A beidou power data intelligent acquisition and dynamic communication scheduling method and system

Abstract

The application provides a Beidou power data intelligent acquisition and dynamic communication scheduling method and system, the method comprising: a master station constructs a database and trains a data priority classification model, a channel quality prediction model and an intelligent compression model; a sub-station acquires power data through an edge computing unit, constructs a skyline candidate set for data screening, and uses the data priority classification model for intelligent classification and marking; the reliability of the data stream is evaluated using a Gaussian mixture model, a compression strategy is selected according to the reliability score, the data type and the priority, and the data frame is packaged; the transmission strategy is determined in combination with the channel quality prediction result and the context-aware intelligent switching protocol, and the data frame is sent; the master station receives the data frame for integrity check, decompression and reorganization, and feeds back the communication state for model updating; the operation state is monitored, early warning is performed based on an abnormality detection model, and a self-healing strategy is triggered. The application improves the Beidou communication resource utilization rate and system reliability.

Description

Technical Field

[0001] This invention relates to the field of power data acquisition and transmission technology, and in particular to a BeiDou power data intelligent acquisition and dynamic communication scheduling method and system. Background Technology

[0002] Data collection and transmission in remote renewable energy plants and small hydropower stations within the power system are crucial components of modern power grid management. Due to geographical limitations, conventional communication methods are insufficient in these remote areas, necessitating reliance on satellite communication systems for data transmission. The BeiDou Navigation Satellite System, my country's independently developed global satellite navigation system, provides not only positioning and navigation functions but also short message communication services, offering a reliable data transmission channel for power facilities in remote areas.

[0003] Currently, data collection and transmission for remote power facilities primarily employs fixed-period collection strategies and preset communication parameters. For example, some systems collect power data at fixed 5-minute or 15-minute intervals and transmit it according to preset priority rules. Another solution uses a multi-level caching mechanism to store data of varying importance in a hierarchical manner, prioritizing the transmission of higher-priority data when communication resources are limited.

[0004] Currently, a rule-based adaptive acquisition and transmission mechanism is employed. This technology dynamically adjusts the acquisition frequency based on preset conditions (such as voltage deviation and load changes), and combines the transmission characteristics of BeiDou short messages to improve transmission efficiency through frame packing and simple data compression. However, this type of technology still has significant limitations when facing the complex and ever-changing operating conditions of power systems and the BeiDou communication environment.

[0005] The existing technology has the following main technical problems: First, rule-based data collection strategies are difficult to cope with the dynamics and uncertainties of power systems and cannot accurately determine the importance of data; second, they lack the ability to predict the quality of BeiDou communication channels and cannot avoid adverse communication conditions in advance. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for intelligent acquisition and dynamic communication scheduling of BeiDou power data, aiming to solve the technical problems in the prior art, such as the difficulty of rule-based acquisition strategies in dealing with the dynamics and uncertainties of power systems, the lack of predictive ability for the quality of BeiDou communication channels, the inability of data compression methods to be dynamically adjusted according to data type and importance, and the lack of intelligent self-healing mechanisms.

[0007] To achieve the above objectives, this invention provides a method for intelligent acquisition and dynamic communication scheduling of BeiDou power data, comprising the following steps:

[0008] The main station performs data cleaning and feature engineering on the acquired historical power data, Beidou communication data, environmental and operating condition data, establishes basic parameter configurations, and writes them into the database.

[0009] The main station trains a data priority classification model, a channel quality prediction model, and an intelligent compression model based on the database.

[0010] The substation collects power data through the edge computing unit, constructs a skyline candidate set by combining the preset EPUS algorithm for data filtering, and uses the data priority classification model from the main station to intelligently classify and label the filtered power data, outputting a data stream with priority labels.

[0011] The substation uses a pre-defined Gaussian mixture model classification framework to perform a reliability assessment on the data stream, obtains a reliability score, and selects a compression strategy based on the reliability score, the data type identifier corresponding to the data stream, and the priority label, and packages the data stream into data frames according to the compression strategy.

[0012] The substation determines a transmission strategy based on the data frame and channel quality prediction results, combined with a preset context-aware intelligent handover protocol, and sends the data frame to the master station according to the transmission strategy. The transmission strategy is used to indicate the transmission timing selection, multi-link intelligent handover, and parameter adaptive adjustment. The channel quality prediction results are obtained by the channel quality prediction model from the master station after predicting the current BeiDou signal strength.

[0013] The master station receives data frames from the substation, performs integrity verification, data decompression and reassembly on the data frames, and feeds back the communication status to the substation, so that the substation updates the database, the data priority classification model, the channel quality prediction model and the intelligent compression model according to the communication status;

[0014] The main station monitors the operating status of itself, the substations, and the edge computing units, provides early warnings of abnormal states based on a pre-trained anomaly detection model, and triggers corresponding self-healing strategies. The abnormal states include data anomalies, communication anomalies, and device anomalies.

[0015] Optionally, the substation collects power data through an edge computing unit, constructs a skyline candidate set using a preset EPUS algorithm for data filtering, and intelligently classifies and labels the filtered power data using the data priority classification model from the main station, outputting a data stream with priority labels, specifically including:

[0016] The edge computing unit detects the electricity meter and monitoring and control device connected via RS-485 and Ethernet interfaces to collect electricity data and records the metadata generated during the collection process to form a raw data stream. The metadata includes the timestamp of data generation, device identifier, associated operating condition information and data type identifier. The electricity data includes voltage, current, power and electrical energy.

[0017] The power data is regarded as data points in a multi-dimensional space, and an importance index space is constructed. The dimensional indicators of the importance index space include the data change rate, the degree of deviation from the rated value, and the predefined importance of the data type.

[0018] Using the preset EPUS algorithm and the importance index space, the importance index of each data point is calculated, and the data point is compared with the data in the current skyline candidate set based on the importance index.

[0019] When the importance index of the data point exceeds that of the data in the current skyline candidate set, the data point is added to the skyline candidate set and updated.

[0020] The data that has been filtered by EPUS and classified into the skyline candidate set are transformed into feature vectors;

[0021] The feature vector is classified using the data priority classification model to generate priority labels and classification confidence scores;

[0022] Meta-information tags are added to the categorized data to obtain the data stream. The meta-information tags include the priority label, the data type identifier, the timestamp, and the classification confidence score.

[0023] Optionally, the substation uses a preset Gaussian mixture model classification framework to perform a reliability assessment on the data stream, obtains a reliability score, and selects a compression strategy based on the reliability score, the data type identifier corresponding to the data stream, and the priority label. The data stream is then packaged into data frames according to the compression strategy, specifically including:

[0024] A feature space representation is established for different types of power data. Numerical features and contextual features are extracted based on the feature space representation. A preset GMM model is trained based on the numerical features and the contextual features to obtain a Gaussian mixture model.

[0025] Using the Gaussian mixture model, the probability density value of each data stream belonging to the classified priority and the probability density value of each other priority category are calculated. The reliability index evaluation value is calculated by comparing the probability density values, and the reliability score is composed of the index evaluation value. The reliability index includes intra-class reliability index, inter-class difference index and anomaly index.

[0026] When the priority label of the data stream is detected to be within a preset low priority range, but the reliability score is higher than a preset high reliability threshold, the priority corresponding to the priority label is increased.

[0027] When the priority label of the data stream is detected to be within a preset high priority range, but the reliability score is lower than a preset low reliability threshold, the context information corresponding to the data stream is appended to the data stream.

[0028] Based on at least one of the priority label, the reliability score, and the data type identifier, a compression strategy is selected, and the data stream is compressed according to the compression strategy to obtain compressed data;

[0029] Based on the frame structure, data size, and priority tag of BeiDou short messages, the compressed data is packaged into data frames.

[0030] Optionally, the compression strategy is used to indicate the compression ratio and / or compression method of the data stream; the step of selecting a compression strategy based on at least one of the priority label, the reliability score, and the data type identifier, and compressing the data stream according to the compression strategy to obtain compressed data, specifically includes:

[0031] Based on the priority label and the reliability score, a compression strategy is determined. Specifically, for a priority label of first priority, when the reliability score is greater than a preset first reliability threshold, a specified first compression ratio is used to compress the data stream; or when the reliability score is not greater than the first reliability threshold, the data stream is combined with simplified data determined from the data stream to generate compressed data. For a priority label of second priority, a specified second compression ratio is used to compress the data stream. The first priority is higher than the second priority, and the first compression ratio is lower than the second compression ratio.

[0032] For power data with time-series correlation, the data difference is differentially encoded by combining the current data change trend and the preset load forecasting model to obtain compressed data. The data difference is determined by the difference between the power data at the current timestamp and the power data at the previous timestamp.

[0033] The frame structure, data size, and priority tag based on BeiDou short messages are used to package the compressed data into data frames, including:

[0034] For compressed data with the priority label set to the first priority, each compressed data is packaged into a separate data frame.

[0035] For compressed data with the priority tag set to the second priority, the compressed data is combined and packaged according to the number of data points in the frame structure to form a data frame;

[0036] For compressed data whose priority label is lower than the second priority, multiple compressed data are concatenated into the same data frame.

[0037] Optionally, the substation determines a transmission strategy based on the data frame and channel quality prediction results, combined with a preset context-aware intelligent handover protocol, specifically including:

[0038] The BeiDou signal strength within a previously specified time period is input into the channel quality prediction model to predict the channel quality changes within a subsequent specified time period, and the channel quality prediction result is output. The channel quality prediction result includes the channel quality prediction level, confidence interval, and trend of change.

[0039] Combining a preset context-aware intelligent handover protocol, the strategy score of the candidate transmission strategy is obtained based on the channel quality score at the current moment, the channel quality prediction level for the specified time period thereafter, the data priority, the resource consumption of the candidate transmission strategy, the current channel quality weight, the predicted channel quality weight, the data priority weight, and the strategy resource consumption weight.

[0040] Based on the channel quality prediction level and the policy score, a transmission policy is determined based on the candidate transmission policy. The transmission policy is used to indicate normal transmission when the channel quality prediction level is greater than a preset high-quality threshold, or to reduce the transmission frequency of data frames with priority markings in the low-priority range when the channel quality prediction level is greater than a preset low-quality threshold but the channel quality prediction level is less than the high-quality threshold, or to suspend the transmission of data frames with priority markings in the low-priority range when the channel quality prediction level is less than the low-quality threshold. The high-quality threshold is greater than the low-quality threshold.

[0041] Optionally, sending the data frame to the master station according to the transmission strategy specifically includes:

[0042] Monitor the BeiDou transmission status, which includes transmission success rate, number of consecutive failures, average latency and backup link availability, 4G signal strength, network latency and available bandwidth;

[0043] When the number of consecutive failures exceeds a preset failure threshold, the transmission success rate drops below a preset success threshold within a specified time, or the channel quality prediction level is lower than a preset quality threshold, at least one available backup link is determined based on the data priority, timeliness requirements, current status of each backup link, energy consumption, and communication cost, and the data frame with the first priority is switched to the backup link.

[0044] When no available backup link is detected, the corresponding transmission parameters are adjusted according to the preset optimal transmission path configuration, and the data frame is transmitted according to the transmission parameters, wherein the transmission parameters include transmission power and data frame length.

[0045] Optionally, the master station receives data frames from the substation, performs integrity verification, data decompression and reassembly on the data frames, and feeds back the communication status to the substation, so that the substation updates the database, the data priority classification model, the channel quality prediction model and the intelligent compression model according to the communication status, specifically including:

[0046] The system receives data frames from the substation and verifies the data integrity of the data frames. It generates a retransmission request based on the corrupted data frames and returns it to the substation until the data integrity verification is passed. The retransmission request is used to instruct the substation to transmit the corrupted data frames.

[0047] Using a decompression strategy that matches the compression strategy adopted by the substation, the verified data frames are decompressed and restored. The split frames are reassembled using sequence number information to restore the complete dataset. The original data is restored by reverse calculation for the differentially encoded time-series data to obtain the complete restored data.

[0048] Based on the restored data, the data transmission quality of each substation is evaluated, and a communication status is generated and fed back to the corresponding substation. This allows the substation to update the database based on the communication status, and to update the data priority classification model, the channel quality prediction model, and the intelligent compression model based on the updated database. The transmission evaluation results include frame loss rate, latency, and data quality evaluation.

[0049] Optionally, the main station monitors the operational status of itself, the substations, and the edge computing units, provides early warnings of abnormal states based on a pre-trained anomaly detection model, and triggers corresponding self-healing strategies, specifically including:

[0050] An anomaly monitoring framework is established, wherein the anomaly monitoring framework includes data anomalies, communication anomalies, and equipment anomalies.

[0051] When the collected power data is detected to exceed the preset normal range, a data anomaly signal is issued;

[0052] When a BeiDou signal terminal is detected and / or the bandwidth of the backup link is lower than a preset bandwidth threshold, a communication abnormality signal is issued;

[0053] When the processor load of the substation exceeds a preset load threshold and / or the storage space of the substation is lower than a preset space threshold, a device abnormality signal is issued.

[0054] Based on the data anomaly signal, the communication anomaly signal, and the device anomaly signal, a corresponding anomaly feature vector is generated. The anomaly feature vector includes anomaly type, anomaly source, anomaly index, time characteristics, correlation information, and severity.

[0055] Using a pre-established reconstruction error model based on an autoencoder, a time-series prediction model based on LSTM, a correlation analysis model based on a graph neural network, and a multi-feature classification model based on XGBoost, the abnormal feature vectors are classified for anomaly type, assessed for severity, and analyzed for root causes, outputting anomaly classification results and severity levels.

[0056] Based on the anomaly classification results and the severity level, a multi-level early warning mechanism is constructed, and early warnings are issued based on the early warning mechanism.

[0057] Optionally, the main station triggers the self-healing strategy through the following steps, specifically including:

[0058] Establish a self-healing strategy library, which includes data anomaly self-healing strategies, communication anomaly self-healing strategies, and device anomaly self-healing strategies.

[0059] Based on the self-healing strategy library, according to the anomaly classification results, the severity level, system status, environmental conditions, historical experience, impact assessment, and resource constraints, the corresponding self-healing strategy is selected, and self-healing operation instructions are output according to the self-healing strategy.

[0060] The self-healing operation is performed according to the self-healing operation instructions, which include data marking repair, link parameter adjustment and switching, service restart, and configuration rollback.

[0061] This invention also provides a BeiDou power data intelligent acquisition and dynamic communication scheduling system, comprising:

[0062] The database construction module is used by the main station to perform data cleaning and feature engineering on the acquired historical power data, Beidou communication data, environmental and operating condition data, establish basic parameter configurations and write them into the database.

[0063] The model building module is used by the main station to train a data priority classification model, a channel quality prediction model, and an intelligent compression model based on the database.

[0064] The data acquisition module is used by the substation to collect power data through the edge computing unit, construct a skyline candidate set by combining the preset EPUS algorithm for data filtering, and use the data priority classification model from the main station to intelligently classify and label the power data after data filtering, and output a data stream with priority labels.

[0065] The data processing module is used by the substation to perform reliability assessment on the data stream using a preset Gaussian mixture model classification framework, obtain a reliability score, and select a compression strategy based on the reliability score, the data type identifier corresponding to the data stream, and the priority label, and package the data stream into data frames according to the compression strategy.

[0066] The data frame sending module is used by the substation to determine a transmission strategy based on the data frame and the channel quality prediction result, combined with a preset context-aware intelligent handover protocol, and to send the data frame to the master station according to the transmission strategy. The transmission strategy is used to indicate the transmission timing selection, multi-link intelligent handover, and parameter adaptive adjustment. The channel quality prediction result is obtained by the channel quality prediction model from the master station after predicting the current BeiDou signal strength.

