiming to solve the problems existing in the current technology.

[0006] In one aspect, the present invention proposes an edge computing-based power grid operation data acquisition system, comprising: Edge computing data acquisition terminals and a central server, wherein the edge computing data acquisition terminals are deployed at each operating node of the power grid; The edge computing data acquisition terminal includes a sensor module, a local computing module, and a communication module. The sensor module is used to collect the operating parameter data of the node, including voltage, current, frequency, circuit breaker status, and temperature and humidity information. The local computing module preprocesses the running parameter data, including data cleaning, outlier removal, time alignment, and feature extraction. It also analyzes the parameter data based on edge AI recognition and identifies event information based on the analysis results. The event information includes mutations, exceeding limits, and disconnections. The communication module, based on event triggering, reports the identified event information to the central server in real time, and periodically uploads the batch-compressed semi-processed data to the central storage system. The compression methods include differential compression, time-series coding, or sliding window aggregation. The central server receives event information from several edge computing data acquisition terminals, performs merge analysis and strategy calculation, generates an operation adjustment strategy based on the calculation results, and sends the operation adjustment strategy to the corresponding edge computing data acquisition terminal in real time. The edge computing data acquisition terminal dynamically adjusts the data acquisition cycle, reporting frequency, model inference threshold, or local control commands according to the operation adjustment strategy. The central server writes all received data and event information into a time-series database.

[0007] Furthermore, when the local computing module preprocesses the running parameter data, it includes: The local computing module performs noise reduction processing on the raw data collected by the sensor. The noise reduction processing includes using median filtering, bilateral filtering or exponential moving average methods to eliminate spikes or abnormal interference values. Sliding window interpolation is used to fill the gaps caused by data collection interruptions; Standardize and convert units for each physical quantity; The feature extraction includes extracting slope, rate of change, periodic fluctuations, and maximum and minimum value intervals, and constructing a time series feature tensor based on the extracted data.

[0008] Furthermore, the local computing module analyzes the parameter data based on edge AI recognition, and when identifying event information based on the analysis results, it includes: The local computing module inputs the time series feature tensor into the edge AI model deployed locally. The edge AI model is a recurrent neural network that has been centrally trained and compressed for optimization. The edge AI model infers from the input data and outputs the operating status category and its confidence score within the current time period; If the confidence level of a mutation event is higher than the preset threshold, it is identified as a mutation event; if the number of sampling points is lower than the abnormal threshold within a continuous time window, it is identified as a disconnection event; if the current physical quantity exceeds the configured range, it is identified as an over-limit event. The local computing module generates a structured event information package for each event type. The event information package includes the event type, node number, trigger time, key feature values, and model confidence.

[0009] Furthermore, when the communication module reports the identified event information to the central server in real time based on event triggering, it includes: The communication module listens to the structured event information packets generated by the local computing module and triggers the event channel communication process after the event information packets are generated. The communication module constructs an event reporting frame containing node ID, event type, timestamp, feature summary and model confidence, and encapsulates it into a signed encrypted data packet; The communication module uploads data packets to the central server via a transmission protocol channel; In the event of network interruption or congestion, the communication module caches event information in a local circular queue and periodically attempts to automatically retransmit until the central server receives an ACK response.

[0010] Furthermore, when the communication module periodically uploads the batch-compressed semi-processed data to the central storage system, it includes: The communication module collects batches of uninferenced data output by the local computing module within a preset time interval and constructs a time-stamped semi-processed data block. A differential compression algorithm is applied to the numerical physical quantities in the semi-processed data block to record the sequence of changes in adjacent values; Time-series coding is used to compress status fields or low-frequency changing data, and repeated or slowly changing values ​​are represented in variable-length format. Perform statistical aggregation processing on the data within a set sliding time window to generate mean, maximum, variance or zero crossover rate indicators and construct compressed summary blocks; The compressed data packets are organized and encapsulated according to node number and time index, and then uploaded to the central server after being cached through a data buffer. When the compression ratio is insufficient, an abnormal event occurs, or the compression process fails, the communication module can switch to the raw data upload mode.

[0011] Furthermore, when the central server receives event information from several edge computing data acquisition terminals and performs merge analysis and strategy calculation, it includes: The central server aggregates and merges event information uploaded by different nodes according to time windows and event types to construct a regional event trend map, identifying high-frequency alarm areas, common anomaly patterns, and coupling relationships with adjacent nodes. Combining power grid topology information and node operation history, a comprehensive assessment of the current situation is conducted based on a rule engine, and the sampling period, upload frequency, model threshold, and control trigger priority of the target node are calculated. Based on the strategy calculation results, the central server generates an operation adjustment strategy and assigns a unique strategy identifier and effective time limit to each target node; The operational adjustment strategy includes scheduling type, adjustment parameters, execution scope and acknowledgment requirements, which are sent to the corresponding edge computing data acquisition terminal through the strategy channel. The central server monitors the effectiveness of the strategy execution and records the feedback data from the edge computing data acquisition terminal after the strategy is executed.

[0012] Furthermore, when the edge computing data acquisition terminal dynamically adjusts the data acquisition cycle, reporting frequency, model inference threshold, or local control commands according to the operation adjustment strategy, it includes: After receiving the operation adjustment strategy, the edge computing data acquisition terminal extracts the scheduling type, target parameters and execution scope information in the operation adjustment strategy, and generates a corresponding parameter update task queue locally. The edge computing data acquisition terminal sets the policy effective time and validity period, stores the operation adjustment policy with a unique identifier, and loads the policy content in real time after the policy takes effect to trigger the acquisition behavior scheduler to update.

[0013] Furthermore, when the edge computing data acquisition terminal executes the operation adjustment strategy, it includes: The edge computing data acquisition terminal dynamically modifies the sensor sampling period and the data upload frequency of the communication module according to the operation adjustment strategy; The communication module adjusts the real-time reporting level and buffering strategy according to the event level in the operation adjustment policy, pushes high-priority data immediately and low-priority data in batches, and adaptively schedules upload tasks according to network bandwidth.

[0014] Furthermore, when the edge computing data acquisition terminal dynamically adjusts the data acquisition cycle, reporting frequency, model inference threshold, or local control commands according to the operation adjustment strategy, it also includes: The edge computing data acquisition terminal modifies the inference threshold parameters of the edge AI model according to the strategy content; When the strategy package contains local control instructions, the edge computing data acquisition terminal triggers local execution actions according to the control instructions. The local execution actions include remote closing / opening of circuit breakers, activation of alarm devices, and camera rotation. The edge computing data acquisition terminal generates an execution status feedback report after the execution of the operation adjustment strategy, and sends it back to the central server through the communication module.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: By constructing a distributed power grid operation monitoring system composed of edge computing data acquisition terminals and a central server, a closed-loop process from local sensing, intelligent identification, event-driven reporting to central aggregation analysis and strategy distribution is achieved. Compared with traditional centralized architectures, this application deploys terminals with local computing capabilities at each operating node to achieve real-time acquisition and edge AI inference analysis of key operating parameters such as voltage, current, and frequency. It can identify key events such as sudden changes, exceeding limits, and line drops within milliseconds and report them to the central server in real time through an event triggering mechanism, improving the timeliness and accuracy of alarm response. At the same time, differential compression, time-series coding, and sliding window aggregation are introduced to efficiently compress non-critical data, effectively reducing bandwidth load. The central server constructs operating strategies based on the merged event information and distributes them to edge terminals in real time as needed, thereby realizing intelligent control of node-level sampling period, reporting frequency, and AI model parameters, improving the system's real-time performance, flexibility, and scalability.

[0016] On the other hand, this application also provides a method for acquiring power grid operation data using edge computing, applied to the aforementioned power grid operation data acquisition system using edge computing, comprising: Operating parameter data are collected at each operating node of the power grid, including voltage, current, frequency, circuit breaker status, and temperature and humidity. The operating parameter data is preprocessed, including denoising, time alignment, standardization and feature extraction, and a time series feature tensor is constructed. Edge AI recognition is performed on the feature tensor to identify event information, including mutations, limit violations, and disconnections, and a structured event information package is generated. Encrypted upload frames are constructed based on event triggers and pushed to the central server in real time. At the same time, semi-processed data blocks that have not triggered events are periodically collected, compressed, and then uploaded. Receive multiple event messages, perform cross-node merging analysis, and combine power grid topology calculations to adjust operational strategies. The operation adjustment strategy is sent to the target edge terminal, and the collection cycle, reporting frequency, model inference threshold or local control command are dynamically adjusted according to the strategy. All received data and events are structured and written into the time series database.

[0017] It is understandable that the aforementioned edge computing-based power grid operation data acquisition methods and systems have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a structural block diagram of a power grid operation data acquisition system based on edge computing, provided in an embodiment of the present invention. Figure 2 A flowchart of a power grid operation data acquisition method for edge computing provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] In traditional power grid operation data acquisition systems, the centralized architecture results in data acquisition response delays exceeding second-level thresholds, failing to meet the real-time requirements of high-frequency monitoring scenarios. All raw data must be uploaded to the central node via a wide area network, causing network bandwidth resource occupancy to exceed 80%, leading to transmission congestion and packet loss in large-scale node deployment scenarios. Event identification relies on static threshold comparison methods, resulting in a false positive rate exceeding 15% for complex anomaly patterns, failing to accurately identify scenarios such as voltage waveform distortion or hidden equipment degradation. Edge nodes lack dynamic policy execution capabilities, causing operational parameter adjustments to lag behind changes in power grid status, with policy effectiveness delays reaching minutes.

[0021] For example, in a regional distribution network with 2,000 monitoring nodes, each node generates 12 physical quantity data points per second, requiring the central server to process 24,000 data streams per second. When a substation experiences a sudden voltage drop, the raw data must be transmitted to the center via three network hops, resulting in a 2.3-second delay in event identification. During this period, edge nodes continuously upload data at a fixed frequency, consuming 60% of the communication bandwidth and hindering the transmission of other critical events. When the central server judges the voltage drop event based on static thresholds, it fails to identify the accompanying harmonic distortion characteristics, leading to a misjudgment as a normal load fluctuation. Policy adjustment instructions are only issued to the edge nodes five minutes after the event occurs, missing the optimal time for voltage recovery.

[0022] If the above problems are not addressed, delays in power grid operation status perception will lead to lag in the action of protection devices, increasing the risk of equipment damage and power outages. Sustained high network bandwidth loads will cause data packet loss, resulting in the loss of historical data integrity. Static threshold mechanisms cannot adapt to the dynamically changing power grid operating environment, leading to an increased rate of missed anomaly reports and affecting fault location accuracy. Delays in policy coordination between edge nodes and the central system will weaken the power grid's adaptive adjustment capabilities, potentially triggering cascading failures in scenarios with high penetration of renewable energy.

[0023] For this, please refer to Figure 1 As shown, this application proposes an edge computing-based power grid operation data acquisition system, including: edge computing data acquisition terminals and a central server. The edge computing data acquisition terminals are deployed at each operating node of the power grid. Each edge computing data acquisition terminal includes a sensor module, a local computing module, and a communication module. The sensor module collects operating parameter data from the nodes, including voltage, current, frequency, circuit breaker status, and temperature and humidity information. The local computing module preprocesses the operating parameter data, including data cleaning, outlier removal, time alignment, and feature extraction. It then analyzes the parameter data based on edge AI recognition, identifying event information based on the analysis results. Event information includes sudden changes, exceeding limits, and line drops. The communication module, based on event triggering, reports the identified event information to the central server in real time and periodically uploads batch-compressed semi-processed data to the central storage system. Compression methods include differential compression, time-series coding, or sliding window aggregation. The central server receives event information from several edge computing data acquisition terminals, performs merge analysis and strategy calculation, generates operation adjustment strategies based on the calculation results, and distributes the operation adjustment strategies to the corresponding edge computing data acquisition terminals in real time. The edge computing data acquisition terminal dynamically adjusts the data acquisition cycle, reporting frequency, model inference threshold, or local control commands according to the operational adjustment strategy. The central server writes all received data and event information into a time-series database.

[0024] Specifically, edge computing data acquisition terminals refer to distributed hardware devices deployed at power grid operation nodes. These can be implemented using embedded systems or industrial controllers, and are used to perform real-time data acquisition and preliminary processing at the data source, reducing latency in data transmission to the central node. The central server is a central processing unit that receives and integrates data from multiple nodes. It can be implemented using a distributed cluster architecture, and is used to coordinate cross-node strategy calculations and global resource scheduling. The sensor module is a detection unit used to acquire power grid operating parameters. It can be implemented using high-precision voltage transformers, current transformers, and digital temperature and humidity sensors, ensuring the accuracy and real-time performance of the acquired data. The local computing module is a data processing unit embedded in the terminal device. It can be implemented using an edge computing framework combined with AI acceleration chips, and is used to perform data cleaning, feature extraction, and event recognition locally, reducing reliance on the central server. Preprocessing includes data cleaning, outlier removal, time alignment, and feature extraction. This can be achieved using a sliding window algorithm combined with statistical analysis methods to eliminate noise interference and extract effective features, improving the accuracy of subsequent analysis. The event information includes sudden changes, limit violations, and line drops. This can be achieved through pattern recognition of time-series data using an edge AI model, enabling rapid location of abnormal states in power grid operation and triggering real-time alarms and responses. The communication module reports based on event triggers, using MQTT or CoAP protocols. This event-driven mechanism prioritizes the transmission of critical alarm information, reducing network overhead for unnecessary data. Periodic batch compression and uploading of semi-processed data can be achieved using differential compression algorithms combined with time-series coding techniques. This reduces transmission bandwidth requirements by compressing redundant data while preserving the recoverability of the original data. Merge analysis and strategy calculation involve aggregating and correlating multi-node events. This can be achieved using a rule engine combined with graph computation algorithms to identify cross-regional anomaly patterns and generate global optimization strategies. The operational adjustment strategy dynamically adjusts the data acquisition cycle and reporting frequency, implemented through the configuration management module. This adaptively optimizes terminal behavior based on the real-time state of the power grid, forming a closed-loop control. Among them, time series database refers to a structured storage system that stores power grid operation data. Specifically, it can be implemented using InfluxDB or TimescaleDB, which supports efficient writing and time window querying to meet the storage and analysis needs of high-frequency data.

[0025] This application utilizes a collaborative architecture of edge computing and a central server to perform real-time preprocessing and event identification at the data acquisition end. Combined with event-driven reporting and periodic compressed upload mechanisms, it achieves low-latency alarms and efficient data transmission. Simultaneously, by dynamically adjusting strategies to form a closed loop of "acquisition-analysis-strategy-execution," it addresses the problems of high latency, high bandwidth pressure, and lack of adaptability inherent in traditional centralized architectures, thereby improving the real-time performance and resource utilization of the power grid monitoring system.

[0026] The working process of this application is as follows: Edge computing data acquisition terminals are deployed at various operating nodes of the power grid. They collect operational parameter data from these nodes via sensor modules, including voltage, current, frequency, circuit breaker status, and temperature and humidity information. Local computing modules preprocess the collected operational parameter data, including data cleaning, outlier removal, time alignment, and feature extraction. Based on edge AI recognition, the modules analyze the parameter data to identify events such as sudden changes, exceeding limits, and line drops. The communication module, based on an event-triggered mechanism, reports the identified event information to the central server in real time, and periodically uploads batch-compressed semi-processed data to the central storage system. The central server receives event information from multiple edge computing data acquisition terminals, performs merge analysis and strategy calculation, generates operational adjustment strategies, and distributes them to the corresponding edge computing data acquisition terminals in real time. The edge computing data acquisition terminals dynamically adjust their data acquisition cycle, reporting frequency, model inference thresholds, or local control commands according to the received operational adjustment strategies. The central server writes all received data and event information into a time-series database.

[0027] The distributed architecture enables local data preprocessing and event recognition, reducing the amount of data that needs to be transmitted. An event-triggered mechanism ensures that critical information is reported in real time, while periodic uploading of compressed, semi-processed data guarantees data integrity. The central server generates and distributes adjustment strategies based on comprehensive analysis of data from multiple terminals, forming a closed-loop control system. This ensures real-time performance while also improving the system's scalability and flexibility.

[0028] As a preferred embodiment, the solution of this application is specifically implemented as follows: Edge computing data acquisition terminals are deployed at key nodes of the power grid, such as substations and distribution rooms. The sensor module includes voltage transformers, current transformers, frequency measuring devices, circuit breaker status detectors, and temperature and humidity sensors. The local computing module uses an embedded processor and runs a real-time operating system. The communication module supports 4G / 5G wireless communication and wired Ethernet.

[0029] During data acquisition, the sensor samples voltage and current waveforms at a frequency of 100Hz, and collects frequency, circuit breaker status, and temperature and humidity data at 1Hz. The local computing module filters and denoises the acquired data, aligns the timestamps, and extracts features such as RMS value, phase angle, and harmonic content. The edge AI model uses a lightweight LSTM (Long Short-Term Memory) network to perform real-time inference on the feature sequences and identify abnormal events such as voltage drops and harmonic distortion.