[0067] The data restoration module is used for the main station to receive data frames from the sub-station, perform integrity verification, data decompression and reassembly on the data frames, and feed back the communication status to the sub-station, so that the sub-station updates the database, the data priority classification model, the channel quality prediction model and the intelligent compression model according to the communication status;

[0068] An anomaly monitoring module is used by the main station to monitor the operating status of itself, the substations and the edge computing units, to provide early warning of abnormal states based on a pre-trained anomaly detection model, and to trigger corresponding self-healing strategies. The abnormal states include data anomalies, communication anomalies and device anomalies.

[0069] The present invention has the following beneficial effects:

[0070] By introducing the edge-assisted parallel uncertain skyline (EPUS) algorithm, data is pre-screened at edge nodes, which effectively reduces the amount of data that needs to be transmitted and reduces the bandwidth pressure of BeiDou short messages.

[0071] The data classification reliability assessment mechanism based on Gaussian mixture model can identify samples with uncertain classification, avoid misclassification of key data, and improve the accuracy of power data classification.

[0072] By introducing the Context-Aware Intelligent Switching (CASH) protocol and combining channel quality prediction results, intelligent switching between BeiDou communication and backup links is achieved, optimizing the power data transmission path and improving the transmission success rate.

[0073] A dedicated differential coding technique was designed for time-series correlated data in power systems. Combined with data change trend prediction, it significantly improved the data compression rate and adapted to the frame length limit of BeiDou short messages.

[0074] An innovative architecture was designed to enable three types of AI models—a data priority classification model, a channel quality prediction model, and an intelligent compression model—to work together. Through information interaction and incremental learning between the models, intelligent optimization of the entire process of power data acquisition and transmission was achieved. Attached Figure Description

[0075] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0076] Figure 1 A flowchart of the BeiDou power data intelligent acquisition and dynamic communication scheduling method provided in this embodiment of the invention;

[0077] Figure 2 The structural block diagram of the Beidou power data intelligent acquisition and dynamic communication scheduling system provided in the embodiment of the present invention. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0079] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0080] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0081] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0082] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0083] See Figure 1 This invention provides a method for intelligent acquisition and dynamic communication scheduling of BeiDou power data, including the following steps 1 to 7.

[0084] Step 1: The main station performs data cleaning and feature engineering on the acquired historical power data, BeiDou communication data, and environmental and operating condition data, establishes basic parameter configurations, and writes them into the database.

[0085] The basic parameter configuration includes model trigger threshold, channel quality threshold, default value of power data acquisition frequency, default frequency level of BeiDou transmission, backup link switching conditions, feature engineering parameters, etc.

[0086] Specifically, the main station collects historical operating data from the substations, including power data (cumulative electrical energy, voltage / current measurements, switch status, etc.), BeiDou communication data (transmission success rate, latency, signal-to-noise ratio, frame loss rate, etc.), and environmental and operating condition data (weather, equipment load rate, power grid fault records, etc.) to establish an initial dataset.

[0087] The collected historical data is preprocessed, including removing outliers (such as voltage spikes and communication delay anomalies), filling in missing values, and normalizing the data to unify the data of different dimensions into the [0,1] interval, so as to ensure the stability of subsequent model training.

[0088] Key features are extracted from the preprocessed data, such as peak electricity consumption characteristics, channel attenuation patterns under severe weather, and abnormal equipment operation modes. The correlation between these features is analyzed to construct a feature vector set, providing high-quality input for AI model training.

[0089] Based on historical data analysis results, key system parameters are preset, including model trigger thresholds (such as the threshold for judging poor channel quality SNR < 10dB), default parameters for BeiDou transmission (such as the default frequency level 12, i.e., 1 minute / time), and backup link switching conditions, to complete the system initialization configuration.

[0090] Step 2: The main station trains a data priority classification model, a channel quality prediction model, and an intelligent compression model based on the database.

[0091] Specifically, based on the historical data preprocessed in step 1, a deep learning model (CNN-LSTM hybrid network) is constructed. The input includes features such as the type of power data, time scale, and equipment status. The model is trained through supervised learning to identify data priorities (levels 1-4), and the output is a validated data priority classification model.

[0092] The time series prediction algorithm (GRU network) is used as input to train the model to predict the channel quality level (good / medium / poor) for the next 5 minutes. The accuracy of the model is evaluated through the validation set, the model parameters are optimized, and the channel quality prediction model is output.

[0093] For different types of power data (structured measurements, unstructured state variables, etc.), a lightweight autoencoder model is trained to achieve a compression rate of 30%-60% while ensuring data accuracy (error ≤ 0.5%), and output a family of compression models for specific data types.

[0094] The trained AI model is quantized, compressed, and optimized for computational acceleration. It is adapted to edge computing hardware (FPGA). The data priority classification model is deployed to the substation edge computing unit, and the channel quality prediction model and intelligent compression model family are deployed to the main station and substation collaborative computing platform.

[0095] Step 3: The substation collects power data through the edge computing unit, constructs a skyline candidate set by combining the preset EPUS algorithm for data filtering, and uses the data priority classification model from the main station to intelligently classify and label the filtered power data, outputting a data stream with priority labels.

[0096] Specifically, step 3 includes:

[0097] The edge computing unit detects the electricity meter and monitoring and control device connected via RS-485 and Ethernet interfaces to collect electricity data and records the metadata generated during the collection process to form a raw data stream. The metadata includes the timestamp of data generation, device identifier, associated operating condition information and data type identifier. The electricity data includes voltage, current, power and electrical energy.

[0098] The power data is regarded as data points in a multi-dimensional space, and an importance index space is constructed. The dimensional indicators of the importance index space include the data change rate, the degree of deviation from the rated value, and the predefined importance of the data type.

[0099] Using the preset EPUS algorithm and the importance index space, the importance index of each data point is calculated, and the data point is compared with the data in the current skyline candidate set based on the importance index.

[0100] When the importance index of the data point exceeds that of the data in the current skyline candidate set, the data point is added to the skyline candidate set and updated.

[0101] The data that has been filtered by EPUS and classified into the skyline candidate set are transformed into feature vectors;

[0102] The feature vector is classified using the data priority classification model to generate priority labels and classification confidence scores;

[0103] Meta-information tags are added to the categorized data to obtain the data stream. The meta-information tags include the priority label, the data type identifier, the timestamp, and the classification confidence score.

[0104] Specifically, multi-source real-time data acquisition serves as the data entry point, enabling real-time data acquisition from power equipment through various communication interfaces. The implementation is as follows: First, the substation is equipped with a multi-interface acquisition unit, including RS-485, Ethernet, and CAN bus interfaces, which connect to different types of power equipment. The RS-485 interface is primarily used to connect traditional power measurement equipment, such as electricity meters and protection devices, using the Modbus RTU protocol; the Ethernet interface connects to new intelligent devices supporting IEC 61850 or Modbus TCP protocols, such as smart circuit breakers and new energy inverters; the CAN bus is mainly used for vehicle-mounted applications or scenarios with high real-time requirements.

[0105] The data acquisition unit actively initiates data requests at a preset acquisition frequency (initially set to: critical quantities every 30 seconds, general quantities every 5 minutes, and statistical quantities every hour) to read raw data from various power equipment. The types of data acquired include: real-time measurements (such as three-phase voltage, current, active power, reactive power, frequency, etc.), status parameters (such as switch position, alarm status, operating mode, etc.), cumulative parameters (such as forward and reverse active energy, four-quadrant reactive energy, etc.), and equipment parameters (such as temperature, humidity, oil level, etc.).

[0106] At the same time, the acquisition unit will add rich metadata to each data point, including:

[0107] Timestamp: The collection time is accurate to the millisecond level;

[0108] Device ID: Unique identifier for the device;

[0109] Data type identifier: distinguishes different types such as measured values ​​and state variables;

[0110] Engineering units: such as V (voltage), A (current), Hz (frequency), etc.

[0111] Related operating information: For example, the output power of a photovoltaic inverter is related to the current light intensity, and the output power of a wind turbine is related to the wind speed, etc.

[0112] This application also integrates environmental sensors (such as weather stations) to collect environmental parameters such as wind speed, temperature, humidity, and light intensity. These parameters, together with power data, form a multi-dimensional data stream, providing a more complete information foundation for subsequent intelligent analysis.

[0113] The acquisition unit has a built-in watchdog mechanism that can detect acquisition anomalies (such as communication interruption, data timeout, etc.) and automatically retry or switch to a backup channel to ensure the continuity of data acquisition.

[0114] The output of this step is the raw data stream, which contains complete power data and its metadata, providing input for subsequent EPUS algorithm filtering.

[0115] Subsequently, the Edge-Assisted Parallel Uncertain Skyline (EPUS) algorithm was applied to power data filtering. This algorithm was originally used to process data in low-latency IoE analysis applications, and it has now been adapted for power scenarios.

[0116] In the core concept of the EPUS algorithm, "skyline data" refers to a point in a multidimensional attribute space that is not dominated by other data points. For power systems, embodiments of this application treat various power parameters as points in a multidimensional space. For example, a measurement data containing parameters such as voltage, current, power, and frequency can be represented as a point in space.

[0117] The specific implementation process is as follows:

[0118] First, define the "importance" metric space for power data. This space includes multiple dimensions, such as the rate of change (the percentage change compared to the last data collection), the degree of deviation from the rated value, and the predefined importance of the data type. For example, for voltage data, its importance is defined as: Where V is the current voltage, It is the rated voltage. It is the rate of change of voltage. These are the predefined importance weights of the parameter, where α, β, and γ are weighting coefficients.

[0119] Then, the EPUS algorithm constructs a "skyline candidate set" on the edge computing nodes (i.e., substation acquisition units). The algorithm calculates the importance index of each newly acquired data point and compares it with the data in the current skyline candidate set. If some dimensions of the new data point exceed those of the points in the candidate set (i.e., "not completely dominated"), it is added to the candidate set; at the same time, if a point in the candidate set is completely dominated by the newly added point, it is removed from the candidate set.

[0120] In this way, the candidate set for the skyline always retains data points that excel in certain important dimensions. For example, data points where the voltage suddenly drops by 25% will be included in the candidate set, while data points that remain stable around the rated value with fluctuations of no more than 1% may be filtered out.

[0121] To adapt to the characteristics of the power system, a "mandatory retention rule" was set up, which means that certain types of data (such as circuit breaker trip signals and protection action information) will be forcibly retained in the candidate set regardless of their calculated importance.

[0122] In addition, this embodiment also implements a "time decay mechanism", which means that data points that have not been updated for a long time will have their importance scores reduced in certain dimensions, ensuring that the latest changes can be monitored.

[0123] Ultimately, the EPUS algorithm outputs a filtered dataset, which consists of either highly important "skyline data" or specific types of data that meet the mandatory retention rules. Compared to the original data stream, the amount of data after filtering is significantly reduced (by 40%-70% in practical applications), significantly reducing the computational burden of subsequent AI processing and preparing for the limited bandwidth of BeiDou communication.

[0124] After obtaining the dataset filtered by the EPUS algorithm, a pre-trained data priority classification model is used to automatically assign a priority label to each data point. This process is completed on the edge computing unit of the substation without requiring real-time communication with the main station.

[0125] The specific implementation process is as follows:

[0126] First, the filtered data is converted into an input format acceptable to the AI ​​model. For each data record, a feature vector is constructed, including: data type encoding (such as one-hot encoding of categories such as voltage, current, and power), data value (normalized), rate of change (compared with historical data), equipment status indicators, time features (weekday / non-weekday, time of day, etc.), and importance indicators calculated by the EPUS algorithm.

[0127] The feature vectors are then fed into a lightweight AI inference engine deployed on an edge computing unit. This engine runs a pre-trained CNN-LSTM hybrid network model (trained in step 2). The CNN layers are responsible for extracting local feature patterns from the data, while the LSTM layers capture dependencies in the time series, making it particularly suitable for processing time-continuous data in power systems.

[0128] The model automatically categorizes data into four priority levels based on input features:

[0129] Level 1 (Emergency): Data directly related to power grid safety, such as switch position signals, overvoltage / undervoltage alarms, and protection action information;

[0130] Level 2 (High Importance): Such as real-time power measurement values ​​(5-minute intervals), load change information, and operating status of important equipment;

[0131] Level 3 (General): Such as hourly cumulative electrical energy value, voltage / current fluctuations within the normal range, etc.

[0132] Level 4 (Low Importance): Such as daily / monthly statistical data, cumulative equipment uptime, etc.

[0133] To improve the efficiency of AI inference at the edge, the model adopts quantization technology (such as 8-bit fixed-point quantization) and is optimized for edge computing hardware (such as FPGA). The inference time for a single data point is controlled in the millisecond level (typical value <10ms), which meets the requirements of real-time processing.

[0134] In addition, the model output includes not only priority categories but also classification confidence scores (values ​​between 0 and 1). When the classification confidence score is below a threshold (default 0.75), the data is marked as "uncertain classification" for special handling in subsequent steps.

[0135] Priority determination results are synchronously recorded in the local log for subsequent incremental learning and performance optimization of the model. The distribution of data for each priority level is analyzed daily to detect whether the model has any biases (for example, if the proportion of level 1 data increases abnormally, it may indicate that the model has misjudged).

[0136] The output of this step is a dataset with priority labels, which provides the basis for subsequent data labeling and association with metadata.

[0137] Specifically, the data categorized into the skyline candidate set after EPUS filtering is transformed into feature vectors. The data priority classification model is then used to classify these feature vectors, generating priority labels and classification confidence scores. Meta-information tags are added to the classified data to obtain the data stream. These meta-information tags include the priority labels, the data type identifier, the timestamp, and the classification confidence score. The substation collects power data through edge computing units, constructs a skyline candidate set using a preset EPUS algorithm for data filtering, and uses the data priority classification model from the main station to intelligently classify and label the filtered power data, outputting a data stream with priority labels.

[0138] After completing the AI-driven priority determination, rich metadata tags are added to the data to build a structured data stream, providing a comprehensive basis for subsequent compression and transmission decisions.

[0139] The specific implementation process is as follows:

[0140] First, a unique identifier (UUID) is generated for each piece of data to ensure accurate tracking of the processing status of each piece of data throughout the entire data flow. This UUID adopts a hierarchical design, containing information such as substation ID, device ID, data type code, and timestamp, ensuring uniqueness even in offline conditions.

[0141] Then, append the following metadata tags to the data:

[0142] Priority labels: AI judgment results from step 3.3 (levels 1-4);

[0143] Classification confidence score: The classification reliability score (value between 0 and 1) output by the AI ​​model;

[0144] Data type identifier: detailed classification code, such as "V_A" (phase A voltage), "P_TOTAL" (total active power), etc.;

[0145] Timestamp: Includes data collection time (accurate to milliseconds) and data entry time;

[0146] Validity period indicator: indicates the timeliness of the data, such as "real-time data" (requires immediate processing), "near real-time data" (allows appropriate delays), and "historical data" (used for statistical analysis).

[0147] Associated data pointer: Linked to other data points related to the current data, such as voltage and current in the corresponding phase;

[0148] Engineering units: Specify the physical unit of the data, such as "V" (volt), "kW" (kilowatt), etc.;

[0149] Data quality markers: Indicate the reliability of data, such as "valid", "questionable", "invalid", etc.

[0150] For specific types of data, the system will also add special tags. For example, for status data (such as switch position), a "status duration" tag will be added; for measured values, a "percentage deviation from the rated value" tag will be added; and for alarm information, "alarm level" and "associated device" tags will be added.

[0151] Furthermore, the data is correlated with its operating environment. For example, power data for photovoltaic power generation equipment is correlated with the current sunlight intensity; for wind power equipment, it is correlated with the current wind speed; and for power distribution line load, it is correlated with the current temperature and workday type. This correlation information helps the master station to more comprehensively understand the environmental context of the data.

[0152] Finally, based on data characteristics and priorities, preset data processing strategy labels are set, such as "compression strategy" (indicating which compression algorithm should be used subsequently), "transmission strategy" (such as "real-time transmission", "batch transmission", "condition-triggered transmission", etc.), and "caching strategy" (such as "local permanent storage", "temporary cache", etc.).

[0153] Through this series of labeling and association processes, the raw data is transformed into a structured, information-rich data stream, which not only contains the value of the data itself, but also comprehensive information such as the source, importance, correlation, and processing strategy of the data, providing a solid foundation for subsequent intelligent processing.

[0154] The output of this step is a structured data stream with complete metadata tags, which is directly used as the input to step 4 (adaptive data compression and frame packing).

[0155] Step 4: The substation uses a preset Gaussian mixture model classification framework to perform a reliability assessment on the data stream, obtains a reliability score, and selects a compression strategy based on the reliability score, the data type identifier corresponding to the data stream, and the priority label, and packages the data stream into data frames according to the compression strategy.

[0156] In one embodiment, specifically, step 4 includes:

[0157] A feature space representation is established for different types of power data. Numerical features and contextual features are extracted based on the feature space representation. A preset GMM model is trained based on the numerical features and the contextual features to obtain a Gaussian mixture model.

[0158] Using the Gaussian mixture model, the probability density value of each data stream belonging to the classified priority and the probability density value of each other priority category are calculated. The reliability index evaluation value is calculated by comparing the probability density values, and the reliability score is composed of the index evaluation value. The reliability index includes intra-class reliability index, inter-class difference index and anomaly index.

[0159] When the priority label of the data stream is detected to be within a preset low priority range, but the reliability score is higher than a preset high reliability threshold, the priority corresponding to the priority label is increased.

[0160] When the priority label of the data stream is detected to be within a preset high priority range, but the reliability score is lower than a preset low reliability threshold, the context information corresponding to the data stream is appended to the data stream.

[0161] Based on at least one of the priority label, the reliability score, and the data type identifier, a compression strategy is selected, and the data stream is compressed according to the compression strategy to obtain compressed data;

[0162] Based on the frame structure, data size, and priority tag of BeiDou short messages, the compressed data is packaged into data frames.

[0163] In this embodiment, the specific process of training the preset GMM model based on numerical features and contextual features is as follows: First, numerical features are extracted from historical power data, including standardized values ​​of voltage amplitude, normalized results of current RMS values, numerical representation of power factor, and quantized values ​​of frequency deviation. Simultaneously, contextual features are extracted, including the encoding of the data acquisition time period (e.g., using values ​​0-23 to represent hourly segments), the encoding of equipment operating status (e.g., normal operation is encoded as 1, abnormal operation as 0), the numerical representation of ambient temperature, and the graded values ​​of load levels.

[0164] Taking voltage data as an example, a feature vector is constructed. Numerical features include the standardized value obtained by dividing the current voltage by the rated voltage, the numerical representation of the voltage change rate, and the quantified value of voltage harmonic content. Contextual features include the numerical encoding of the acquisition time, the numerical representation of the equipment load rate, and the numerical characteristics of environmental factors such as temperature and humidity. These features are combined to form a multi-dimensional feature vector, with one feature vector corresponding to each power data sample.

[0165] During training, training sets are built for power data of different priorities. For Level 1 priority data, all historically marked power data as emergency and their corresponding feature vectors are collected; for Level 2 priority data, feature vectors of all historically marked high importance are collected, and so on. For each priority category, the Expectation-Maximization (EM) algorithm is used to estimate the parameters of the Gaussian Model (GMM), and the mean vector, covariance matrix, and mixing weights of each Gaussian component are determined through an iterative optimization process.

[0166] In the specific training process, the model parameters are first initialized, such as setting the number of Gaussian components to 3 to 5, and randomly initializing the mean and covariance of each component. Then, the iterative training phase begins. In the expectation step, the posterior probability of each training sample belonging to each Gaussian component is calculated, and in the maximization step, the model parameters are updated based on these probabilities. The changes in the log-likelihood function value are monitored during the training process, and training stops when the change is less than a preset threshold or the maximum number of iterations is reached. After training is completed, the model is validated using an independent validation dataset to evaluate the model's classification accuracy and the reliability of the probability density estimation, ensuring that the model can accurately reflect the feature distribution patterns of data with different priorities.

[0167] It should be noted that Gaussian mixture models are a type of unsupervised machine learning model based on probability and statistics. They assume that the data is a probability distribution model composed of a mixture of multiple Gaussian distributions. The core idea of ​​this model is to represent complex data distributions as linear combinations of several simple Gaussian distributions, each of which is called a component. The probability distribution of the actual data is approximated by weighted combinations of different components.

[0168] The mathematical foundation of the Gaussian Mixture Model (GMM) is built upon probability and statistics. Each Gaussian component is defined by three key parameters: the mean vector describes the central location of the component in the feature space; the covariance matrix characterizes the dispersion and correlation of the data across various feature dimensions; and the mixture weights represent the importance of the component in the overall mixture model. The overall probability density function of the model is a weighted sum of the probability density functions of all Gaussian components, and the sum of the weights must equal 1 to satisfy the basic requirements of the probability distribution.

[0169] The training of the GMM model employs the Expectation-Maximization (EM) algorithm, an iterative optimization algorithm particularly suitable for parameter estimation of probabilistic models containing latent variables. The algorithm consists of two alternating steps: the expectation step calculates the posterior probability of each data point belonging to each Gaussian component under the current parameters, which is equivalent to soft assignment of the latent variables; the maximization step updates the model parameters based on these posterior probabilities, including recalculating the mean, covariance, and mixture weights of each component. Through multiple iterations, the algorithm progressively optimizes the model parameters, causing the likelihood function value of the model with respect to the training data to continuously increase, eventually converging to a local optimum.

[0170] In intelligent power data acquisition applications, the Gaussian Mixture Model (GMM) has the advantage of capturing the complex distribution characteristics of data with different priorities. Data in power systems often exhibits multimodal distribution characteristics; for example, the distribution of data under normal operating conditions differs significantly from that under abnormal conditions, making it difficult to accurately model with a single Gaussian distribution. GMM, through the combination of multiple Gaussian components, can effectively describe this complex distribution pattern, providing a solid theoretical foundation for data classification reliability assessment. Furthermore, the GMM model can output the probability value of each data point belonging to different categories; this soft classification result provides a quantitative basis for subsequent reliability assessment and priority adjustment.

[0171] After receiving the priority-labeled data stream output from step 3, a classification framework based on Gaussian mixture model is introduced to conduct an in-depth evaluation of the reliability of the data classification results in order to avoid misclassification of important data.

[0172] The specific implementation process is as follows:

[0173] First, feature space representations are established for different types of power data. For each data type (such as voltage, current, power, etc.), multiple feature dimensions are extracted, including numerical features (such as standardized measurements, rate of change, etc.) and contextual features (such as acquisition time, associated operating conditions, etc.). These features together constitute a multi-dimensional feature space, in which each data record forms a point.

[0174] Then, a Gaussian Mixture Model (GMM) is applied to model these data points. The GMM assumes that the data distribution for each priority category (levels 1-4) can be represented by a weighted sum of multiple Gaussian distributions. For each data category, the system pre-trains the corresponding GMM model, and the model parameters (mean vector, covariance matrix, and weights of each Gaussian distribution) have been determined in the model training phase of step 2.

[0175] During the real-time processing phase, the probability density value P(x|c) of each data point belonging to its classification priority is calculated, where x is the feature vector of the data and c is its classification priority category. Simultaneously, the probability density values ​​of the data belonging to each of the other priority categories are calculated. By comparing these probability density values, the following reliability metrics are derived:

[0176] Intra-class reliability: The position of a data point in the probability density distribution of its classified category, reflecting the degree of matching between the data and the typical characteristics of that category;

[0177] Inter-class variability: The difference between the probability density of a data point in its class and the probability density of its next-highest class, reflecting the certainty of the classification;

[0178] Anomaly score: The maximum probability density of a data point across all categories. If it is generally low, it indicates that the data may be anomalous.

[0179] Based on these metrics, a comprehensive reliability score (a value between 0 and 1) is calculated. Data with a reliability score below the threshold (default 0.7) is specially marked, and different processing strategies are adopted according to specific circumstances:

[0180] If data was originally classified as low priority (level 3-4) but has low reliability, it should be raised to a higher priority (usually level 2) to prevent important data from being erroneously delayed in processing.

[0181] If the data was originally classified as high priority (level 1-2) but has low reliability, its priority remains unchanged, but more contextual information is added during transmission to help the main station make further judgments;

[0182] In addition, for obviously abnormal data (such as measurements that deviate significantly from historical data), a "suspicious data" label is marked and an additional verification process is triggered, such as requesting other parameters of the relevant equipment or increasing the sampling frequency for confirmation.

[0183] Therefore, this embodiment solves the problem that traditional deep learning classifiers have difficulty reliably evaluating uncertain samples by applying GMM to power data classification reliability assessment, and is particularly suitable for error-prone scenarios such as data importance judgment in power systems.

[0184] The output is a dataset with reliability metrics, which will directly affect the subsequent compression strategy selection and transmission decisions.

[0185] Based on the data reliability assessment results, intelligent compression strategy selection is realized. Different compression methods are adopted for data with different priorities, types and reliability to maximize transmission efficiency while ensuring data quality.

[0186] The specific implementation process is as follows:

[0187] First, a multi-layered compression strategy library was established, containing various compression algorithms and their parameter configurations:

[0188] Lossless compression algorithms, such as Huffman coding, LZ77 / LZ78, and Deflate, are suitable for critical data where no information loss is permissible.

[0189] Lossy compression algorithms, such as wavelet transform-based compression, piecewise linear approximation, and dynamic quantization, are suitable for measurement data where a certain degree of accuracy loss is permissible.

[0190] Specialized compression algorithms: such as run-length encoding for digital signals and Fourier transform compression for periodic data, optimized for specific types of data;

[0191] Then, based on the data's priority, type, and reliability score, the following decision-making process is executed:

[0192] For Level 1 data (urgent):

[0193] High reliability (≥0.9): Lossless compression with a low compression ratio (≤30%) is used to ensure data integrity, such as by using optimized Huffman coding;

[0194] In terms of reliability (0.7-0.9): lossless compression is still used, but additional parity bits are added to improve transmission reliability;

[0195] Low reliability (<0.7): Redundant coding method is used to transmit both raw and simplified data simultaneously to ensure that critical information is not lost;

[0196] For Level 2 data (high importance):

[0197] High reliability: It adopts a hybrid compression with a medium compression ratio (30%-40%), which performs lossless compression on critical fields and lossy compression on secondary fields;

[0198] Medium / Low Reliability: Reduce compression ratio, increase redundant information, and ensure reliable data recovery;

[0199] For level 3-4 data (general / low importance):

[0200] High reliability: It adopts lossy compression with a high compression ratio (40%-60%), such as using significant bit compression for electrical energy data (retaining 3-4 significant digits).

[0201] For reliability: appropriately reduce the compression ratio (35%-50%) to retain more details;

[0202] Low reliability: Re-evaluate the data, perhaps by prioritizing it or by employing a specialized anomalous data compression strategy;

[0203] For specific types of data, the system also employs customized compression strategies. For example:

[0204] For rate-based electricity data (including electricity values ​​for multiple time periods such as peak, flat, and valley), joint encoding compression is used to improve compression efficiency by leveraging the correlation between electricity values ​​in different time periods.

[0205] For three-phase voltage / current data, compression is performed using the three-phase balance characteristics, transmitting only the unbalanced component and reference phase information;

[0206] For state data (such as switch positions), use state change timestamps for compression and transmit only when the state changes.

[0207] During compression, the identifier and key parameters of the compression algorithm used are recorded and appended as metadata to the compressed data to ensure that the main station can decompress it correctly. Simultaneously, a CRC-16 checksum is appended to the compressed data to verify its integrity.

[0208] In addition, the compression effect is dynamically monitored, and the actual compression ratio and compression / decompression time of various types of data are recorded for continuous optimization of the compression strategy. If the compression effect of a certain type of data is found to be poor (such as the compression ratio being lower than expected), the compression parameters are automatically adjusted or the compression algorithm is switched.

[0209] Subsequently, a dynamic priority enhancement mechanism based on reliability scores was implemented to prevent important data from being misclassified as low priority, leading to transmission delays. The specific implementation process is as follows: First, a low-priority range is defined, typically including level 3 (generally important) and level 4 (low important) data. Under normal circumstances, this data will be placed in a lower transmission priority sequence. Simultaneously, a high reliability threshold is set, with a default value of 0.85, indicating that the GMM model has a high confidence level in classifying the data.

[0210] When a data stream initially classified as a Level 3 or 4 priority is detected, but its reliability score exceeds 0.85, a priority reassessment process is triggered. The feature vector of this data is analyzed, focusing on whether its numerical characteristics exhibit abnormal patterns, such as a sudden voltage deviation exceeding 10% of the rated value, a sharp drop in power factor, or frequency fluctuations outside the normal range. If these abnormal characteristics are found, the data is considered to potentially contain important system status information, and its priority should be increased.

[0211] Specifically, the priority improvement strategy employs a gradual adjustment method. For data originally at level 4, if the reliability score is between 0.85 and 0.90, it is improved to level 3; if the score is between 0.90 and 0.95, it is improved to level 2; and if the score exceeds 0.95 and the data shows obvious anomalies, it is improved to level 1. For data originally at level 3, a score exceeding 0.85 is improved to level 2, and a score exceeding 0.92 is improved to level 1. Detailed information for each priority improvement is recorded, including the original priority, the improved priority, the reliability score, and the specific characteristics triggering the improvement, for subsequent model optimization and performance evaluation.

[0212] In addition, a priority escalation safety mechanism is implemented. If a substation triggers priority escalation more than a preset number (default 20 times) within a short period (e.g., within 1 hour), an alarm signal is issued, indicating potential data collection anomalies or model classification bias, requiring manual intervention. Simultaneously, the priority escalation frequency for various data types is statistically analyzed. If the escalation frequency for a certain type of data remains consistently high, a model retraining process will be triggered to improve classification accuracy.

[0213] Furthermore, a context information appending strategy is implemented for high-priority data with low reliability scores to ensure that this data carries sufficient background information during transmission for the main station to make accurate judgments. The specific implementation process is as follows: First, a high-priority range is defined, including Level 1 (urgent) and Level 2 (highly important) data, which typically require priority transmission and processing. Simultaneously, a low reliability threshold is set, with a default value of 0.65, indicating that the GMM model has relatively low confidence in the classification results for this data.

[0214] When a data stream is detected to be classified as a Level 1 or Level 2 priority but has a reliability score below 0.65, context information collection and additional processes are initiated. Multi-dimensional context information related to the data is automatically collected, including temporal context such as the precise timestamp of data acquisition, comparison information with historical data from the same period, and the system operating mode at the time of acquisition. Device context includes the model, years of operation, recent maintenance records, and current operating parameters of the data source equipment. Environmental context covers meteorological conditions such as ambient temperature, humidity, and wind speed at the time of acquisition, as well as power grid load levels and the operating status of other equipment.

[0215] In addition, this application includes supplementary contextual information related to data quality, including technical parameters such as signal strength, communication delay, and retransmission count during the data acquisition process. For measurement data, it includes supplementary information such as measurement accuracy level, calibration date, and range information. For status data, it includes supplementary information such as the duration of status changes, comparison of device parameters before and after the changes, and response status of associated devices. This contextual information is organized in a structured manner to form supplementary fields in the data packet.

[0216] During the context information appending process, an intelligent compression strategy is employed to control the amount of data in the appended information. For critical context information, lossless compression is used to ensure information integrity; for secondary background information, lossy compression is used to reduce the amount of data. The level of detail in the appended information is dynamically adjusted according to the frame length limitations of BeiDou short messages. When there is sufficient available frame space, complete context information is appended; when frame space is limited, the most critical context elements, such as the time of the anomaly and the status of associated devices, are appended first.