[0030] When an abnormal event is detected, the communication module immediately constructs an event information packet and encrypts it before transmitting it to the central server. Simultaneously, every 10 minutes, the raw data that has not triggered an event is differentially encoded, compressed, and uploaded in batches. The central server receives event information from multiple terminals, performs correlation analysis based on the power grid topology, and generates policy instructions such as adjusting the sampling rate and updating model parameters, which are then sent to the relevant terminals.

[0031] The edge terminal dynamically adjusts parameters such as sensor sampling frequency and event recognition threshold based on received policy instructions. For example, when increased voltage fluctuations are detected, the sampling rate is increased to 200Hz and the event trigger threshold is lowered. All raw data and event information are ultimately stored in a distributed time-series database, supporting subsequent big data analysis.

[0032] Through the above-described scheme, this application achieves rapid acquisition and intelligent analysis of power grid operation data. The edge computing architecture reduces data transmission volume and alleviates network bandwidth pressure. The event-driven reporting mechanism ensures that critical information can be transmitted to the control center within milliseconds, improving system response speed. The application of edge AI models enhances the accuracy of identifying complex anomaly patterns. The dynamic strategy adjustment mechanism enables the system to optimize operating parameters in real time based on power grid conditions, enhancing the power grid's adaptive adjustment capabilities. The distributed intelligent architecture lays the foundation for the deployment of large-scale smart grid monitoring systems, effectively supporting the evolution of the power grid towards a higher level of intelligence.

[0033] In some of the solutions described above in this application, the raw data collected by the sensor module suffers from noise interference, gaps caused by intermittent collection, inconsistent units for different physical quantities, and missing feature dimensions. This results in the local computing module being unable to effectively extract discriminative time-series features, affecting the accuracy of the edge AI model in event recognition.

[0034] This application further proposes that the local computing module preprocesses the operating parameter data, including: denoising the raw data collected by the sensors, which includes using median filtering, bilateral filtering, or exponential moving average methods to eliminate spikes or abnormal interference values; filling gaps caused by data acquisition interruptions using sliding window interpolation; standardizing and converting the units of each physical quantity; and extracting features including slope, rate of change, periodic fluctuations, and maximum / minimum value intervals, and constructing a time-series feature tensor based on the extracted data.

[0035] The denoising process employs median filtering to replace anomalous spikes within a window by taking the median value; bilateral filtering to smooth high-frequency noise while preserving edge features; and exponential moving average to weight historical data and suppress random fluctuations. Sliding window interpolation fills gaps by linearly interpolating valid data before and after the window, maintaining the continuity of the time series. Standardization converts physical quantities with different dimensions into a unified range, and unit conversion converts non-standard units into preset benchmark units. Feature extraction integrates multidimensional features into a time series feature tensor by calculating the slope and rate of change between adjacent sampling points, statistically analyzing the fluctuation amplitude within the period, and recording the maximum and minimum value intervals.

[0036] Specifically, the voltage and current signals collected by the sensors are filtered by median to eliminate instantaneous spike interference, and temperature and humidity data are processed using exponential moving averages to reduce random fluctuations. When current acquisition is interrupted, linear interpolation is used to fill the gaps using a sliding window of three sampling points before and after the interruption. Voltage data is converted to kilovolts, and current data is converted to amperes, and normalized to the 0-1 range. The slope of the current within a 10-second window is extracted as a trend feature, and the maximum fluctuation of the voltage within the period is calculated to be ±5%. A time-series tensor containing voltage, current, and frequency features is constructed as input to the edge AI model. Thus, the preprocessed data eliminates noise interference and unit differences, fills data gaps, extracts multi-dimensional time-series features, and improves the accuracy and robustness of the edge AI model in event recognition.

[0037] As a preferred embodiment, the solution of this application is specifically implemented as follows: The local computing module performs noise reduction processing on the raw data collected by the sensors. The noise reduction process uses median filtering to eliminate spikes or abnormal interference values. Specifically, for each sampling point, the median of the five points before and after it is taken as the filtering result for that point, effectively removing sudden noise.

[0038] Furthermore, sliding window interpolation was used to fill the gaps caused by data collection interruptions. The sliding window size was set to 60 seconds, and linear interpolation was used to fill the missing data points within the window.

[0039] Therefore, the physical quantities are standardized and their units are converted. For example, voltage values ​​are converted to kilovolts (kV), current values ​​to amperes (A), and frequency values ​​to hertz (Hz).

[0040] Feature extraction includes extracting slope, rate of change, periodic fluctuations, and the range of maximum and minimum values. Specifically, the slope between two adjacent sampling points is calculated, the rate of change within 10 minutes is statistically analyzed, periodic fluctuation features are extracted using Fast Fourier Transform, and the range of maximum and minimum values ​​within one hour is recorded. A time series feature tensor is constructed based on the extracted data, with the tensor dimensions being [time step, number of features, number of physical quantities].

[0041] Through the above technical solutions, this application achieves effective preprocessing and feature extraction of raw data. Denoising improves data quality, interpolation ensures data continuity, standardization facilitates subsequent analysis, and feature extraction provides rich input information for anomaly identification. These preprocessing steps lay the foundation for subsequent edge AI recognition, helping to improve the accuracy and real-time performance of event recognition.

[0042] In some of the solutions described above in this application, when the local computing module performs event identification on the preprocessed running parameter data, there are problems such as a simple event discrimination logic and a lack of intelligent analysis capabilities, resulting in low accuracy in identifying complex abnormal behaviors and difficulty in effectively distinguishing event types such as mutations, exceeding limits, and disconnections.

[0043] This application further proposes a local computing module that inputs time-series feature tensors into a locally deployed edge AI model. The edge AI model is a recurrent neural network trained centrally and optimized through compression. The edge AI model infers from the input data and outputs the operational status category and its confidence score for the current time period. If the confidence score of a mutation event is higher than a preset threshold, it is identified as a mutation event. If the number of sampling points is lower than an anomaly threshold within a continuous time window, it is identified as a disconnection event. If the current physical quantity exceeds the configured range, it is identified as an out-of-limit event. The local computing module generates a structured event information package for each event type. The event information package includes the event type, node number, trigger time, key feature values, and model confidence score.

[0044] The time-series feature tensor is constructed from multi-dimensional physical quantities after standardization and unit conversion, including slope, rate of change, and periodic fluctuation features. The recurrent neural network employs a long short-term memory (LSTM) unit structure, compressed using knowledge distillation technology to adapt to the storage and computing power limitations of edge computing terminals. The model inference process uses a sliding window mechanism to predict the state of continuous time series, outputting category labels and probability distributions including mutation, normal, and out-of-limit categories. The confidence threshold is dynamically adjusted based on the historical false alarm rate to ensure the reliability of event identification. The structured event information package uses a binary encoding format, with key feature values ​​containing statistical indicators within the time window before and after the event trigger.

[0045] Specifically, the preprocessed time-series feature tensor is input into the edge AI model. The model models the temporal dependencies using a recurrent neural network, outputting the running state category and confidence score. When the confidence score exceeds a preset threshold, it is determined to be a sudden event. When the number of valid sampling points within a continuous time window is lower than a set threshold, a disconnection event detection is triggered. Real-time physical quantity values ​​are compared with preset safety ranges to generate limit-crossing judgments. The event information package encapsulates node identifiers, timestamps, and model inference basis through structured fields, where key feature values ​​include the mean, variance, and extreme value data within a five-minute window before and after the event. The compressed and optimized recurrent neural network maintains over 85% classification accuracy while reducing the model size to 30% of its original size, and controlling the inference latency to within 200 milliseconds, meeting the real-time processing requirements of edge terminals.

[0046] As a preferred embodiment, the solution of this application is specifically implemented as follows: The local computing module inputs the time-series feature tensor into the edge AI model deployed locally. The edge AI model is a centrally trained and compressed-optimized recurrent neural network. Specifically, this recurrent neural network employs a long short-term memory (LSTM) structure, containing two LSTM layers and one fully connected layer. Each LSTM layer has 64 units, and the fully connected layer has 32 neurons. The model input is a 10×5-dimensional time-series feature tensor, where 10 represents the time step and 5 represents the feature dimensions, including voltage, current, frequency, circuit breaker state, and temperature.

[0047] The edge AI model infers from the input data and outputs the operational status category and its confidence score for the current time period. The operational status categories include four types: normal, mutation, exceeding limits, and offline. The model uses a softmax activation function to output the probability distribution for each category and selects the category with the highest probability as the prediction result. The confidence score is the probability value of that category.

[0048] Furthermore, if the confidence level of a mutation-type event is higher than a preset threshold, it is identified as a mutation event. For example, if the confidence threshold for a mutation event is set to 0.8, and the probability of the model outputting a mutation category is greater than 0.8, it is determined to be a mutation event. If the number of sampling points is lower than an abnormal threshold within a continuous time window, it is identified as a disconnection event. Specifically, if the time window is set to 10 seconds and the abnormal threshold is 3 sampling points, and fewer than 3 valid data points are collected within 10 seconds, it is determined to be a disconnection event. If the current physical quantity exceeds the configured range, it is identified as an over-limit event. Among them, the voltage over-limit range is ±7% of the rated value, and the frequency over-limit range is 50±0.2Hz.

[0049] The local computing module generates structured event information packets based on event types. Each event information packet includes the event type, node number, trigger time, key feature values, and model confidence. Therefore, the data structure of the event information packet can be defined as: {"event_type":"mutation","node_id":"N001","trigger_time":"2023-05-0110:30:15.123","key_features":{"voltage":220.5,"current":100.2,"frequency":49.98},"confidence":0.92} Through the above technical solution, this application achieves intelligent analysis and event identification of power grid operation data at the edge. By deploying a lightweight recurrent neural network model, the dynamic characteristics of time-series data can be effectively captured, improving the accuracy of identifying abnormal events such as sudden changes and exceeding limits. The combination of preset thresholds and rules allows for flexible handling of different types of anomalies. Structured event information packages are generated, facilitating subsequent data transmission and analysis. This edge intelligence approach reduces the computational burden on the central server, improves the system's real-time response capability, and provides a more efficient and intelligent solution for power grid operation monitoring.

[0050] In some of the solutions described above in this application, the communication module may face data security risks and data loss due to network instability when reporting event information in real time. In particular, when the network is interrupted or congested, the event information cannot be reliably transmitted to the central server, affecting the system's real-time response capability.

[0051] This application further proposes a communication module that listens to structured event information packets generated by the local computing module and triggers the event channel communication process after the event information packets are generated. The communication module constructs an event reporting frame containing node ID, event type, timestamp, feature summary, and model confidence, and encapsulates it into a signed encrypted data packet. The communication module uploads the data packet to the central server via a transport protocol channel. In the event of network interruption or congestion, the communication module caches the event information in a local circular queue and periodically attempts to automatically retransmit it until the central server receives an ACK response.

[0052] The event channel communication process utilizes a listening mechanism to achieve real-time event-driven transmission, avoiding the latency caused by polling. Signed encrypted data packets employ an asymmetric encryption algorithm to digitally sign event information, preventing data tampering. The transmission protocol channel can choose MQTT or CoAP for low-power transmission. A local circular queue uses a first-in, first-out (FIFO) strategy to manage cached data, ensuring priority retransmission of unacknowledged events upon network recovery. The automatic retransmission mechanism uses an exponential backoff algorithm to adjust the retry interval, preventing further network congestion.

[0053] Specifically, once a structured event information packet is generated, the communication module immediately triggers the event channel communication process, encrypting and encapsulating the event reporting frame containing node identifiers and event characteristics. For example, the data content is encrypted using the AES-256 algorithm, and the data packet is signed and verified using an RSA private key. The data packet is transmitted to the central server via the MQTT protocol. If the network connection is normal, the central server returns an ACK response to confirm receipt. If a network interruption is detected, the event information is written to a local circular queue buffer, with the queue capacity set to store event data from the most recent 24 hours. The communication module attempts to re-establish the connection and send the queue header data every 30 seconds. After each failed retry, the interval is doubled until the maximum number of retries is reached or an ACK response is received. Signature verification and encrypted transmission ensure the integrity and confidentiality of event information during transmission. Buffering and adaptive retransmission mechanisms ensure data reliability under extreme network conditions.

[0054] As a preferred embodiment, the solution of this application is specifically implemented as follows: The communication module listens for structured event packets generated by the local computing module and triggers the event channel communication process after the event packet is generated. Specifically, the communication module sets up an event listener to continuously monitor the output buffer of the local computing module. When a new event packet is detected, the subsequent reporting process is immediately triggered.

[0055] The communication module constructs an event reporting frame containing a node ID, event type, timestamp, feature summary, and model confidence score, and encapsulates it into a signed encrypted data packet. The node ID is a 16-bit integer used to uniquely identify the edge terminal. The event type is encoded using a 4-bit enumeration value. The timestamp is accurate to milliseconds. The feature summary contains the maximum, minimum, and mean values ​​of key physical quantities. The model confidence score is a floating-point number between 0 and 1. The data packet is encrypted using the AES-256 algorithm and appended with a SHA-256 hash signature.

[0056] The communication module uploads data packets to the central server via a transport protocol channel. Furthermore, the MQTT protocol is used as the transport layer, with a QoS level of 1 to ensure at least-once delivery. Message topics are constructed in the format "event / [node ID] / [event type]" to facilitate message routing and processing by the central server.

[0057] In the event of network outage or congestion, the communication module caches event information in a local circular queue and periodically attempts to automatically retransmit until the central server receives an ACK response. Specifically, the circular queue capacity is set to 1000 messages, employing an LRU (Least Recently Used) eviction policy. The retransmission interval uses an exponential backoff algorithm, with an initial interval of 1 second and a maximum interval of 5 minutes. Upon receiving an ACK response from the server, the corresponding message is removed from the queue.

[0058] Through the above technical solution, this application implements an event-driven real-time reporting mechanism for edge terminals. This ensures that important event information is transmitted to the central server in a timely manner, improving the system's response speed to anomalies. Simultaneously, local caching and automatic retransmission mechanisms effectively address network instability, guaranteeing data transmission reliability. Furthermore, encryption and signature technologies enhance data transmission security, preventing information tampering or theft. Overall, this solution optimizes the efficiency and quality of power grid operation data collection, providing strong support for real-time monitoring and rapid decision-making in the power grid.

[0059] In some of the solutions described above in this application, when the communication module periodically uploads semi-processed data blocks, directly using a fixed compression algorithm to process different types of data may result in insufficient compression efficiency and an inability to effectively reduce the data volume. Furthermore, if the compression process fails and the transmission mode cannot be switched, data loss or transmission interruption may occur.

[0060] This application further proposes a method for the communication module to periodically upload batch-compressed semi-processed data to the central storage system, including: Within a preset time interval, the communication module collects batches of uninferenced data output from the local computing module and constructs time-stamped semi-processed data blocks. Differential compression algorithms are applied to numerical physical quantities in the semi-processed data blocks, recording the incremental change sequence between adjacent values. Time-series encoding is used to compress state fields or low-frequency changing data, representing repetitive or slowly varying values ​​in a variable-length format. Statistical aggregation processing is performed on the data within a set sliding time window to generate mean, maximum, variance, or zero-crossing rate indices and construct a compressed summary block. The compressed data packets are organized and encapsulated according to node number and time index, cached in a data buffer, and then uploaded to the central server. When the compression ratio is insufficient, an abnormal event occurs, or compression processing fails, the communication module can switch to raw data upload mode.

[0061] Numerical physical quantities are processed using a differential compression algorithm, recording only the changes in adjacent data to reduce redundant storage. State-type data is compressed using time-series coding, leveraging repetitive or slowly varying characteristics to reduce data length. A sliding window statistical aggregation is used to generate summary blocks, extracting key statistical indicators to replace the original data. The compressed data packets are organized using node numbers and time indices for easy retrieval by the central server. A data buffer caches the encapsulated data packets to balance network transmission pressure. In case of compression failure or anomalies, the original data upload mode is switched to ensure data integrity.

[0062] Specifically, within a preset time interval, batches of un-inferenced data are collected and time-stamped, forming semi-processed data blocks. Numerical physical quantities are processed using differential compression algorithms; for example, continuous sampled values ​​of current or voltage are recorded only as incremental change sequences between adjacent time points, reducing the storage space for repetitive values. Status fields, such as circuit breaker status or low-frequency temperature and humidity data, are converted into variable-length encoding formats through time-series coding, further compressing the data length. For data within a sliding time window, statistical aggregation calculations are performed, such as calculating the average and maximum current values ​​within a 10-second window, generating compressed summary blocks containing key indicators. The compressed data packets are encapsulated according to node number and time index; for example, data from node A between 09:00 and 09:10 is encapsulated into independent data packets, temporarily stored in a buffer, and then uploaded in batches. When the compression ratio is detected to be lower than a preset threshold or an abnormal event occurs, the communication module automatically switches to the raw data upload mode; for example, when a sudden data change causes differential compression to fail, the raw sampled values ​​are directly uploaded to avoid data loss. This process improves data transmission efficiency and reliability by adapting differential compression algorithms to data types and combining dynamic transmission mode switching.