[0217] A priority ranking mechanism for context information has been established. For Level 1 data, security-related context information, such as the triggering reason for protection actions, fault duration, and scope of impact, is attached first. For Level 2 data, operational-related context information, such as load change trends, equipment operating efficiency, and expected recovery time, is attached first. All attached context information is digitally signed or verified using a checksum to ensure it is not tampered with during transmission. After receiving this data with rich context information, the master station can more accurately determine the true importance of the data and avoid processing errors caused by uncertain classification.

[0218] After completing the basic compression strategy selection, differential coding and time-series compression are implemented for power data with time-series correlation to further improve compression efficiency, which is particularly important for bandwidth-constrained communication methods such as BeiDou short messages.

[0219] The specific implementation process is as follows:

[0220] First, identify data types with time-series correlations. In power systems, many data exhibit significant time-series correlations, such as:

[0221] Continuous measurements such as voltage, current, and power typically change gradually over a short period of time;

[0222] The cumulative electrical energy shows a monotonically increasing trend, and the increments at adjacent time points usually exhibit a regularity.

[0223] The operating status of equipment, such as transformer temperature and battery SOC (state of charge), usually changes according to physical laws;

[0224] Then, different differential encoding strategies are applied according to the data type:

[0225] For measured data (such as voltage, current, etc.), first-order differential encoding is used, that is, the original value Xi is not transmitted directly, but the difference with the previous time step is transmitted. Due to the physical inertia of power systems, these differences are often much smaller than the original values ​​and can be represented with fewer bits. For example, voltage may fluctuate around 220V, with a fluctuation range of ±5V. With differential encoding, only this ±5V change needs to be represented, instead of the full 220V.

[0226] For cumulative data (such as electrical energy), incremental encoding is used, which transmits the incremental values ​​at adjacent time points and combines them with a load forecasting model. A simplified load forecasting model is established based on historical data from the same period, which can predict the approximate range of increments. Only the deviation between the actual increment and the predicted increment needs to be transmitted, significantly reducing the amount of data.

[0227] For periodically changing data, Fourier transform-based encoding is applied. By performing Fourier analysis on historical data, the main frequency components are extracted, and only the deviation between the current data and the data reconstructed based on these main frequency components is transmitted. This approach is suitable for data such as load curves with obvious periodicity.

[0228] To further improve compression efficiency, this application also implements predictive coding:

[0229] Linear prediction-based: Using a simplified autoregressive (AR) model, the current value is predicted based on the first N sampling points, and only the prediction error is transmitted;

[0230] Pattern matching based: For data with typical patterns (such as daily load curves), multiple typical pattern templates are pre-stored, and only the identifier of the best matching template and key deviation points need to be transmitted;

[0231] Physical model-based: For specific equipment (such as photovoltaic power generation), a simplified physical model is established by combining meteorological data and equipment characteristics to predict the output value, transmitting only the prediction error;

[0232] To ensure decoding reliability, the following strategies were adopted:

[0233] Regularly transmit complete reference points (e.g., once per hour) to avoid the accumulation of differential errors;

[0234] For critical data points (such as load spikes), transmit the complete value instead of the differential value;

[0235] Dynamically adjust differential precision and adaptively select the number of encoding bits based on the degree of data change.

[0236] Furthermore, this application implements multi-scale time-series compression, that is, it employs different compression strategies for data at different time scales. For example, for voltage data, there may be simultaneous data requirements at the second, minute, and hour levels:

[0237] Second-level data: Utilizes high-precision differential encoding to capture rapidly changing data;

[0238] Minute-level data: represented using downsampling and statistical features (such as mean, maximum, and minimum values);

[0239] Hourly data: using coarser-grained statistical features and trend information;

[0240] The output of this step is a data stream that has undergone differential coding and temporal compression, which further reduces the amount of data compared to the original compressed data (in practical applications, it can reduce the amount of data by an additional 30%-50%), thus preparing it for the final frame packaging.

[0241] After completing data compression and differential coding, intelligent frame packaging and optimization are implemented to maximize communication efficiency, taking into account the special limitations of BeiDou short messages (1835 bits / frame for regional services).

[0242] The specific implementation process is as follows:

[0243] First, the compressed data is categorized and organized in preparation for packaging. The data is grouped according to priority (levels 1-4), destination (e.g., different main station systems), and data type (measurements, status variables, statistics, etc.) to facilitate subsequent intelligent combination.

[0244] Next, a frame structure template library was established. According to the technical specifications of BeiDou short message service, each data frame includes a frame header (containing source address, destination address, frame type, etc.), a payload, and a frame trailer (verification information). Several frame structure templates are predefined, such as "emergency alarm frame," "routine measurement frame," and "statistical data frame," each with a specific field arrangement and encoding method.

[0245] The core intelligent packaging algorithm operates based on the following strategy:

[0246] Priority-driven frame allocation:

[0247] Level 1 data (urgent): Packed separately, one data point per frame, ensuring the fastest possible transmission;

[0248] Level 2 data (high importance): can be combined and packaged, but the number of data points per frame is limited (usually ≤5) to ensure faster processing;

[0249] Level 3-4 data (general / low importance): Multiple data points are combined and packaged to maximize the use of frame space;

[0250] Data combinations with perceived relevance:

[0251] Analyze the correlations between data and group related data within the same frame. For example, three-phase voltage data are typically processed together and should be packaged into the same frame;

[0252] For data with contextual relationships (such as multiple parameters of a device), it is preferable to combine them into the same frame for easier processing at the receiving end;

[0253] Dynamic frame length optimization:

[0254] Based on channel quality prediction results, the frame length is dynamically adjusted. When the channel quality is good, the frame length is close to the maximum (1835 bits); when the channel quality is poor, the frame length is reduced (e.g., reduced to around 900 bits) to improve the single-frame transmission success rate.

[0255] For high-priority data, maintain a shorter frame length even when the channel is good to reduce transmission time;

[0256] Intelligent frame splitting and serialization:

[0257] When a single logical data unit (such as a set of related measurements) exceeds the frame length limit, it is automatically split.

[0258] Add sequence information to the split frames, including "frame number / total number of frames" (e.g., "frame 1 / 3") and frame association identifier to ensure that the receiver can correctly reassemble them;

[0259] For split, multi-frame data, assess its integrity requirements. If complete reception is necessary for meaningful data reception, all frames should use the same priority. If partial reception is also valuable, decreasing priority can be used.

[0260] Frame header compression and multiplexing:

[0261] For multiple frames transmitted consecutively, frame header compression technology is used, so that subsequent frames only transmit frame header fields that are different from those of the previous frames, thereby reducing redundancy;

[0262] For periodic data with a fixed format, predefined frame template IDs are used instead of detailed field descriptions to reduce metadata overhead.

[0263] Service quality assurance mechanism:

[0264] Add redundant check bits to high-priority frames to improve transmission reliability;

[0265] Important frames are set with an automatic retransmission flag, triggering a retransmission if no acknowledgment is received;

[0266] Key data is packaged using an interleaved approach to prevent continuous data loss due to sudden interference.

[0267] During the frame packing process, detailed packing logs are recorded, including which frame each data point is assigned to, what compression method is used, and the expected transmission time. This information is used for system performance evaluation and optimization.

[0268] Finally, the optimized data frame set is output, with each frame containing a clear priority marker, sequence information, integrity check code, and necessary parsing instructions, providing input for the dynamic scheduling of BeiDou communication in step 5. Through intelligent frame packaging, the transmission efficiency of effective information is maximized within the limited bandwidth of BeiDou short messages.

[0269] In some embodiments, the compression strategy is used to indicate the compression ratio and / or compression method of the data stream; the step of selecting a compression strategy based on at least one of the priority label, the reliability score, and the data type identifier, and compressing the data stream according to the compression strategy to obtain compressed data, specifically includes:

[0270] Based on the priority label and the reliability score, a compression strategy is determined. Specifically, for a priority label of first priority, when the reliability score is greater than a preset first reliability threshold, a specified first compression ratio is used to compress the data stream; or when the reliability score is not greater than the first reliability threshold, the data stream is combined with simplified data determined from the data stream to generate compressed data. For a priority label of second priority, a specified second compression ratio is used to compress the data stream. The first priority is higher than the second priority, and the first compression ratio is lower than the second compression ratio.

[0271] For power data with time-series correlation, the data difference is differentially encoded by combining the current data change trend and the preset load forecasting model to obtain compressed data. The data difference is determined by the difference between the power data at the current timestamp and the power data at the previous timestamp.

[0272] In determining the compression strategy, a hierarchical decision-making mechanism is adopted, comprehensively considering two key factors: data priority labels and reliability scores. For first-priority data, i.e., level 1 urgent data, a first reliability threshold of 0.90 is set. When the data reliability score is greater than this threshold, it indicates that the data classification is accurate and the content is reliable. At this time, a specified first compression rate is applied for processing. The first compression rate is usually set to a low value, such as 15% to 25%, to ensure data integrity and accuracy and avoid loss of critical information due to over-compression.

[0273] When the reliability score of first-priority data is no greater than 0.90, the data is considered to have potential classification uncertainty or content anomalies. In this case, a special dual-protection compression strategy is adopted. The complete information of the original data stream is preserved, while a simplified version of the data stream is generated. The simplified data includes the most important information such as core key fields, device identifier, anomaly type, and occurrence time. The two are combined to generate the final compressed data packet, ensuring that even if some data is lost during transmission, the receiving end can still obtain basic key information. Although this strategy increases the data volume, this redundancy protection is necessary for urgent data with questionable reliability.

[0274] For second-priority data, also known as Level 2 high-importance data, the system uses a specified second compression rate, typically set between 30% and 45%, which is higher than the first compression rate. Since first priority is higher than second priority, resource allocation prioritizes the transmission quality of higher-priority data; therefore, the first compression rate must be lower than the second compression rate to ensure better fidelity for urgent data. Fine-tuning is performed based on the specific type of second-priority data; for example, a 35% compression rate is used for real-time power measurement data, and a 40% compression rate is used for equipment operating status data, achieving refined compression control.

[0275] In determining the compression strategy, a hierarchical decision-making mechanism is adopted, comprehensively considering two key factors: data priority labels and reliability scores. For first-priority data, i.e., level 1 urgent data, a first reliability threshold of 0.90 is set. When the data reliability score is greater than this threshold, it indicates that the data classification is accurate and the content is reliable. At this time, a specified first compression rate is applied for processing. The first compression rate is usually set to a low value, such as 15% to 25%, to ensure data integrity and accuracy and avoid loss of critical information due to over-compression.

[0276] When the reliability score of first-priority data is no greater than 0.90, the data is considered to have potential classification uncertainty or content anomalies. In this case, a special dual-protection compression strategy is adopted. The complete information of the original data stream is preserved, while a simplified version of the data stream is generated. The simplified data includes the most important information such as core key fields, device identifier, anomaly type, and occurrence time. The two are combined to generate the final compressed data packet, ensuring that even if some data is lost during transmission, the receiving end can still obtain basic key information. Although this strategy increases the data volume, this redundancy protection is necessary for urgent data with questionable reliability.

[0277] For second-priority data, i.e., level 2 highly important data, a specified second compression rate is used. This compression rate is typically set between 30% and 45%, higher than the first compression rate. Since first priority is higher than second priority, resource allocation prioritizes the transmission quality of higher-priority data; therefore, the first compression rate must be lower than the second compression rate to ensure better fidelity for urgent data. Fine-tuning is performed based on the specific type of second-priority data; for example, a 35% compression rate is used for real-time power measurement data, and a 40% compression rate is used for equipment operating status data, achieving refined compression control.

[0278] The frame structure, data size, and priority tag based on BeiDou short messages are used to package the compressed data into data frames, including:

[0279] For compressed data with the priority label set to the first priority, each compressed data is packaged into a separate data frame.

[0280] For compressed data with the priority tag set to the second priority, the compressed data is combined and packaged according to the number of data points in the frame structure to form a data frame;

[0281] For compressed data whose priority label is lower than the second priority, multiple compressed data are concatenated into the same data frame.

[0282] During the data frame packaging stage, differentiated packaging strategies are implemented based on the frame structure limitations of BeiDou short messages and the transmission requirements of data with different priorities. The single frame capacity of the BeiDou short message system is limited to 1835 bits, and it is necessary to maximize transmission efficiency under this constraint.

[0283] For first-priority compressed data, an independent packetization strategy is employed, packaging each compressed data piece into a separate complete data frame. This strategy ensures that urgent data receives the fastest transmission speed and the highest transmission priority, avoiding processing delays caused by merging with other data. Each independent data frame contains complete frame header information, data payload, and checksum information, guaranteeing the independence and integrity of individual urgent data even under network congestion or transmission errors. Special flag bits are also assigned to these independent frames to instruct the receiver to prioritize their processing.

[0284] For second-priority compressed data, a combination and packaging strategy is implemented. Based on the data capacity limitations of the frame structure and the actual size of the data points, multiple related data points are combined into the same data frame. The correlation between data points is analyzed, and multiple parameters from the same device or the same measurement time are prioritized for combination and packaging to facilitate unified processing and analysis at the receiving end. During combination and packaging, it is ensured that the utilization rate of each data frame reaches more than 80% to avoid wasting frame space, while reserving sufficient space for necessary metadata and verification information.

[0285] For compressed data of lower priority than the second, including level 3 and 4 data, a high-density splicing and packaging strategy is adopted to splice as many data points as possible into the same data frame, maximizing the information capacity of a single frame. This strategy is suitable for statistical and historical data with relatively low timeliness requirements but large data volumes. During the splicing process, data is grouped according to data type, and data of the same type are packaged together for subsequent batch processing. At the same time, detailed index information is added to these high-density packaged data frames, including the starting position and length of each data point in the frame, to ensure that the receiving end can accurately parse and extract each data item.

[0286] In all packetization strategies, the actual frame size is dynamically adjusted based on the current channel quality prediction results. When the channel quality is good, the frame size is kept as close as possible to the maximum frame length limit to improve transmission efficiency; when the channel quality is poor, the frame length is appropriately reduced to improve the success rate of single-frame transmission and avoid transmission failures and retransmission overhead caused by excessively large frames.

[0287] Step 5: The substation determines a transmission strategy based on the data frame and channel quality prediction results, combined with a preset context-aware intelligent handover protocol, and sends the data frame to the master station according to the transmission strategy. The transmission strategy is used to indicate the transmission timing selection, multi-link intelligent handover, and parameter adaptive adjustment. The channel quality prediction results are obtained by the channel quality prediction model from the master station after predicting the current BeiDou signal strength.

[0288] In one embodiment, the substation determines a transmission strategy based on the data frame and channel quality prediction results, combined with a preset context-aware intelligent handover protocol, specifically including:

[0289] The BeiDou signal strength within a previously specified time period is input into the channel quality prediction model to predict the channel quality changes within a subsequent specified time period, and the channel quality prediction result is output. The channel quality prediction result includes the channel quality prediction level, confidence interval, and trend of change.

[0290] Combining a preset context-aware intelligent handover protocol, the strategy score of the candidate transmission strategy is obtained based on the channel quality score at the current moment, the channel quality prediction level for the specified time period thereafter, the data priority, the resource consumption of the candidate transmission strategy, the current channel quality weight, the predicted channel quality weight, the data priority weight, and the strategy resource consumption weight.

[0291] Based on the channel quality prediction level and the policy score, a transmission policy is determined based on the candidate transmission policy. The transmission policy is used to indicate normal transmission when the channel quality prediction level is greater than a preset high-quality threshold, or to reduce the transmission frequency of data frames with priority markings in the low-priority range when the channel quality prediction level is greater than a preset low-quality threshold but the channel quality prediction level is less than the high-quality threshold, or to suspend the transmission of data frames with priority markings in the low-priority range when the channel quality prediction level is less than the low-quality threshold. The high-quality threshold is greater than the low-quality threshold.

[0292] In this embodiment, channel quality prediction based on time-series analysis is implemented, using historical BeiDou signal data to predict future communication environment trends. First, a specified time period is defined, typically setting the past 30 minutes as the historical data input window and the next 15 minutes as the prediction target time period. This time window selection is based on a balance between the statistical regularity of BeiDou signal changes and the real-time requirements of power data transmission.

[0293] The system collects BeiDou signal strength data from the previous 30 minutes, including multi-dimensional parameters such as signal received power, signal-to-noise ratio, Doppler shift, and pseudorange measurement accuracy. These raw signal parameters are converted into standardized signal strength indices to eliminate the influence of equipment differences and environmental noise. Simultaneously, relevant environmental factor data are collected, such as current meteorological conditions, ionospheric activity index, and satellite geometric distribution factor, all of which affect the propagation quality of BeiDou signals.

[0294] The channel quality prediction model employs a deep learning architecture, combining the advantages of long short-term memory networks and convolutional neural networks to simultaneously capture the temporal characteristics and spatial correlations of signal changes. The model uses historical signal strength data as temporal input and extracts deep feature patterns of signal changes through a multi-layered neural network structure, identifying different types of signal evolution characteristics such as periodic variation patterns, sudden interference patterns, and gradual trends. The model training process utilizes a large amount of historical data, covering signal change patterns across different seasons, weather conditions, and time periods, ensuring the accuracy and generalization ability of the predictions.

[0295] The prediction output comprises three key elements: The channel quality prediction level uses a five-level classification system, from Level 1 (Excellent) to Level 5 (Very Poor), with each level corresponding to a specific signal strength range and expected transmission success rate. The confidence interval represents the reliability range of the prediction result, usually given as a percentage; for example, a 95% confidence interval means the prediction result has a 95% probability of falling within the specified range. The trend describes the evolution direction of channel quality over a future time period, including four basic trend types: stable, rising, falling, and fluctuating, providing forward-looking guidance for transmission strategy formulation.

[0296] Furthermore, a context-aware intelligent handover protocol is used to comprehensively evaluate and quantitatively score multiple candidate transmission strategies. This protocol integrates several key factors and calculates a comprehensive score for each candidate strategy through weighted calculation, providing a quantitative basis for the final transmission decision.

[0297] The current channel quality score is calculated based on real-time monitored BeiDou signal parameters, using a scoring system from 0 to 100. A score of 90 or above indicates excellent channel quality, 70 to 90 indicates good, 50 to 70 indicates average, 30 to 50 indicates poor, and below 30 indicates very poor. By analyzing real-time indicators such as signal received power, bit error rate, and transmission delay, the instantaneous channel quality score is comprehensively calculated to reflect the true state of the current transmission environment.

[0298] The channel quality prediction level for the specified time period is then derived from the prediction results and converted into corresponding numerical scores: Level 1 corresponds to 95 points, Level 2 to 85 points, Level 3 to 70 points, Level 4 to 50 points, and Level 5 to 30 points. This conversion allows the prediction results to be compared and calculated with the current quality score on the same scale.

[0299] Data priority factors are quantified based on the importance level of the data to be transmitted. Level 1 urgent data corresponds to a priority score of 100 points, Level 2 highly important data corresponds to 80 points, Level 3 general data corresponds to 60 points, and Level 4 low-importance data corresponds to 40 points. This scoring system ensures that high-priority data receives a higher weight in the selection of transmission strategies.

[0300] The resource consumption evaluation of candidate transmission strategies covers multiple dimensions, including transmit power consumption, processor computational load, storage space usage, and network bandwidth requirements. A resource consumption model is established for each strategy to quantify the resource demand intensity of different strategies. Strategies with lower resource consumption receive higher scores, reflecting the system's consideration of optimizing resource utilization.

[0301] The weighting coefficients reflect the relative importance of different factors in transmission decisions. The current channel quality weight is typically set to 0.3, emphasizing the direct impact of the current environment; the predicted channel quality weight is set to 0.25, reflecting forward-looking considerations of future trends; the data priority weight is set to 0.35, highlighting the decisive role of data importance; and the policy resource consumption weight is set to 0.1, balancing resource utilization efficiency. These weights can be dynamically adjusted based on actual application needs and system operation experience.

[0302] The overall score is calculated using a weighted summation formula, which multiplies each scoring factor by its corresponding weight and then sums the results to obtain the final score for each candidate transmission strategy. Strategies with higher scores are considered more suitable for the current context, and the system will prioritize high-scoring strategies as the final transmission scheme.

[0303] It should be noted that, based on the channel quality prediction level and policy scoring results, an intelligent decision-making algorithm is used to determine the optimal transmission policy. First, two key thresholds are set: the excellent threshold is set to level 2 (corresponding to a score of 85 points), and the poor threshold is set to level 4 (corresponding to a score of 50 points). These two thresholds divide the channel quality into three intervals, corresponding to different transmission policy modes.

[0304] When the predicted channel quality level is greater than the excellent threshold, i.e., the predicted channel quality is level 1 or 2 (excellent or good), the normal transmission mode is adopted. In this mode, all priority data frames are sent normally according to the preset transmission plan, making full use of the good channel conditions and maximizing data transmission throughput. In normal transmission mode, level 1 data is sent immediately, level 2 data follows immediately after level 1 data is sent, and level 3 and level 4 data are transmitted according to standard timing arrangements to ensure that all data can be delivered to the master station in a timely manner.

[0305] When the channel quality prediction level falls between the excellent and poor thresholds (i.e., predicted as level 3 (general), the system enters a selective transmission mode. In this mode, the normal transmission frequency of high-priority data is maintained, but the transmission frequency of low-priority data frames is reduced. Specifically, level 1 and 2 data continue to be transmitted normally, the transmission frequency of level 3 data is reduced to 60% to 70% of its original frequency, and the transmission frequency of level 4 data is reduced to 40% to 50% of its original frequency. This strategy ensures timely transmission of critical data while avoiding excessive consumption of communication resources under generally poor channel conditions, thus reducing the risk of transmission conflicts and failures.

[0306] When the predicted channel quality level is below the poor threshold, i.e., predicted as level 4 or 5 (poor or very poor), the system activates a protective transmission mode. In this mode, transmission of data frames in the low-priority range is suspended, and limited communication resources are concentrated entirely on the transmission of high-priority data. Level 4 data transmission is completely suspended, and Level 3 data is also temporarily suspended; only Level 1 and Level 2 data continue to be transmitted. For Level 1 urgent data, enhanced transmission mechanisms are also activated, including increasing the number of retransmissions, increasing transmission power, and adopting more reliable coding methods to ensure that critical information can be successfully transmitted under adverse channel conditions.

[0307] This embodiment also establishes a dynamic strategy adjustment mechanism, which continuously optimizes the accuracy of the transmission strategy based on the comparison between actual transmission results and prediction results. If the actual channel quality is significantly better or worse than the prediction results, the current transmission mode is adjusted in real time to avoid resource waste or transmission delays caused by prediction deviations. Simultaneously, detailed data for each strategy execution is recorded, including indicators such as transmission success rate, latency, and resource consumption, to continuously improve the performance of the prediction model and strategy algorithm.

[0308] In one embodiment, sending the data frame to the master station according to the transmission strategy specifically includes:

[0309] Monitor the BeiDou transmission status, which includes transmission success rate, number of consecutive failures, average latency and backup link availability, 4G signal strength, network latency and available bandwidth;

[0310] When the number of consecutive failures exceeds a preset failure threshold, the transmission success rate drops below a preset success threshold within a specified time, or the channel quality prediction level is lower than a preset quality threshold, at least one available backup link is determined based on the data priority, timeliness requirements, current status of each backup link, energy consumption, and communication cost, and the data frame with the first priority is switched to the backup link.

[0311] When no available backup link is detected, the corresponding transmission parameters are adjusted according to the preset optimal transmission path configuration, and the data frame is transmitted according to the transmission parameters. The transmission parameters include the transmit power parameter that the substation can adjust under non-BeiDou communication link and the application layer length of the data frame generated according to the compressed data volume.

[0312] In the multi-link adaptive communication mechanism of this invention, in addition to data transmission via the BeiDou short message link, the substation is also configured with backup communication links, including 4G public mobile communication links, LoRa links, and microwave links. These backup links typically allow the terminal side to dynamically adjust the transmission power of the wireless transmission module based on network conditions, energy consumption strategies, or transmission priorities in terms of hardware structure and communication protocols. For example, the 4G LTE communication module supports configuring the transmit power level (TX Power Level) via AT commands or a driver interface to reduce the retransmission probability under weak signal conditions or reduce power consumption when channel quality is good.

[0313] Therefore, the "transmission power parameter" mentioned in this invention does not refer to the physical layer transmission power of the BeiDou short message satellite link, but specifically refers to the transmission power control quantity that can be adjusted by the communication module on the substation side in the backup link. This is used to meet the energy consumption control, reliability assurance, and collision avoidance requirements of data with different priorities under different links. When the BeiDou link cannot meet the current data transmission requirements and the backup link is available, the system can dynamically adjust the above-mentioned transmission power parameter based on the current channel conditions, data priority, and energy consumption strategy to achieve reliable transmission.

[0314] In the data frame construction process of this invention, the "data frame length" refers to the application layer data frame length generated by the substation based on the size of the compressed power data, priority tags, and frame structure rules. This data frame length differs from the fixed frame length of BeiDou short messages in the satellite communication protocol; it is the data encapsulation length within the terminal, used to form a complete structured message before entering the BeiDou short message payload or before transmission on the backup link.

[0315] The length of an application layer data frame is determined by multiple factors, including the size of the compressed data, the data type, whether it contains context information, and whether it contains redundant information. For example, for high-priority data with a high reliability score, this invention can package a single compressed data into an application layer frame; while for low-priority data or data that allows for a certain degree of error, multiple data can be spliced ​​together to form a longer data frame, thereby improving the overall link utilization.

[0316] Therefore, the "data frame application layer length" in the claim is entirely a data structure that can be defined and adjusted at the terminal application layer, does not involve the physical layer frame structure of the BeiDou satellite link, and does not conflict with the fixed frame mechanism of the BeiDou short message protocol.

[0317] The aforementioned design, based on adjustable transmission power parameters of the backup link and variable data frame length at the application layer, enables this invention to achieve an intelligent communication scheduling strategy that balances reliability and energy consumption optimization in complex power field environments (such as mountainous terrain, severe weather, and link fading). When the BeiDou link is temporarily unable to meet communication needs, the system can utilize the power adjustment capability of the backup link to improve the probability of reliable transmission of critical data. Simultaneously, by dynamically adjusting the application layer data frame length, it can flexibly adapt to the data carrying capacity of BeiDou and the backup link under different bandwidth conditions, thereby improving overall communication efficiency and achieving optimal scheduling of data resources.

[0318] In this embodiment, a comprehensive BeiDou transmission status monitoring system was established. Through real-time tracking and analysis of multi-dimensional indicators, the operational status of the communication system was fully understood. Transmission success rate monitoring employed a sliding window statistical method, calculating the percentage of successfully transmitted frames in the most recent 100 transmission attempts and updating the success rate value in real time. An alarm mechanism was triggered when the success rate fell below 85%, and a link switching evaluation procedure was initiated when it fell below 70%. Furthermore, the transmission success rates of data with different priorities were also statistically analyzed to provide a basis for subsequent differentiated processing strategies.

[0319] The consecutive failure count monitoring uses a real-time counter to track the current number of consecutive transmission failures. The counter is reset to zero and restarts after each successful transmission. Different failure thresholds trigger different levels of response measures: a retransmission mechanism is activated after 3 consecutive failures, link quality is evaluated after 5 consecutive failures, and a backup link switching procedure is forcibly initiated after 8 consecutive failures. For Level 1 emergency data, the failure threshold is set more strictly, and backup link evaluation begins after 2 consecutive failures.

[0320] Average latency monitoring employs an exponentially weighted moving average algorithm to statistically analyze recent transmission latency data, reflecting both current latency levels and maintaining sensitivity to historical trends. It records the complete latency from data frame transmission to reception acknowledgment, including signal propagation time, satellite processing time, and ground station forwarding time. Under normal circumstances, the average latency of BeiDou short messages should be controlled within 20 seconds. When the average latency exceeds 30 seconds, the system begins to assess the reasons for link performance degradation.

[0321] Backup link availability monitoring covers the status assessment of multiple alternative communication methods, including 4G cellular networks, WiFi networks, and wired Ethernet. Probe signals are periodically sent to test key parameters such as connectivity, bandwidth capacity, and signal strength of each backup link, establishing a real-time status database for backup links. 4G signal strength monitoring assesses cellular network availability by receiving signal strength indicators and metrics such as signal-to-interference-plus-noise ratio (SNR). Signal strength below -100 dBm is marked as unavailable.

[0322] Network latency monitoring assesses the latency characteristics of each backup link through periodic ping tests and packet round-trip time measurements. Network latency to different target servers, including the main server, regional data centers, and cloud platform access points, is tested to establish latency characteristic profiles. Available bandwidth monitoring employs a combination of active and passive methods. Active probing evaluates the maximum available bandwidth of the link by sending test data streams, while passive monitoring assesses bandwidth utilization by analyzing trends in actual data transmission throughput.

[0323] Subsequently, an intelligent link switching strategy based on multi-factor comprehensive evaluation is implemented, automatically selecting the optimal backup transmission path when the performance of the BeiDou main link degrades. A consecutive failure threshold of 8 times, a success rate threshold of 70%, and a quality threshold of level 4 (poor) are set. The link switching evaluation procedure is triggered when any of these conditions are met.

[0324] The switchover decision-making process first stratifies data based on priority. Level 1 urgent data enjoys the highest switchover priority, receiving the best backup link resources. Level 2 highly important data follows, while Level 3 and 4 data may need to wait or use lower-quality backup links when resources are scarce. Timeliness requirements assessment determines the urgency of the switchover based on the real-time requirements of different data types. For example, protection action information needs to be transmitted within one minute, while statistical data can tolerate delays of several hours.

[0325] The current status assessment of each backup link employs a multi-dimensional scoring system, including a comprehensive evaluation of indicators such as signal strength, network latency, available bandwidth, and historical reliability. The assessment of 4G networks focuses on signal coverage quality and data traffic cost, while the assessment of WiFi networks focuses on connection stability and security, and the assessment of wired networks focuses on bandwidth capacity and physical connection reliability. A dynamic scoring model is established for each backup link, updating the availability score of each link in real time.

[0326] Energy consumption assessment considers the power consumption differences of different transmission methods. BeiDou communication has relatively low power consumption but limited transmission capacity, 4G communication has higher power consumption but stronger transmission capabilities, and wired communication has the lowest power consumption but requires a physical connection. Power consumption models for various transmission methods were established, and energy efficiency was taken as an important factor when selecting backup links, especially for remote sites powered by batteries.

[0327] Communication cost assessment encompasses both direct communication expenses and indirect operation and maintenance costs. 4G data transmission involves traffic fees, satellite communication involves satellite leasing fees, and leased line communication involves line leasing fees. A cost model is established to prioritize the backup link with the best cost-effectiveness while ensuring transmission requirements are met.

[0328] The link switching process employs seamless switching technology. First, a communication connection is established on the backup link, and test transmissions are performed. Once the backup link is confirmed to be functioning correctly, the Level 1 data stream is switched over. During the switching process, data transmission continuity is maintained to prevent data loss or duplicate transmissions due to the switching operation. After the switching is complete, the status of the primary link is continuously monitored. When the primary link returns to normal, it automatically switches back to the primary link to conserve backup link resources.