[0063] As a preferred embodiment, the solution of this application is specifically implemented as follows: Within a preset time interval, the communication module collects batches of un-inferenced data output from the local computing module and constructs time-stamped semi-processed data blocks. Differential compression algorithms are applied to numerical physical quantities within the semi-processed data blocks, recording the incremental change sequence between adjacent values. Time-series encoding is used to compress state fields or low-frequency changing data, representing repetitive or slowly varying values ​​in a variable-length format. Statistical aggregation processing is performed on data within a set sliding time window to generate mean, maximum, variance, or zero-crossing rate indices and construct compressed summary blocks. The compressed data packets are organized and encapsulated according to node number and time index, cached in a data buffer, and then uploaded to the central server. When the compression ratio is insufficient, an abnormal event occurs, or compression processing fails, the communication module can switch to raw data upload mode.

[0064] Specifically, the communication module collects batches of un-inferenced data output from the local computing module every 10 minutes. For numerical physical quantities such as voltage and current, a differential compression algorithm is used to record the changes between adjacent sampling points. For example, the voltage data sequence [220.1, 220.3, 220.2, 220.4] is compressed and recorded as [220.1, +0.2, -0.1, +0.2]. For low-frequency change data such as circuit breaker status, a time-series encoding method is used for compression. For example, the circuit breaker status sequence [0, 0, 0, 1, 1, 1, 0, 0] can be compressed into [03, 13, 02]. Within a 5-minute sliding time window, the mean, maximum value, variance, and other statistical indicators of each physical quantity are calculated to construct a compressed summary block. Finally, the compressed data is encapsulated according to node number and timestamp, and temporarily stored in a 1MB circular buffer. Upload is triggered when the buffer reaches 80% capacity. If the compression ratio of a batch of data is lower than 30%, or a sudden event is detected, the system automatically switches to the original data upload mode.

[0065] Through the above technical solutions, this application achieves efficient compression and flexible uploading of power grid operation data. Differential compression and time-series coding reduce data transmission volume, while statistical aggregation further extracts key features. Batch caching and adaptive switching mechanisms balance real-time performance and network load. This improves data transmission efficiency, reduces storage and bandwidth overhead, while ensuring timely reporting of critical information, enhancing system reliability and flexibility.

[0066] In some of the above-mentioned schemes in this application, after the central server receives event information from multiple edge computing data acquisition terminals, the lack of a comprehensive analysis mechanism for the correlation of cross-node events makes it difficult to accurately identify regional abnormal patterns during the strategy calculation process, and makes it impossible to effectively coordinate the operation adjustment actions between multiple nodes, thus affecting the overall power grid control efficiency.

[0067] This application further proposes a central server that aggregates and merges event information uploaded by different nodes according to time windows and event types to construct a regional-level event trend map, identifying high-frequency alarm areas, common anomaly patterns, and coupling relationships with adjacent nodes. Combining power grid topology information and node operation history, a rule engine is used to comprehensively evaluate the current situation, calculating the sampling period, upload frequency, model threshold, and control trigger priority of target nodes. Based on the strategy calculation results, operational adjustment strategies are generated, and a unique strategy identifier and effective time limit are assigned to each target node. The operational adjustment strategy includes scheduling type, adjustment parameters, execution scope, and feedback requirements, and is distributed to the corresponding edge computing data acquisition terminal through the strategy channel. The execution effect of the strategy is monitored, and feedback data from the edge computing data acquisition terminal is recorded after strategy execution.

[0068] The aggregation and merging mechanism uses a sliding time window to match events along the time dimension, while event type classification uses a hash table index for fast retrieval. The regional event trend map stores the spatial relationships between nodes and event propagation paths through a graph database. The rule engine incorporates a power grid topology weight matrix and a historical event association rule library, and fuzzy logic algorithms are introduced to handle uncertainties during comprehensive evaluation. The execution adjustment strategy uses binary protocol encoding, with the strategy identifier including a combination of timestamp and node hash value. The strategy channel uses a message queue for asynchronous communication, and execution effect monitoring is achieved through heartbeat packets and data verification mechanisms.

[0069] Specifically, after receiving event information from multiple nodes, the central server first aligns the events according to a preset time window, classifying events within the same time period and storing them in an in-memory database. Next, based on the node connections in the power grid topology, a visual map containing event distribution density and propagation direction is constructed to identify high-frequency alarm areas and abnormal coupling relationships between adjacent nodes. Subsequently, the rule engine calls the power grid topology weight matrix, combined with node fault modes recorded in the historical event database, to comprehensively score the current event, calculating the sampling period, upload frequency, and model threshold parameters that each node needs to adjust. Based on the scoring results, a binary-formatted policy package is generated, each carrying a unique identifier and an effective timestamp, and distributed to the target nodes via a message queue. After a node executes the policy, the central server continuously receives heartbeat data uploaded by the node to verify the policy execution status and writes the verification results to the log database. For example, when a regional trend map shows that three adjacent nodes in a substation have experienced voltage over-limit events, the rule engine automatically increases the sampling frequency of nodes in that region to 100 times per second, while simultaneously lowering the model threshold by 5% to enhance anomaly detection sensitivity.

[0070] As a preferred embodiment, the solution of this application is specifically implemented as follows: The central server receives event information from multiple edge computing data acquisition terminals. This event information is aggregated and merged according to time windows and event types to construct a regional-level event trend map. The map identifies high-frequency alarm areas, common anomaly patterns, and the coupling relationship between adjacent nodes. For example, for voltage surge events within a 10-minute time window, event information from adjacent substations is aggregated to identify the voltage fluctuation propagation path.

[0071] Furthermore, by combining power grid topology information with node operation history, a comprehensive assessment of the current situation is performed based on a rule engine. Specifically, first, a preset power grid topology map is loaded, and the types and connections of each node are marked. Then, node operation data from the past 24 hours is read, and key indicators such as voltage stability and load factor are calculated. Finally, real-time event information is compared and analyzed with historical data to assess the degree of anomaly.

[0072] Therefore, parameters such as sampling period, upload frequency, model threshold, and control trigger priority are calculated for the target node. For example, for a substation that detects voltage fluctuations, the sampling period can be shortened from 5 seconds to 1 second, the upload frequency can be increased from 1 minute to 10 seconds, and the identification threshold for voltage surge events can be lowered.

[0073] Based on the strategy calculation results, the central server generates an operational adjustment strategy. This includes the scheduling type (e.g., "sampling period adjustment"), adjustment parameters (e.g., "1 second / time"), execution scope (e.g., "substations A, B, and C"), and feedback requirements (e.g., "feedback within 5 minutes after execution"). Each target node is assigned a unique strategy identifier and an effective time limit, such as "Strategy_001_2024080112" representing strategy number 001 that takes effect at 12:00 on August 1, 2024.

[0074] The adjustment strategy is distributed to the corresponding edge computing data acquisition terminal via an encrypted channel. The central server continuously monitors the strategy's execution effect, recording information such as the execution status and newly acquired data from the terminals, which is used to evaluate the strategy's effectiveness and make dynamic adjustments.

[0075] Through the above technical solutions, this application realizes intelligent and distributed acquisition and management of power grid operation data. The central server can perform global situational awareness based on multi-source event information and quickly generate targeted operation adjustment strategies. The dynamic adjustment mechanism improves the flexibility and targeting of data acquisition, enabling the system to better respond to sudden events and abnormal situations in power grid operation. At the same time, through hierarchical processing and intelligent decision-making, the computational burden on the central server is reduced, improving the response speed and scalability of the entire system. In addition, the structured strategy package design facilitates rapid parsing and execution by edge terminals, further enhancing the real-time performance and reliability of the system.

[0076] In some of the solutions described above in this application, if the edge computing data acquisition terminal directly executes parameter update operations after receiving the operation adjustment strategy, the parameter update tasks may become disordered due to the complexity of the strategy content or execution timing conflicts, which in turn causes the acquisition behavior scheduler to run in chaos, affecting the stability of data acquisition and the effectiveness of strategy execution.

[0077] This application further proposes that after receiving the operation adjustment strategy, the edge computing data acquisition terminal extracts the scheduling type, target parameters, and execution scope information from the operation adjustment strategy, and generates a corresponding parameter update task queue locally. The edge computing data acquisition terminal sets the strategy's effective time and validity period, stores a unique identifier for the operation adjustment strategy, and loads the strategy content in real time after the strategy takes effect to trigger the acquisition behavior scheduler to update.

[0078] The parameter update task queue generation process includes decomposing scheduling types into independent task items, prioritizing execution based on target parameter types, and mapping the execution scope to local device functional modules. Setting the policy effective time and validity period involves parsing the timestamp field in the policy package, combining it with the local clock to generate a time window, and locking the policy version within the window. Unique identifier storage uses a hash algorithm to generate a digest value for the policy content, serving as the storage index key to prevent duplicate policies from overwriting each other. The triggering logic for the collection behavior scheduler update includes calling the scheduler interface to inject new parameters when the policy effective time arrives and suspending currently unfinished task processes.

[0079] Specifically, when the edge computing data acquisition terminal receives the operation adjustment strategy issued by the central server, it first parses the scheduling type field in the strategy package, breaking down the composite strategy into multiple independent task items. For example, if the strategy includes adjusting the acquisition cycle and modifying the model inference threshold, acquisition cycle update tasks and model threshold update tasks are generated respectively. The task items are sorted according to a preset priority, with control-type instructions having higher priority than parameter-type instructions. Subsequently, based on the effective timestamp and effective period fields in the strategy package, a time window is generated locally to ensure that the strategy takes effect within the specified time period. To avoid strategy conflicts, a hash algorithm is used to generate a unique identifier for the strategy content, and the uniqueness of the identifier is verified during storage to prevent duplicate strategies from overwriting each other. When the system clock reaches the strategy's effective time, the scheduler interface is called, new parameters are injected into the acquisition behavior scheduler, and currently unfinished task processes are suspended until the new strategy is loaded and execution resumes. Thus, parameter update tasks are executed sequentially, the strategy version remains stable within its validity period, and the update process of the acquisition behavior scheduler is orderly and controllable.

[0080] As a preferred embodiment, the solution of this application is specifically implemented as follows: After receiving the operation adjustment strategy, the edge computing data acquisition terminal extracts the scheduling type, target parameters, and execution scope information, and generates a corresponding parameter update task queue locally. Specifically, the edge terminal parses the received strategy data packet, extracting the scheduling type field (such as "sampling period adjustment", "reporting frequency adjustment", etc.), target parameter values ​​(such as new sampling interval, reporting time interval, etc.), and execution scope information (such as specific sensor ID, time period, etc.). This information is organized into a task queue, with each task item containing attributes such as scheduling type, target value, and execution scope.

[0081] Furthermore, the edge computing data acquisition terminal sets the policy's effective time and validity period, stores a unique identifier for each policy adjustment, and loads the policy content in real time after the policy takes effect, triggering an update by the acquisition behavior scheduler. For example, the terminal assigns a unique policy ID to each received policy and records its effective time and validity period. The policies are stored in a local database, using the policy ID as an index. When the policy's effective time arrives, the terminal loads the corresponding policy content from the database and triggers the acquisition behavior scheduler to update. The scheduler dynamically adjusts the operating parameters and behaviors of relevant modules based on the loaded policy content.

[0082] Therefore, edge computing data acquisition terminals can flexibly adjust their data acquisition and processing behavior according to the operation adjustment strategies issued by the center, thereby achieving intelligent and adaptive monitoring of the power grid operation status.

[0083] Through the above technical solution, this application achieves efficient parsing and execution of operational adjustment strategies by edge terminals. Edge terminals can accurately extract key information from the strategies and transform it into an executable task queue. By setting the strategy's effective time and validity period, and employing a unique identifier storage mechanism, the timeliness and traceability of strategy execution are ensured. Real-time loading of strategy content and triggering updates to the data acquisition scheduler enable edge terminals to respond promptly to adjustment instructions from the center, dynamically optimizing their data acquisition and processing behavior. This improves the flexibility and adaptability of the power grid operation data acquisition system, allowing it to quickly adjust acquisition strategies according to actual operating conditions and needs, thereby better supporting the intelligent operation and management of the power grid.

[0084] In some of the solutions described above in this application, the edge computing data acquisition terminal can dynamically adjust the data acquisition cycle, reporting frequency, model inference threshold, or local control commands according to the operation adjustment strategy. However, during the execution process, the communication module lacks differentiated processing for the reporting real-time level and buffering strategy of different priority events, and does not consider the impact of dynamic changes in network bandwidth on the upload task. This results in high-priority events being delayed in reporting due to network congestion, while low-priority data occupies bandwidth resources, affecting the overall response efficiency of the system.

[0085] This application further proposes that the communication module adjusts the real-time level and buffering strategy of the reported data according to the event level in the operation adjustment strategy, pushes high-priority data immediately and pushes low-priority data in batches, and adaptively schedules the upload task according to the network bandwidth.

[0086] The adjustments to real-time performance levels and buffering strategies include allocating independent transmission channels or queues for different event levels, employing a preemptive transmission mechanism for high-priority events, and triggering batch transmission of low-priority data after it accumulates in the buffer to a set threshold. Adaptive scheduling of upload tasks involves real-time monitoring of network bandwidth fluctuations and dynamically adjusting the transmission rate or fragment size. When network congestion occurs, the data fragment size of low-priority tasks is reduced to prioritize bandwidth utilization of high-priority channels. For example, when network bandwidth falls below a preset threshold, the communication module adjusts the fragment size of low-priority data from the default 1MB to 256KB, while simultaneously suspending the upload of non-urgent tasks.

[0087] Specifically, when an event is defined as high priority in the operation adjustment strategy, the communication module immediately interrupts the current low-priority data transmission task, encapsulates the high-priority event information into an independent data packet, and pushes it to the central server through a dedicated channel. Simultaneously, the low-priority data is temporarily stored in a circular buffer and uploaded in batches according to a set time window after the high-priority task is completed or network bandwidth is restored. When the network bandwidth monitoring module detects that the available bandwidth has dropped below 30% of the original bandwidth, the communication module automatically triggers a transmission rate adjustment algorithm, limiting the upload rate of low-priority data to 10% of the bandwidth, reserving the remaining bandwidth for the high-priority event channel. Therefore, when high-risk events such as circuit breaker tripping occur at power grid nodes, relevant alarm information can be reported within 50ms, while low-frequency data such as environmental temperature and humidity are transmitted in compressed fragments with delayed transmission when bandwidth is limited, ensuring timely response to critical events while avoiding excessive consumption of network resources.

[0088] As a preferred embodiment, the solution of this application is implemented as follows: After receiving the operation adjustment strategy issued by the central server, the edge computing data acquisition terminal parses the preset sampling period parameters and reporting priority rules in the strategy. When the strategy requires improving the real-time performance of data acquisition, the sampling period of the sensor module is adjusted from the default 5 seconds / time to 1 second / time, and the upload frequency of the communication module is adjusted from compressed upload every 10 minutes to incremental upload every 30 seconds. For high-priority event data, an independent transmission channel is established using the MQTT protocol to achieve instant push, and the data packet is encapsulated and sent within 200 milliseconds after generation. Low-priority data is cached in a local circular buffer, and is batch compressed in batches of one hour before being transmitted via the FTP protocol. The communication module monitors the network bandwidth utilization in real time. When the bandwidth usage exceeds 70%, the data packet fragmentation mechanism is automatically enabled to limit the amount of data uploaded at one time to within 500KB, and flow control is achieved by dynamically adjusting the TCP window size.

[0089] Through the above technical solutions, this application effectively solves the problems of high data transmission latency and low bandwidth resource utilization in traditional centralized architectures. By dynamically adjusting sampling and reporting strategies, invalid transmission of non-critical data is reduced, prioritizing the real-time reporting needs of emergencies. The adaptive scheduling mechanism based on network status avoids channel congestion and improves data transmission success rate. The hierarchical buffering strategy optimizes the allocation of storage resources for edge terminals while ensuring the real-time performance of critical data, enabling the stable operation of the power grid operation monitoring system in complex network environments.

[0090] In some of the solutions described above in this application, the lack of synchronous adjustment of the inference threshold of the edge AI model during the dynamic adjustment of the data collection cycle and reporting frequency leads to insufficient synergy between event recognition sensitivity and reporting strategies. Furthermore, the absence of a triggering mechanism for local execution actions prevents direct activation of terminal devices such as circuit breakers and alarm devices after strategy adjustments. In addition, the lack of feedback verification of strategy execution effects affects the real-time performance and accuracy of subsequent strategy optimization.

[0091] This application further proposes that when the edge computing data acquisition terminal dynamically adjusts the data acquisition cycle, reporting frequency, model inference threshold, or local control commands according to the operation adjustment strategy, it also includes: the edge computing data acquisition terminal modifying the inference threshold parameters of the edge AI model according to the strategy content. When the strategy package contains local control commands, the edge computing data acquisition terminal triggers local execution actions according to the control commands. These local execution actions include remote closing / opening of circuit breakers, activation of alarm devices, and camera reversal. After the operation adjustment strategy is completed, the edge computing data acquisition terminal generates an execution status feedback report and transmits it back to the central server via the communication module.