[0329] Furthermore, when all backup links are detected to be unavailable, a transmission parameter optimization mechanism is activated to improve the transmission success rate by adjusting the technical parameters of BeiDou communication. This is based on a preset optimal transmission path configuration database, which contains the best parameter combinations under different channel conditions, derived through historical data analysis and simulation optimization.

[0330] It should be noted that transmit power adjustment is one of the key parameter optimization methods. The transmit power level is dynamically adjusted based on the current channel quality assessment results. Standard power is used when channel quality is good to save energy, while transmit power is gradually increased when channel quality deteriorates to enhance signal penetration. The adjustment range of transmit power is typically between 80% and 150% of the standard power. Too low a power level cannot guarantee transmission quality, while too high a power level will waste energy and may cause interference problems.

[0331] Data frame length adjustment is based on statistical analysis of channel error rate and transmission success rate. Shortening the data frame length can improve the success probability of single-frame transmission when channel conditions are poor. A model relating frame length to transmission success rate was established. The standard frame length was shortened to 60% of its original length when the channel quality was level 5 (very poor), and to 80% when it was level 4 (relatively poor). Although shorter frame lengths increase the number of transmissions, they significantly improve the overall transmission success rate.

[0332] In addition, other key transmission parameters were adjusted, including the number of retransmissions, retransmission interval, and coding redundancy. The number of retransmissions was dynamically set based on data priority and timeliness requirements: Level 1 data was retransmitted a maximum of 10 times, Level 2 data a maximum of 5 times, and Level 3 and 4 data a maximum of 3 times. The retransmission interval adopted an exponential backoff algorithm, with the initial retransmission interval being 5 seconds, and subsequent retransmission intervals gradually increasing to a maximum of 30 seconds to avoid channel congestion caused by frequent retransmissions.

[0333] Coding redundancy adjustment improves the anti-interference capability of data packets by increasing the redundancy information of forward error correction coding. Appropriately increasing the redundancy coding ratio when channel quality deteriorates, although increasing data volume, can significantly improve transmission reliability. Redundancy is dynamically adjusted according to the bit error rate (BER). Standard redundancy is used when the BER is below one-thousandth, and the redundancy is increased to 150% of the original when the BER exceeds one percent.

[0334] Optimal transmission path configuration also includes intelligent selection of transmission timing. This involves analyzing the geometric distribution and visibility windows of BeiDou satellites to select the time period with the best satellite geometry for data transmission. Through satellite orbit prediction algorithms, the optimal transmission window for the next few hours can be calculated in advance, allowing for the rational scheduling of non-urgent data transmission and maximizing the success rate.

[0335] Step 6: The master station receives data frames from the sub-station, performs integrity verification, data decompression and reassembly on the data frames, and feeds back the communication status to the sub-station, so that the sub-station updates the database, the data priority classification model, the channel quality prediction model and the intelligent compression model according to the communication status.

[0336] Specifically, in one embodiment, step 6 includes:

[0337] The system receives data frames from the substation and verifies the data integrity of the data frames. It generates a retransmission request based on the corrupted data frames and returns it to the substation until the data integrity verification is passed. The retransmission request is used to instruct the substation to transmit the corrupted data frames.

[0338] Using a decompression strategy that matches the compression strategy adopted by the substation, the verified data frames are decompressed and restored. The split frames are reassembled using sequence number information to restore the complete dataset. The original data is restored by reverse calculation for the differentially encoded time-series data to obtain the complete restored data.

[0339] Based on the restored data, the data transmission quality of each substation is evaluated, and a communication status is generated and fed back to the corresponding substation. This allows the substation to update the database based on the communication status, and to update the data priority classification model, the channel quality prediction model, and the intelligent compression model based on the updated database. The transmission evaluation results include frame loss rate, latency, and data quality evaluation.

[0340] In this embodiment, the master station establishes a robust data frame reception and integrity verification system to ensure the accuracy and reliability of power data transmitted from each substation. The master station first monitors data transmission from each substation via the BeiDou short message receiving module. Each received data frame carries a unique frame identifier, containing key information such as the sending substation's identifier, frame sequence number, and timestamp, facilitating frame management and tracking by the system.

[0341] Data integrity verification employs a multi-layered verification mechanism. First, a frame structure integrity check is performed, verifying that the format of the frame header, data payload, and frame trailer conforms to predefined protocol specifications. The frame length is checked to ensure it matches the length declared in the frame header, and the position and format of each field are correct, ensuring that no structural damage has occurred during data frame transmission. Next, data integrity verification is performed using a Cyclic Redundancy Check (CRC) code to verify the entire data frame content. The calculation result is compared with the checksum at the frame trailer; any mismatch indicates that an error has occurred during data transmission.

[0342] For data frames using forward error correction coding, an error correction algorithm is run to attempt to repair detected errors. When the number of error bits is within the error correction capability, error correction is automatically performed and the frame is marked as "corrected." When the error exceeds the error correction capability, the frame is marked as "uncorrectable," and a retransmission request mechanism is triggered. In addition, a logical integrity check is performed to verify whether the data content conforms to the physical laws and value range of the power system, such as whether the voltage value is within a reasonable range and whether the power value matches the voltage and current values.

[0343] When a corrupted data frame is detected, a standardized retransmission request message is immediately generated. This message contains detailed information such as the identifier of the corrupted frame, a description of the corruption type, and the specific content requested for retransmission. The retransmission request is sent back to the substation using a high-priority transmission channel to ensure timely notification of data retransmission. A retransmission request tracking mechanism is established to record the sending time and response status of each retransmission request. If the substation fails to respond to the retransmission request within a specified time, the system will send a reminder message or contact the substation using a backup communication channel.

[0344] The retransmission verification process continues until the data integrity verification is fully passed. A maximum retransmission limit is set for each data frame, typically 5 times. If this limit is exceeded, the frame is marked as "transmission failed" and recorded in the exception log. For Level 1 urgent data, a special expedited retransmission mechanism is used to shorten the retransmission waiting time and increase the retransmission limit, ensuring that critical data can be successfully transmitted. In addition, the reasons for retransmission failures are analyzed to distinguish whether they are due to channel quality issues, substation equipment failures, or problems with the data itself, providing a basis for subsequent system optimization.

[0345] Furthermore, the master station implements a decompression process that fully matches the compression strategy of the substations to ensure accurate restoration of the original power data. First, the compression strategy type used for each data frame is identified. This information, including the compression algorithm type, compression parameters, and compression ratio, is stored in the compression identifier field of the frame header. The master station maintains a complete decompression strategy library, synchronized with the substations' compression strategy libraries, ensuring that each compression method has a corresponding decompression algorithm.

[0346] The decompression process employs different decompression algorithms depending on the compression strategy. For data frames using lossless compression, such as those compressed using Huffman coding or LZ series algorithms, the corresponding decompression algorithm completely restores the original data without any information loss. For data frames using lossy compression, the decompression process references the quantization parameters and precision settings used during compression to restore the data as accurately as possible, and marks the degree of precision loss in the restored data.

[0347] Frame reassembly is a crucial step in the decompression process. When a large dataset is split into multiple data frames for transmission, accurate reassembly based on frame sequence numbers is necessary. A frame reassembly buffer is established to temporarily store individual frames belonging to the same dataset. Once all frames are collected, they are reassembled in sequence. During reassembly, the system checks the continuity and completeness of the frames. If a missing frame is found, a targeted retransmission request is sent, requiring the substation to retransmit the specific frame.

[0348] For differentially encoded time-series data, a reverse calculation and reconstruction mechanism is implemented to gradually reconstruct the complete time-series data sequence based on the reference point data and differential information. Reference point information for each data type at each substation is maintained, including the most recent complete data value and its corresponding timestamp. When differentially encoded data is received, the differential value is added to the corresponding reference value to obtain the actual value at the current moment. Then, the reference point information is updated to the current value to prepare for the next differential decoding.

[0349] During the time-series data restoration process, data continuity checks are also performed to verify the continuity of timestamps and the rationality of data changes. If timestamp jumps or abnormally drastic data changes are found, these data are marked as "suspicious data" and handled specially. For differentially encoded data of long-term series, complete reference point data is received periodically to prevent error accumulation. When the reference point data is inconsistent with the accumulated calculation results, the subsequent differential decoding process is recalibrated based on the reference point data.

[0350] After decompression and reassembly are complete, a final integrity and consistency check is performed on the restored data to ensure that no errors or loss occurred during the entire process. The restored data is tagged with metadata such as decompression timestamp, data source, and decompression method, and then stored in the main site's database for subsequent analysis and application.

[0351] Furthermore, the main station comprehensively evaluates the data transmission quality of each substation based on the restored data and generates detailed communication status feedback information, driving continuous optimization and improvement of the entire system. The transmission quality evaluation adopts a multi-dimensional indicator system to comprehensively reflect all aspects of the data transmission process.

[0352] Specifically, the frame loss rate assessment is calculated by statistically analyzing the data frame loss within a specific time window. This involves recording the total number of data frames sent and the number of successfully received frames for each substation, and then calculating the frame loss rate percentage. The assessment process differentiates between frame loss rates for data of different priorities: Level 1 data requires a frame loss rate below 1%, Level 2 data below 3%, and Level 3 and 4 data can tolerate a frame loss rate below 5%. Furthermore, the temporal distribution characteristics of the lost frames are analyzed to identify whether the loss is random or continuous. Continuous frame loss typically indicates a systemic problem requiring close attention.

[0353] Latency assessment covers the complete end-to-end transmission latency, from the generation of data at the substation to the successful reception and decompression by the master station. Transmission latency for data of different priorities is statistically analyzed, calculating average latency, maximum latency, and latency variance. The average latency for Level 1 urgent data must be controlled within 30 seconds, for Level 2 data within 2 minutes, and for Level 3 and 4 data, longer latency is acceptable. Latency assessment also includes latency stability analysis; excessive latency fluctuations indicate an unstable transmission path requiring optimization.

[0354] Data quality assessment measures the quality level of received data from multiple perspectives, including data completeness, accuracy, and consistency. Completeness assessment checks the integrity of data fields, identifying any missing fields or truncated data. Accuracy assessment identifies obviously abnormal or unreasonable data values ​​through comparative analysis with historical data. Consistency assessment examines the logical relationships between different parameters at the same time, such as whether three-phase voltages are balanced and whether the calculated relationship between power and voltage / current is correct.

[0355] Communication status feedback information is generated in a structured format, including detailed results of transmission quality assessment, analysis of identified problems, and improvement suggestions. Feedback information is sent to the corresponding substations through multiple channels, including BeiDou short message service, 4G network, and WiFi network, ensuring timely receipt of feedback information by the substations. The priority of feedback information is determined according to the severity of the problem; urgent issues are sent immediately through high-priority channels, while general feedback information is sent periodically through regular channels.

[0356] After receiving communication status feedback, the substation updates the transmission performance statistics in its local database based on the feedback, including key indicators such as success rate, latency, and quality level. This updated statistical information serves as new training data for retraining and optimizing the data priority classification model, channel quality prediction model, and intelligent compression model. Model updates employ incremental learning, incorporating new empirical data while maintaining existing knowledge, gradually improving the model's accuracy and adaptability.

[0357] Furthermore, this application establishes a transmission quality trend analysis mechanism. By using long-term quality monitoring data, it identifies trends in system performance and predicts potential problems and optimization opportunities. When a substation's transmission quality is found to be continuously declining, a diagnostic request is proactively sent to assist the substation in troubleshooting and performance optimization. This feedback-based continuous improvement mechanism ensures that the entire system can continuously self-optimize during operation, improving the reliability and efficiency of data transmission.

[0358] Step 7: The main station monitors the operating status of itself, the substation and the edge computing unit, provides early warning of abnormal states based on the pre-trained anomaly detection model, and triggers corresponding self-healing strategies. The abnormal states include data anomalies, communication anomalies and device anomalies.

[0359] Specifically, in one embodiment, step 7 includes:

[0360] An anomaly monitoring framework is established, wherein the anomaly monitoring framework includes data anomalies, communication anomalies, and equipment anomalies.

[0361] When the collected power data is detected to exceed the preset normal range, a data anomaly signal is issued;

[0362] When a BeiDou signal terminal is detected and / or the bandwidth of the backup link is lower than a preset bandwidth threshold, a communication abnormality signal is issued;

[0363] When the processor load of the substation exceeds a preset load threshold and / or the storage space of the substation is lower than a preset space threshold, a device abnormality signal is issued.

[0364] Based on the data anomaly signal, the communication anomaly signal, and the device anomaly signal, a corresponding anomaly feature vector is generated. The anomaly feature vector includes anomaly type, anomaly source, anomaly index, time characteristics, correlation information, and severity.

[0365] Using a pre-established reconstruction error model based on an autoencoder, a time-series prediction model based on LSTM, a correlation analysis model based on a graph neural network, and a multi-feature classification model based on XGBoost, the abnormal feature vectors are classified for anomaly type, assessed for severity, and analyzed for root causes, outputting anomaly classification results and severity levels.

[0366] Based on the anomaly classification results and the severity level, a multi-level early warning mechanism is constructed, and early warnings are issued based on the early warning mechanism.

[0367] In this embodiment, a comprehensive anomaly monitoring framework is established. This framework adopts a hierarchical and categorized monitoring architecture to achieve intelligent monitoring of the entire lifecycle of the BeiDou power data acquisition system. The anomaly monitoring framework is built upon three core anomaly categories, each containing multiple subdivided monitoring dimensions and indicators. Data anomaly monitoring focuses on the rationality and consistency of power parameters, including subcategories such as measured value anomalies, rate of change anomalies, and logical relationship anomalies. Communication anomaly monitoring targets the status and performance of various communication links, covering anomalies in different communication methods such as the BeiDou main link, 4G backup link, and WiFi link. Equipment anomaly monitoring focuses on the operational status of substation hardware and software systems, including the health status of key components such as processors, memory, sensors, and communication modules.

[0368] The framework employs a monitoring strategy that combines real-time monitoring with periodic evaluation. Key indicators such as the transmission status of Level 1 data and device CPU temperature are monitored in real-time at a frequency down to the second. General indicators such as storage space utilization and network traffic statistics are monitored periodically, with monitoring cycles ranging from minutes to hours. The system establishes an anomaly threshold management mechanism, setting dynamic anomaly thresholds for each monitoring indicator based on factors such as different device types, operating environments, and historical statistical data. These thresholds are adaptively adjusted according to seasonal changes, equipment aging, and environmental conditions.

[0369] The framework also integrates multi-source data fusion technology, correlating and analyzing information from different monitoring points to avoid the limitations of single-point monitoring. For example, when a voltage anomaly is detected, it simultaneously analyzes related parameters such as current, power, and frequency, as well as influencing factors such as ambient temperature and equipment load, to form a comprehensive anomaly judgment. The monitoring framework has tiered response capabilities, triggering different levels of response measures based on the severity and scope of the anomaly, ranging from simple log recording to emergency manual intervention notifications.

[0370] Subsequently, a data anomaly detection mechanism based on multi-dimensional rules and intelligent algorithms is implemented to ensure timely detection of various anomalies in power data. Data anomaly detection first establishes normal range benchmarks for power parameters. These benchmarks are determined comprehensively based on factors such as equipment rated parameters, historical operating data, and industry standards. For voltage parameters, the normal range is typically set at ±10% of the rated voltage; exceeding this range is considered abnormal. For current parameters, the normal range is dynamically adjusted based on the equipment's rated current and actual load. For power parameters, the system checks whether the calculated relationship between power values ​​and voltage / current is reasonable.

[0371] Furthermore, this application implements a dynamic threshold adjustment mechanism, which adjusts the anomaly detection threshold in real time based on factors such as load changes, seasonal factors, and equipment status. For example, during the high temperatures of summer, the operating temperature threshold of the transformer will be increased accordingly; during peak load periods, the upper limit of the normal current range will be appropriately widened. The dynamic threshold adjustment, based on statistical analysis of historical data and machine learning algorithms, can adapt to different operating conditions.