[0092] Modifying the inference threshold parameters of the edge AI model can be achieved by adjusting the confidence score threshold or the anomaly detection sensitivity parameters. For example, the confidence threshold for mutation events can be lowered from 0.85 to 0.75 to improve recognition sensitivity. When a local execution action is triggered, the control command may include the target status code of the circuit breaker operation, the activation duration of the alarm device, and the camera turning angle parameters. The execution status feedback report includes the policy identifier, execution timestamp, actual adjusted parameter values, and device response status code.

[0093] Specifically, when the adjustment strategy includes a model inference threshold adjustment instruction, the edge computing data acquisition terminal parses the strategy parameters and updates the threshold configuration of the local AI model. For example, during peak grid load periods, the threshold for determining over-limit events is lowered, allowing the model to identify potential over-limit risks earlier. If the strategy includes a circuit breaker tripping instruction, the terminal sends a tripping signal to the circuit breaker through its internal control interface, simultaneously activating the alarm device and recording the action time. After the strategy is executed, the terminal summarizes the model threshold change records, device response status, and execution time to generate a feedback report, which is then encapsulated into an encrypted data packet and uploaded to the central server via the communication module. The central server parses the execution status code in the feedback report and compares it with the actual parameters to determine the expected effect of the strategy. If the circuit breaker fails to respond according to the instruction or the false alarm rate increases after the model threshold adjustment, the strategy correction process is triggered to regenerate the adjustment parameters.

[0094] As a preferred embodiment, the solution of this application is implemented as follows: When the operation adjustment strategy includes local control instructions, the edge computing data acquisition terminal parses the action type and target device identifier in the instructions. For example, if the strategy instruction requires a tripping operation to be performed on circuit breaker numbered N-152, the local computing module converts the instruction into a device drive signal and sends it to the target circuit breaker controller via the RS-485 interface, triggering the execution of the tripping action. Simultaneously, the local computing module calls the camera control interface and adjusts the camera's pan-tilt-zoom (PTZ) rotation according to the preset angle parameters in the instruction, so that the monitoring image focuses on the circuit breaker operation area. After the action is executed, the local computing module collects the circuit breaker status signal and camera position feedback data, generates a feedback report containing the execution result code, timestamp, and device status snapshot, and encapsulates it into a JSON format data packet via the MQTT protocol and sends it back to the central server.

[0095] Through the above technical solutions, this application achieves rapid response and autonomous execution capabilities of power grid edge nodes to control commands, avoiding the command delay problem in traditional centralized processing modes. By dynamically adjusting the model inference threshold and linking with local devices, the accuracy and efficiency of abnormal event handling are improved. The execution status feedback mechanism ensures closed-loop verification of the strategy execution process, providing reliable real-time data support for the central server's global decision-making.

[0096] The above embodiments construct a distributed power grid operation monitoring system composed of edge computing data acquisition terminals and a central server, realizing a closed-loop process from local perception, intelligent identification, event-driven reporting to central aggregation analysis and strategy distribution. Compared with traditional centralized architectures, this application deploys terminals with local computing capabilities at each operating node to achieve real-time acquisition and edge AI inference analysis of key operating parameters such as voltage, current, and frequency. It can identify key events such as sudden changes, exceeding limits, and line drops within milliseconds and report them to the central server in real time through an event triggering mechanism, improving the timeliness and accuracy of alarm response. At the same time, differential compression, time-series coding, and sliding window aggregation are introduced to efficiently compress non-critical data, effectively reducing bandwidth load. The central server constructs operating strategies based on the aggregated event information and distributes them to edge terminals in real time as needed, thereby realizing intelligent control of node-level sampling period, reporting frequency, and AI model parameters, improving the system's real-time performance, flexibility, and scalability.

[0097] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides a power grid operation data acquisition method based on edge computing, applied to the aforementioned power grid operation data acquisition system based on edge computing, including: S100: Collects operating parameter data at each operating node of the power grid. The operating parameter data includes voltage, current, frequency, circuit breaker status, and temperature and humidity.

[0098] S200: Preprocess the runtime parameter data, including denoising, time alignment, standardization and feature extraction, and construct a time series feature tensor.

[0099] S300: Performs edge AI recognition on feature tensors, identifies event information including mutations, limit violations, and disconnections, and generates structured event information packages.

[0100] S400: Constructs encrypted upload frames based on event triggers and pushes them to the central server in real time. At the same time, it periodically collects semi-processed data blocks that have not triggered events, compresses them, and then uploads them.

[0101] S500: Receives multiple event messages, performs cross-node merging analysis, and combines this with power grid topology calculations to implement operational adjustment strategies. The operational adjustment strategies are then distributed to the target edge terminals, dynamically adjusting the data acquisition cycle, reporting frequency, model inference thresholds, or local control commands based on the strategies.

[0102] S600: Structure all received data and events and write them into the time series database.

[0103] Specifically, operational parameter data includes voltage, current, frequency, circuit breaker status, and temperature and humidity. Preprocessing includes denoising, time alignment, standardization, and feature extraction. Edge AI recognition classifies events by constructing time-series feature tensors; event information includes abrupt changes, exceeding limits, and line drops. Encrypted upload frames are encapsulated using signature encryption and pushed via a transmission protocol channel when an event is triggered. Semi-processed data blocks are compressed using differential compression, time-series coding, or sliding window aggregation. Merge analysis incorporates power grid topology calculation strategies, including adjustments to the acquisition cycle, reporting frequency, model thresholds, and control commands. The time-series database stores data in a structured format.

[0104] Specifically, sensor modules deployed at each operating node of the power grid collect operating parameter data such as voltage, current, and circuit breaker status. This data is transmitted to a local computing module for preprocessing. During preprocessing, the raw data undergoes median filtering to eliminate spike noise, sliding window interpolation to fill data gaps, and standardization to unify dimensions. The feature extraction module calculates slope, rate of change, and periodic fluctuation indicators to construct a multi-dimensional time-series feature tensor. An edge AI model infers from the feature tensor, outputting the operating status category and confidence score. When the confidence score exceeds a preset threshold, a structured event information packet is generated. The event information packet is encapsulated as an encrypted data frame and uploaded to the central server in real time via an event channel. Semi-processed data that has not triggered events is compressed in batches at preset intervals. A differential algorithm is used to record numerical increments, time-series encoding is used to compress low-frequency data, and sliding window aggregation generates statistical indicators. The compressed data blocks are then uploaded via a buffer. The central server performs time window aggregation on multi-node events and, combined with common anomaly patterns from power grid topology analysis, generates an operating strategy that includes adjustments to the acquisition cycle and optimization of model thresholds. The strategy is distributed to edge terminals via a strategy channel. The terminals dynamically modify sensor sampling frequencies, adjust communication module reporting priorities, update AI model inference parameters, and trigger circuit breaker opening and closing controls based on the strategy. All processed data and event information are written to a time-series database using timestamp indexes, forming a complete data loop. For example, when a voltage over-limit event is detected, the central server calculates and generates a strategy to shorten the acquisition cycle. The edge terminal adjusts the sampling interval from 1 second to 0.5 seconds and simultaneously increases the event reporting priority to ensure real-time uploading of abnormal data. This method reduces data transmission volume through edge preprocessing, improves real-time performance using an event-driven mechanism, and achieves closed-loop control through strategy feedback, effectively solving the problems of high latency, large bandwidth consumption, and insufficient intelligence inherent in traditional solutions.

[0105] As a preferred embodiment, the solution of this application is implemented as follows: An edge computing terminal is deployed at the 110kV main transformer node of the power grid substation. Three-phase current data is collected through Hall sensors, the open / closed status of the circuit breaker is read using an RS485 bus, and environmental data of the equipment is obtained using temperature and humidity sensors. The collected raw data is preprocessed. First, high-frequency interference in the current signal is eliminated by an exponential moving average algorithm. Then, the sampling timestamps of different sensors are aligned using linear interpolation, and the current values ​​are converted into per-unit values ​​to generate a normalized tensor. The feature tensor is input into a pre-set LSTM model for inference. When a current mutation confidence exceeds 0.92, a mutation event packet is generated. The event packet is encrypted with AES and pushed to the central server in real time via the MQTT protocol. At the same time, the current waveform data that has not triggered an event is differentially compressed every 5 minutes to form a compressed block containing time series increments and uploaded to the central storage. After receiving over-limit events from multiple nodes, the central server identifies three nodes in the same power supply area that simultaneously have voltage over-limit events based on the power grid topology and generates an adjustment strategy to shorten the sampling period from 1 second to 200 milliseconds. After receiving the policy instruction, the target terminal immediately modifies the sampling frequency of the current acquisition module and adjusts the model inference threshold from 0.9 to 0.85 to improve event sensitivity.

[0106] Through the above technical solutions, this application achieves real-time intelligent processing and closed-loop control of power grid operation data. By preprocessing data and identifying events at the edge, the computational load on the central node is effectively reduced, shortening the response time to sudden events to the millisecond level. The transmission mechanism, combining event-driven and periodic compression, reduces network bandwidth usage by approximately 60%. The dynamic strategy adjustment mechanism enables the system to adapt to changes in power grid operating conditions, rapidly improving monitoring accuracy in the early stages of faults and preventing the escalation of abnormal situations. Simultaneously, a complete technical closed loop of data acquisition, intelligent analysis, and strategy feedback execution is constructed, enhancing the autonomous decision-making capability and operational reliability of the power grid operation monitoring system.

[0107] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0108] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0109] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0110] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A power grid operation data acquisition system based on edge computing, characterized in that, include: Edge computing data acquisition terminals and a central server, wherein the edge computing data acquisition terminals are deployed at each operating node of the power grid; The edge computing data acquisition terminal includes a sensor module, a local computing module, and a communication module. The sensor module is used to collect the operating parameter data of the node, including voltage, current, frequency, circuit breaker status, and temperature and humidity information. The local computing module preprocesses the running parameter data, including data cleaning, outlier removal, time alignment, and feature extraction. It also analyzes the parameter data based on edge AI recognition and identifies event information based on the analysis results. The event information includes mutations, exceeding limits, and disconnections. The communication module, based on event triggering, reports the identified event information to the central server in real time, and periodically uploads the batch-compressed semi-processed data to the central storage system. The compression methods include differential compression, time-series coding, or sliding window aggregation. The central server receives event information from several edge computing data acquisition terminals, performs merge analysis and strategy calculation, generates an operation adjustment strategy based on the calculation results, and sends the operation adjustment strategy to the corresponding edge computing data acquisition terminal in real time. The edge computing data acquisition terminal dynamically adjusts the data acquisition cycle, reporting frequency, model inference threshold, or local control commands according to the operation adjustment strategy. The central server writes all received data and event information into a time-series database.

2. The edge computing power grid operation data acquisition system according to claim 1, characterized in that, When the local computing module preprocesses the runtime parameter data, it includes: The local computing module performs noise reduction processing on the raw data collected by the sensor. The noise reduction processing includes using median filtering, bilateral filtering or exponential moving average methods to eliminate spikes or abnormal interference values. Sliding window interpolation is used to fill the gaps caused by data collection interruptions; Standardize and convert units for each physical quantity; The feature extraction includes extracting slope, rate of change, periodic fluctuations, and maximum and minimum value intervals, and constructing a time series feature tensor based on the extracted data.

3. The edge computing power grid operation data acquisition system according to claim 2, characterized in that, The local computing module analyzes the parameter data based on edge AI recognition, and when identifying event information based on the analysis results, it includes: The local computing module inputs the time series feature tensor into the edge AI model deployed locally. The edge AI model is a recurrent neural network that has been centrally trained and compressed for optimization. The edge AI model infers from the input data and outputs the operating status category and its confidence score within the current time period; If the confidence level of a mutation event is higher than the preset threshold, it is identified as a mutation event; if the number of sampling points is lower than the abnormal threshold within a continuous time window, it is identified as a disconnection event; if the current physical quantity exceeds the configured range, it is identified as an over-limit event. The local computing module generates a structured event information package for each event type. The event information package includes the event type, node number, trigger time, key feature values, and model confidence.

4. The edge computing power grid operation data acquisition system according to claim 3, characterized in that, When the communication module, based on event triggering, reports the identified event information to the central server in real time, it includes: The communication module listens to the structured event information packets generated by the local computing module and triggers the event channel communication process after the event information packets are generated. The communication module constructs an event reporting frame containing node ID, event type, timestamp, feature summary and model confidence, and encapsulates it into a signed encrypted data packet; The communication module uploads data packets to the central server via a transmission protocol channel; In the event of network interruption or congestion, the communication module caches event information in a local circular queue and periodically attempts to automatically retransmit until the central server receives an ACK response.

5. The edge computing power grid operation data acquisition system according to claim 1, characterized in that, When the communication module periodically uploads the batch-compressed semi-processed data to the central storage system, it includes: The communication module collects batches of uninferenced data output by the local computing module within a preset time interval and constructs a time-stamped semi-processed data block. A differential compression algorithm is applied to the numerical physical quantities in the semi-processed data block to record the sequence of changes in adjacent values; Time-series coding is used to compress status fields or low-frequency changing data, and repeated or slowly changing values ​​are represented in a variable-length format. Perform statistical aggregation processing on the data within a set sliding time window to generate mean, maximum, variance or zero crossover rate indicators and construct compressed summary blocks; The compressed data packets are organized and encapsulated according to node number and time index, and then uploaded to the central server after being cached through a data buffer. When the compression ratio is insufficient, an abnormal event occurs, or the compression process fails, the communication module can switch to the raw data upload mode.

6. The edge computing power grid operation data acquisition system according to claim 4, characterized in that, When the central server receives event information from several edge computing data acquisition terminals and performs merge analysis and strategy calculation, it includes: The central server aggregates and merges event information uploaded by different nodes according to time windows and event types to construct a regional event trend map, identifying high-frequency alarm areas, common anomaly patterns, and coupling relationships with adjacent nodes. Combining power grid topology information and node operation history, a comprehensive assessment of the current situation is conducted based on a rule engine, and the sampling period, upload frequency, model threshold, and control trigger priority of the target node are calculated. Based on the strategy calculation results, the central server generates an operation adjustment strategy and assigns a unique strategy identifier and effective time limit to each target node; The operational adjustment strategy includes scheduling type, adjustment parameters, execution scope and acknowledgment requirements, which are sent to the corresponding edge computing data acquisition terminal through the strategy channel. The central server monitors the effectiveness of the strategy execution and records the feedback data from the edge computing data acquisition terminal after the strategy is executed.

7. The edge computing power grid operation data acquisition system according to claim 6, characterized in that, When the edge computing data acquisition terminal dynamically adjusts the data acquisition cycle, reporting frequency, model inference threshold, or local control commands according to the operation adjustment strategy, it includes: After receiving the operation adjustment strategy, the edge computing data acquisition terminal extracts the scheduling type, target parameters and execution scope information in the operation adjustment strategy, and generates a corresponding parameter update task queue locally. The edge computing data acquisition terminal sets the policy effective time and validity period, stores the operation adjustment policy with a unique identifier, and loads the policy content in real time after the policy takes effect to trigger the acquisition behavior scheduler to update.

8. The edge computing power grid operation data acquisition system according to claim 7, characterized in that, When the edge computing data acquisition terminal executes the operation adjustment strategy, it includes: The edge computing data acquisition terminal dynamically modifies the sensor sampling period and the data upload frequency of the communication module according to the operation adjustment strategy; The communication module adjusts the real-time reporting level and buffering strategy according to the event level in the operation adjustment policy, pushes high-priority data immediately and low-priority data in batches, and adaptively schedules upload tasks according to network bandwidth.

9. The edge computing power grid operation data acquisition system according to claim 8, characterized in that, When the edge computing data acquisition terminal dynamically adjusts the data acquisition cycle, reporting frequency, model inference threshold, or local control commands according to the operation adjustment strategy, it also includes: The edge computing data acquisition terminal modifies the inference threshold parameters of the edge AI model according to the strategy content; When the strategy package contains local control instructions, the edge computing data acquisition terminal triggers local execution actions according to the control instructions. The local execution actions include remote closing / opening of circuit breakers, activation of alarm devices, and camera rotation. The edge computing data acquisition terminal generates an execution status feedback report after the execution of the operation adjustment strategy, and sends it back to the central server through the communication module.

10. A method for acquiring power grid operation data using edge computing, applied to the power grid operation data acquisition system using edge computing as described in any one of claims 1-9, characterized in that, include: Operating parameter data are collected at each operating node of the power grid, including voltage, current, frequency, circuit breaker status, and temperature and humidity. The operating parameter data is preprocessed, including denoising, time alignment, standardization and feature extraction, and a time series feature tensor is constructed. Edge AI recognition is performed on the feature tensor to identify event information, including mutations, limit violations, and disconnections, and a structured event information package is generated. Encrypted upload frames are constructed based on event triggers and pushed to the central server in real time. At the same time, semi-processed data blocks that have not triggered events are periodically collected, compressed, and then uploaded. Receive multiple event messages, perform cross-node merging analysis, and combine power grid topology calculations to adjust operational strategies. The operational adjustment strategy is sent to the target edge terminal, and the collection cycle, reporting frequency, model inference threshold or local control command are dynamically adjusted according to the strategy. All received data and events are structured and written into the time series database.

Images