[0372] Data anomaly detection includes the identification of various anomaly patterns. Point anomaly detection identifies outliers in a single data point, such as sudden voltage jumps or abnormal current spikes. Sequence anomaly detection identifies anomaly patterns in time series data, such as sustained voltage drops or abnormal power fluctuations. Correlation anomaly detection identifies abnormal relationships between multiple parameters; for example, if the voltage is normal but the power is abnormally low, it may indicate a problem with the power factor.

[0373] When a data anomaly is detected, a standardized data anomaly signal is immediately generated. This signal contains detailed information such as the anomaly parameter name, anomaly value, normal range, anomaly time, and anomaly severity. The anomaly signal also includes contextual information, such as the equipment operating status, environmental conditions, and other relevant parameter values ​​at the time of the anomaly, providing comprehensive background information for subsequent anomaly analysis. A prioritization mechanism for anomaly signals is established, determining signal priority based on the type and severity of the anomaly. High-priority anomaly signals will immediately trigger alarm and response procedures.

[0374] Furthermore, an anomaly monitoring system covering all communication links was established to monitor the operational status of the BeiDou main link and various backup links in real time. BeiDou signal monitoring evaluates signal quality across multiple dimensions, including key indicators such as signal strength, signal-to-noise ratio, bit error rate, and transmission delay. When the BeiDou signal strength falls below a preset threshold (typically -130 dBm), it is considered an anomaly; multiple consecutive transmission failures or a bit error rate exceeding one in a thousand also trigger a communication anomaly signal.

[0375] The BeiDou signal terminal monitoring system employs a heartbeat detection mechanism, periodically sending detection signals to the BeiDou terminal to monitor its response time and quality. Under normal circumstances, the BeiDou terminal should respond to the detection signal within 10 seconds; failure to respond within this timeframe is considered a terminal malfunction. In addition, the system monitors the BeiDou terminal's operating status indicator lights, self-test reports, and other hardware status information to comprehensively assess the terminal's health.

[0376] Backup link monitoring covers the status monitoring of various communication methods, including 4G cellular networks, WiFi wireless networks, and wired Ethernet. 4G network monitoring focuses on indicators such as signal strength, network registration status, and data connection status. An abnormal signal is triggered when the 4G signal strength falls below a preset threshold (typically -100dBm) or the network connection is lost. WiFi network monitoring focuses on parameters such as wireless signal quality, connection stability, and network latency. An abnormal signal is issued when WiFi signal quality deteriorates or the connection frequently drops.

[0377] Bandwidth monitoring employs a combination of active and passive detection. Active detection measures the actual available bandwidth of each link by sending test data streams, while passive monitoring assesses bandwidth utilization by analyzing the transmission rate of actual service data. A communication anomaly signal is immediately issued when the available bandwidth of any backup link falls below a preset bandwidth threshold. The bandwidth threshold is set according to different link types and service requirements; typically, it is set at 100kbps for 4G networks, 1Mbps for WiFi networks, and 10Mbps for wired networks.

[0378] Communication anomaly signals include detailed descriptions of the anomaly, such as the type of abnormal link, specific abnormal parameters, duration of the anomaly, and affected data transmission. Furthermore, network environment information at the time of the anomaly is recorded, such as the currently connected base station, Wi-Fi hotspot, and network topology, providing a basis for anomaly cause analysis.

[0379] Furthermore, a comprehensive equipment health monitoring system has been established to monitor the operational status of key hardware and software components in substations in real time. Processor load monitoring obtains key indicators such as CPU utilization, memory utilization, and process status through system calls. Under normal circumstances, CPU utilization should be controlled below 70%, and memory utilization should be controlled below 80%. When the processor load continuously exceeds the preset load threshold (usually set to 85%), a device abnormality signal is issued, indicating that the processor may be overloaded or that abnormal processes are consuming resources.

[0380] Processor monitoring also includes temperature monitoring, which uses a built-in temperature sensor to monitor the operating temperature of the CPU and motherboard. The normal operating temperature should be kept below 70 degrees Celsius. When the temperature exceeds the safe threshold, an overheating alarm signal is issued, and automatic frequency reduction or shutdown protection measures may be triggered. In addition, the processor's operating frequency, voltage, and other parameters are monitored to ensure that the processor operates within its normal technical specifications.

[0381] Storage space monitoring covers the capacity usage of various storage devices, including the space occupancy rate of different partitions such as system disk, data disk, and journal disk. It sets multi-level storage space thresholds, issuing an alert when storage space utilization exceeds 80%, an abnormal signal when it exceeds 90%, and triggering an emergency cleanup mechanism when it exceeds 95%. Storage monitoring also includes disk health checks, using SMART technology to monitor hard drive health indicators such as the number of bad sectors, read / write error rate, and temperature, promptly identifying potential hard drive failure risks.

[0382] Equipment anomaly monitoring extends to other critical components, including the operational status of network interface cards, the connection status of serial communication modules, and the power supply quality of power modules. Network interface monitoring checks indicators such as network card connection status, traffic statistics, and error packet statistics; serial port monitoring checks the communication status and error rate of RS485 interfaces; and power supply monitoring checks the stability of input voltage and the operating temperature of power modules.

[0383] When a device malfunction is detected, a detailed malfunction signal is generated, including information such as the type of malfunctioning component, specific malfunction parameters, severity of the malfunction, and potential impact range. Furthermore, an initial fault diagnosis is attempted, such as distinguishing between hardware and software failures, and between temporary and persistent issues, providing reference information for subsequent fault handling.

[0384] Subsequently, abnormal signals from different monitoring modules are integrated and analyzed to construct structured abnormal feature vectors, providing standardized input data for subsequent intelligent analysis. The abnormal feature vectors employ a multi-dimensional information encoding method, with each dimension containing specific types of abnormal information, ensuring the completeness and analyzability of the abnormal information.

[0385] The anomaly type coding adopts a hierarchical classification system. The primary classification includes three main categories: data anomalies, communication anomalies, and equipment anomalies. The secondary classification further subdivides specific anomaly subtypes. Data anomalies include measurement value anomalies, rate of change anomalies, and logical relationship anomalies; communication anomalies include signal quality anomalies, connection interruption anomalies, and insufficient bandwidth anomalies; equipment anomalies include processor anomalies, storage anomalies, and interface anomalies. Each anomaly type is assigned a unique code identifier to facilitate unified processing and analysis by the system.

[0386] The anomaly source identifier records the specific location and equipment information where the anomaly occurred, including detailed information such as substation identifier, equipment type, component number, and sensor location. This precise anomaly source localization facilitates rapid troubleshooting and repair. The anomaly index quantifies and records specific numerical information about the anomaly, such as the measured value at the time of the anomaly, the degree of deviation from the normal range, and the duration of the anomaly, providing a data foundation for assessing the severity of the anomaly.

[0387] The time characteristics include information from multiple time dimensions: the anomaly occurrence time records the precise timestamp of the anomaly; the anomaly duration records the time span from the start to the end of the anomaly; and the anomaly occurrence frequency statistics show the frequency of similar anomalies in historical periods. The time characteristics also include periodic analysis, such as pattern information like whether anomalies occur in clusters during specific time periods or whether they are related to peak load periods.

[0388] Correlation analysis examines the relationships between anomalies and other factors, including concurrent anomalies, changes in the status of related equipment, and the impact of environmental factors. An anomaly correlation database is established to record patterns of association between different anomalies; for example, a certain equipment anomaly is often accompanied by specific data anomalies, or a certain communication anomaly may lead to a chain of equipment anomalies. Correlation information facilitates root cause analysis and preventative maintenance.

[0389] The severity assessment employs a multi-factor comprehensive scoring mechanism, considering factors such as the degree of impact of the anomaly on system functionality, the level of security risk, the difficulty of remediation, and the scope of impact. Severity is divided into five levels, with level 1 representing minor impact and level 5 representing severe impact. The scoring algorithm comprehensively considers both the direct impact and potential risks of the anomaly; for example, data anomalies may affect the accuracy of decision-making, communication anomalies may lead to monitoring blind spots, and equipment anomalies may trigger security incidents.

[0390] Optionally, this application employs four different artificial intelligence models for deep analysis of anomalous feature vectors, improving the accuracy and comprehensiveness of anomaly analysis through multi-model fusion. The autoencoder-based reconstruction error model is specifically designed to detect anomalous data patterns. This model learns the feature representations of normal data and calculates the reconstruction error of anomalous data; a larger reconstruction error indicates a more anomalous data. During the training phase, the autoencoder model uses a large amount of normal operating data for unsupervised learning, learning the data distribution characteristics and parameter correlations during normal power system operation.

[0391] During the anomaly detection phase, the model attempts to reconstruct the input anomalous feature vector. If the input data conforms to a normal pattern, the reconstruction error is small; if the input data contains anomaly patterns, the reconstruction error increases significantly. The system sets a reconstruction error threshold; data exceeding the threshold is considered anomalous. Autoencoder models are particularly suitable for detecting novel, previously unseen anomalous patterns and exhibit good generalization ability.

[0392] The LSTM-based time-series prediction model focuses on analyzing the temporal development patterns of anomalies and predicting future trends. This model is trained using historical anomaly data to learn the temporal patterns of anomaly occurrence, development, and regression. The Long Short-Term Memory (LSTM) mechanism of the LSTM network can capture long-term dependencies in the anomaly evolution process, predicting the development trend and potential impact range of anomalies. The time-series prediction model is of significant value for preventative maintenance and risk assessment, enabling early warnings of potentially worsening anomalies.

[0393] A graph neural network-based correlation analysis model is specifically designed to analyze the relationships and propagation paths between different anomalies. This model models each component and anomaly type in the power system as nodes in a graph network, and the physical connections and logical relationships between components as edges. When an anomaly occurs at a node, the model can analyze the propagation path of the anomaly's impact within the graph network and predict other nodes that may be affected. Graph neural network models are particularly effective for root cause analysis and impact assessment of systemic failures.

[0394] The XGBoost-based multi-feature classification model is responsible for accurate anomaly classification and severity assessment. This model uses a large amount of labeled historical anomaly data for supervised learning, learning feature patterns and severity determination rules for different anomaly types. The XGBoost algorithm has powerful feature selection and combination capabilities, automatically identifying the most important feature combinations for anomaly classification. The model output includes detailed information such as the specific anomaly category, confidence score, and severity level.

[0395] The fusion analysis of the four models employs an ensemble learning approach. Each model provides independent analysis results, which are then combined using techniques such as weighted voting or stacking to arrive at the final analytical conclusion. The fusion strategy considers the domain expertise and historical accuracy of each model, dynamically adjusting the weights of different models. The final output of anomaly classification results and severity levels is a comprehensive analysis of the four models, resulting in higher reliability and accuracy.

[0396] LSTM is a special recurrent neural network architecture specifically designed to address the vanishing and exploding gradient problems that traditional recurrent neural networks encounter when processing long sequences. By introducing gating mechanisms, including forget gates, input gates, and output gates, LSTM can selectively remember or forget historical information, thus effectively capturing long-term dependencies. In power data anomaly detection applications, LSTM can analyze the temporal patterns of anomalies, identify their development trends, and predict their possible evolution directions.

[0397] The process of constructing an LSTM model in power system anomaly monitoring includes key steps such as data preprocessing, network architecture design, and training optimization. In the data preprocessing stage, the system collects time-series data of historical anomaly events, including multi-dimensional information such as anomaly occurrence time, anomaly type, anomaly severity, and related equipment status, organizing this data into a time-series format. For example, for voltage anomaly events, the system collects relevant parameters such as voltage measurements, load changes, and ambient temperature over a period before and after the anomaly, forming a multivariate time-series dataset. Data standardization ensures that parameters with different dimensions can be analyzed on the same scale.

[0398] The LSTM network architecture is optimized based on the characteristics of power system anomalies. The input layer receives multi-dimensional temporal feature vectors, including information such as power parameters, equipment status, and environmental factors. The hidden layers employ multi-layer LSTM units. The first LSTM layer extracts basic temporal features, while subsequent layers progressively abstract higher-level temporal patterns. Within each LSTM unit, the forget gate determines which historical information needs to be discarded, the input gate controls the degree of integration of new information, and the output gate adjusts the intensity of the output information. The network also integrates an attention mechanism, enabling the model to automatically focus on the time periods and feature dimensions most important for anomaly prediction.

[0399] The model training process utilizes a large amount of labeled historical anomaly data, aiming to learn the temporal patterns of anomaly occurrence, development, and resolution. The loss function combines classification and regression losses, aiming to accurately predict both anomaly types and their severity variations. The training process employs a sliding time window approach, dividing long-term series into fixed-length training samples, each containing a historical time series segment and its corresponding future anomaly state. The backpropagation algorithm calculates gradients through time expansion, and gradient pruning prevents gradient explosion. After training, the model can predict potential future anomaly types and severity based on current and historical system states.

[0400] Graph Neural Networks (GNNs) are a class of deep learning models specifically designed for processing graph-structured data. They can simultaneously consider node features and graph topology information, learning node and edge representations through message passing and aggregation mechanisms. In power system anomaly analysis, GNNs model power equipment, measurement points, and communication nodes as nodes in a graph, and physical connections, electrical relationships, and communication links between equipment as edges, thereby enabling the analysis of anomaly propagation paths and impact ranges within the system.

[0401] The construction of a power system graph neural network first requires establishing a graph representation model of the system. Node types include different types of system components such as generating equipment, transformers, transmission lines, load nodes, and communication equipment. Each node has specific attributes, such as equipment capacity, operating status, geographical location, and importance. Edge types include electrical connection edges, communication connection edges, and control relationship edges, each with corresponding attributes such as impedance value, transmission capacity, and reliability level. The graph construction also considers the time dimension, forming a dynamic graph structure that can reflect the time-varying characteristics of the system topology and state.

[0402] The specific architecture of a graph neural network employs advanced models such as graph convolutional networks or graph attention networks. The network progressively aggregates neighborhood information through multiple layers of graph convolutional operations, with each layer updating the node's representation vector. The first layer of graph convolution collects information from directly adjacent nodes, and subsequent layers gradually expand the receptive field, ultimately fusing information from multi-hop neighborhoods into the representation of each node. The message passing mechanism determines the strength of information transmission based on the edge type and weight, with critical connections receiving higher weights. Aggregation functions fuse information from different neighbors; common aggregation methods include summation, averaging, and maximization.

[0403] The model training uses historical anomaly propagation cases as supervision signals. Training data includes information such as the initial location of the anomaly, its propagation path, the range of affected devices, and its propagation speed. The training objective is to learn the propagation patterns of anomalies in a graph structure, including which types of anomalies are prone to propagation, preferred propagation paths, and propagation attenuation patterns. The loss function is designed as a combination of link prediction loss and node classification loss, aiming to predict both whether an anomaly will propagate along a certain edge and the probability of each node being affected by the anomaly. The training process employs a mini-batch approach, with each batch containing multiple subgraph samples, using graph sampling techniques to handle large-scale graph data. After training, the model can quickly analyze the possible propagation paths and impact range of an anomaly when it is detected, providing decision support for emergency response.

[0404] XGBoost is an ensemble learning algorithm based on gradient boosting decision trees. It achieves strong predictive performance by sequentially constructing multiple weak learners and combining their predictions. XGBoost incorporates optimizations such as regularization, parallel computation, and missing value handling on top of traditional gradient boosting, making it particularly suitable for classification and regression problems involving structured data. In anomaly analysis applications, XGBoost is responsible for accurately classifying anomalous feature vectors and determining the specific type and severity level of the anomaly.

[0405] The XGBoost model construction process begins with feature engineering, transforming each dimension of the anomaly feature vector into a format suitable for decision tree processing. Numerical features, such as anomaly index values ​​and durations, are used directly, while categorical features, such as anomaly type and device type, are encoded. Temporal features are extracted into periodic features such as hours, days of the week, and seasons. The feature selection process uses methods such as correlation analysis and importance assessment to screen the most valuable feature subset for anomaly classification, removing redundant and noisy features to improve the model's generalization ability.

[0406] The model training utilizes a large number of historically labeled anomaly cases. Each training sample contains a complete anomaly feature vector and corresponding label information, including the anomaly's precise classification and severity level. XGBoost employs a boosting learning strategy, first training a base decision tree model, and then progressively adding new decision trees to correct the prediction errors of the previous model. Each training round calculates the prediction residuals of the current model, and the newly added decision trees are specifically used to fit these residuals, thereby continuously improving the overall prediction performance. The objective function includes a prediction loss and a regularization term. The prediction loss measures the model's fit, while the regularization term controls model complexity to prevent overfitting.

[0407] Key parameters during training include the learning rate, which controls the contribution of each tree; a smaller learning rate requires more trees but usually yields better performance. Tree depth controls the complexity of a single tree; too deep a tree can lead to overfitting, while too shallow a tree may result in underfitting. Subsample ratios and feature sampling ratios introduce randomness, improving the model's robustness. Regularization parameters control the complexity of leaf node weights, preventing the model from becoming overly complex. An early stopping mechanism is used during training; training stops when validation set performance no longer improves, avoiding overfitting.

[0408] Model optimization also includes hyperparameter tuning, using methods such as grid search, random search, or Bayesian optimization to find the optimal parameter combination. Cross-validation is used to evaluate the model's generalization performance, ensuring good classification accuracy even on unseen data. Feature importance analysis identifies the features that contribute most to anomaly classification, providing insights into anomaly mechanism analysis. The trained XGBoost model can quickly and accurately classify new anomaly feature vectors, outputting detailed information such as the specific anomaly type, confidence score, and severity level, providing a reliable basis for anomaly handling decisions.

[0409] It should be noted that a multi-layered intelligent early warning mechanism is constructed based on the anomaly analysis results, and differentiated early warning strategies and response measures are formulated according to factors such as the type, severity, and scope of impact of the anomaly. The multi-layered early warning mechanism adopts a five-level early warning system, corresponding to different response intensities and processing procedures.

[0410] Level 1 warnings correspond to minor anomalies, primarily deviations from normal ranges that do not affect the safe operation of the system, such as small fluctuations in certain measured values ​​or minor overheating of non-critical equipment. Level 1 warnings are logged, recording the anomaly information in the system log without triggering immediate notifications. They are checked and addressed during regular system inspections. These warnings are mainly used for long-term trend analysis and the development of preventative maintenance plans.

[0411] Level 2 alerts correspond to general anomalies, including situations that may affect system performance but do not currently threaten security, such as degraded communication quality, high storage space utilization, and heavy device load. Level 2 alerts trigger automated processing steps, such as cleaning up temporary files to free up storage space, adjusting data transmission strategies to reduce communication burden, and optimizing task scheduling to reduce processor load. Simultaneously, email or SMS notifications are sent to operations and maintenance personnel, but immediate action is not required.

[0412] Level 3 alerts correspond to critical anomalies, involving situations that may affect the normal functioning of the system, such as failure to transmit critical data, excessively high temperatures of important equipment, or communication link interruptions. A Level 3 alert will immediately notify maintenance personnel, requiring them to address the issue within a predetermined timeframe (usually within 2 hours). The system will automatically activate backup plans, such as switching to a backup communication link, initiating equipment cooling measures, and increasing the data retransmission frequency.

[0413] Level 4 alerts correspond to severe anomalies, involving situations that may jeopardize system security or cause significant impact, such as simultaneous failure of multiple critical devices, large-scale loss of important data, or malfunction of security protection devices. Level 4 alerts will immediately notify relevant personnel through various means, including telephone, SMS, email, and mobile application push notifications, requiring immediate response and handling. The system will automatically activate emergency response procedures, such as isolating faulty equipment, activating emergency communication plans, and preserving critical data.

[0414] Level 5 alerts correspond to emergency anomalies, involving extreme situations that may cause serious safety accidents or systemic failures, such as fires, floods, cybersecurity attacks, and large-scale equipment failures. A Level 5 alert will trigger the highest level of emergency response, immediately notifying all relevant personnel, including management, and activating emergency plans and disaster recovery procedures. The system will automatically execute emergency measures such as safe shutdown, data backup, and personnel evacuation.

[0415] The early warning mechanism also includes an intelligent upgrade function. When a low-level anomaly is not effectively handled within a certain period and shows a worsening trend, the early warning level is automatically upgraded to ensure that important issues receive timely attention. Simultaneously, the early warning mechanism has learning and optimization capabilities, continuously optimizing the early warning thresholds and response strategies based on the handling results and effects of historical early warnings, thereby improving the accuracy and effectiveness of early warnings.

[0416] In some embodiments, the main station triggers the self-healing strategy through the following steps, specifically including:

[0417] Establish a self-healing strategy library, which includes data anomaly self-healing strategies, communication anomaly self-healing strategies, and device anomaly self-healing strategies.

[0418] Based on the self-healing strategy library, according to the anomaly classification results, the severity level, system status, environmental conditions, historical experience, impact assessment, and resource constraints, the corresponding self-healing strategy is selected, and self-healing operation instructions are output according to the self-healing strategy.

[0419] The self-healing operation is performed according to the self-healing operation instructions, which include data marking repair, link parameter adjustment and switching, service restart, and configuration rollback.

[0420] In this embodiment, a comprehensive self-healing strategy library was established. This library is built based on actual operational experience and fault handling cases of the power data acquisition system, covering various possible anomalies and their corresponding automated processing solutions. The data anomaly self-healing strategy mainly addresses issues such as abnormal power parameter measurements, incorrect data logic relationships, missing or corrupted data, and other problems by developing corresponding repair solutions. These strategies include various technical means such as data smoothing and filtering, interpolation repair based on historical data, abnormal data isolation and re-acquisition, and multi-source data cross-validation, which can automatically identify and repair most common data quality problems.

[0421] Communication anomaly self-healing strategies focus on fault recovery and performance optimization of various communication links, including signal optimization of the BeiDou main link, intelligent switching of backup links, dynamic adjustment of network parameters, and adaptive selection of communication protocols. These strategies can automatically select the most suitable recovery scheme based on different types of communication anomalies, such as power adjustment strategies for signal quality degradation, flow control strategies for network congestion, and fast switching strategies for link interruptions. Communication self-healing strategies also include preventative measures that can proactively optimize and adjust when a trend of declining communication quality is detected, avoiding complete communication outages.

[0422] Device anomaly self-healing strategies encompass various fault handling solutions for hardware devices and software systems, including load balancing strategies for processor overload, cleanup and expansion strategies for insufficient storage space, heat dissipation and frequency reduction strategies for overheating, and restart and recovery strategies for software anomalies. These strategies place particular emphasis on system stability and security, prioritizing the protection of basic system functions and data security during self-healing operations to avoid causing more serious problems due to the self-healing process itself. Device self-healing strategies also include a gradual processing mechanism, prioritizing less impactful solutions and only resorting to more drastic measures when mild intervention proves ineffective.

[0423] Subsequently, a multi-dimensional factor analysis is used to select the most suitable self-healing strategy, ensuring the accuracy and effectiveness of the self-healing operation. The anomaly classification results provide basic information on the nature and type of the problem, and candidate solutions are selected from the corresponding strategy branches based on the specific anomaly category. For example, for data measurement anomalies, data repair strategies are prioritized; for communication interruption anomalies, link switching strategies are prioritized; for equipment overheating anomalies, heat dissipation and load adjustment strategies are prioritized. The severity level determines the intensity and urgency of the self-healing operation; minor anomalies may only require simple parameter adjustments, while severe anomalies may require strong intervention measures such as service restarts or device isolation.

[0424] Status assessment is a crucial basis for strategy selection, including information such as the current system load level, available resource status, and the operational status of critical services. Under high load, self-healing strategies avoid operations that might further increase system burden; when resources are scarce, they prioritize lightweight solutions with lower resource consumption. Environmental condition analysis includes current external environmental factors, such as temperature, humidity, and electromagnetic interference—parameters that may affect equipment operation—as well as operational environmental factors such as power grid load and communication quality. Harsh environmental conditions may limit the effectiveness of certain self-healing strategies, requiring adjustments to strategy selection accordingly.

[0425] The historical experience database records the handling results and effectiveness evaluations of similar past anomalies. Machine learning algorithms analyze this historical data to identify the most effective handling strategies under specific conditions. Historical experience also includes the analysis of failure cases to avoid repeating previously ineffective or negatively impactful solutions. Impact assessment analyzes the potential impact of anomalies on various aspects of the system, including data quality, communication performance, equipment security, and business continuity, selecting self-healing strategies that minimize negative impacts.

[0426] Resource constraint analysis considers the various resources currently available for self-healing operations, including computing resources, storage resources, network resources, and human resources. Some self-healing strategies may require significant computing resources or network bandwidth; under resource constraints, the system will choose alternatives with lower resource requirements. Furthermore, considering the time window constraint of self-healing operations, strategies that can be completed within the allowed time are prioritized.

[0427] Furthermore, based on the selected self-healing strategy, detailed self-healing operation instructions are generated. These instructions include all the technical parameters and execution steps required to perform the self-healing operation. Data labeling and repair instructions primarily target data quality issues, including marking outlier data, correcting erroneous data, and completing missing data. Corresponding repair instructions are generated based on the specific type of data anomaly; for example, statistical filtering algorithms are used for smoothing measurement anomalies, correlation checking and correction algorithms are used for logical relationship anomalies, and historical pattern-based interpolation algorithms are used for missing data. The data repair instructions also include verification steps for the repair effect to ensure that the repaired data meets quality requirements.

[0428] Link parameter adjustment and switching commands involve the configuration optimization and link management of the communication system, including operations such as adjusting BeiDou communication parameters, activating backup links, and reconfiguring network routes. When a decline in communication quality is detected, parameter adjustment commands are generated, such as increasing transmission power, adjusting data frame length, and modifying retransmission strategies. When the primary link completely fails, a link switching command is generated, automatically activating the optimal backup communication link and reconfiguring relevant parameters. These commands also include data protection measures during the switching process to ensure that critical data is not lost during link switching.

[0429] The service restart command is used to resolve software system anomalies and resource cleanup issues, including restarting specific service processes, clearing system caches, and releasing memory resources. The scope and method of restarting depend on the specific nature of the anomaly; minor software anomalies may only require restarting specific service processes, while severe system anomalies may require a full system restart. The restart command also includes data protection measures, automatically saving important data and configuration information before restarting and automatically restoring the system state and service configuration after restarting.

[0430] The configuration rollback command is used to restore the system to a previous stable state. When a recent configuration change is detected as potentially causing system anomalies, a rollback command is automatically generated to restore the relevant configuration to its state before the anomaly occurred. The configuration rollback function is based on the system's version control mechanism, automatically recording detailed information and timestamps of each configuration change, enabling precise rollback to a specified historical version. The rollback operation also includes compatibility checks to ensure that the rolled-back configuration is compatible with the current system environment and data state.

[0431] The self-healing operation is executed gradually and in a controlled manner, starting with operations that have less impact and observing their effects before deciding whether further intervention is needed. Each self-healing operation is equipped with a timeout mechanism and a rollback plan. If the operation does not achieve the expected results within the predetermined time or an unexpected situation occurs, the current operation is automatically stopped and alternative solutions are tried. The entire process of the self-healing operation is logged in detail, including the triggering reason, execution steps, parameter settings, and execution results, providing data support for subsequent effect evaluation and strategy optimization.

[0432] Application example of a power data intelligent acquisition and dynamic communication dispatch system in a prefecture-level city:

[0433] Deployment background and overall architecture:

[0434] Taking a prefecture-level city in a certain region as an example, the city administers six counties and districts. Its power system includes one 220kV main substation, 15 110kV substations, 120 35kV substations, as well as 85 small hydropower stations, 30 photovoltaic power stations, and 12 wind farms distributed in mountainous areas and remote rural villages. Due to the complex terrain, some power generation sites are located in deep mountain valleys or island areas, making coverage difficult with traditional communication methods. Data transmission has always been a problem that plagues power grid dispatching.

[0435] Adopting a "1+N" deployment architecture, the municipal power dispatch center serves as the main station, responsible for the aggregation, analysis, and intelligent decision-making of power data across the city. Substations are set up at each power generation site and important substation, equipped with edge computing units and BeiDou communication terminals. The main station deploys a high-performance server cluster, running AI algorithms such as data priority classification models, channel quality prediction models, and intelligent compression models. The substations utilize industrial-grade embedded computing platforms, integrating functional modules such as data acquisition, intelligent analysis, and communication dispatch.

[0436] Database construction and model training practices:

[0437] In the initial stage, the main station collected nearly three years of power operation data for the city, including approximately 5 million power measurement records, 2 million BeiDou communication records, and 1.5 million environmental condition data points. The data cleaning process identified and processed approximately 15% of the abnormal data, mainly including outliers caused by sensor malfunctions, data gaps due to communication interruptions, and human error. Feature engineering extracted more than 200 key feature variables, including load fluctuation characteristics, seasonal variation patterns, and equipment aging characteristics.

[0438] Based on the cleaned data, three core AI models were trained. The data priority classification model, based on a CNN-LSTM hybrid network architecture, was trained using 300,000 labeled samples covering various scenarios such as substation protection actions, line faults, equipment anomalies, and load surges. The model achieved a classification accuracy of 94.2% on the test set. The channel quality prediction model, employing a GRU network structure and trained using 18 months of BeiDou signal quality data, can predict channel quality changes 15 minutes in advance with an accuracy of 87.5%. The intelligent compression model family comprises 12 sub-models targeting different data types, achieving an average compression rate of 45% and keeping data restoration errors within 0.3%.

[0439] Edge data acquisition and intelligent processing:

[0440] Taking a small hydropower station located in a district of the city as an example, the station is equipped with a multi-protocol data acquisition unit. It connects to the governor, excitation device, and protection devices of the turbine generator unit via an RS-485 interface, and to newly installed smart energy meters and fault recorders via an Ethernet interface. The data acquisition frequency is set according to the importance of the parameters: key parameters such as generator voltage and current are collected every 30 seconds; important parameters such as switch status are collected every 2 minutes; and statistical parameters such as cumulative electricity generation are collected every 15 minutes.

[0441] The edge computing unit runs an improved EPUS algorithm, mapping the collected power data to a three-dimensional importance index space. Dimensions include data change rate, deviation from normal values, and predefined importance of data type. Taking a specific equipment malfunction as an example, where the generator stator current suddenly increased from the rated value of 95A to 110A within 5 minutes, its importance index was calculated as follows: change rate dimension score 0.85 (change magnitude 15.8%), deviation dimension score 0.92 (out of normal range), and type importance dimension score 0.88 (critical operating parameter). The overall importance score was 0.88, successfully placing it into the skyline candidate set.

[0442] The AI ​​classification model prioritizes the filtered data. The feature vector of the abnormal current data is input into the model, resulting in a priority level of 1 (urgent) with a confidence score of 0.94. Complete metadata is then added to the data, including device ID, timestamp, data type, geographical location, and associated device status, forming a structured data stream.

[0443] Intelligent compression and dynamic frame packing:

[0444] For this emergency data, a Gaussian mixture model (GMM) was used for reliability assessment. The GMM analysis showed that the probability density of the data belonging to the Level 1 emergency category was 0.91, while the probability density of other categories was all below 0.15. The intra-category reliability index was 0.91, the inter-category difference index was 0.76, the anomaly index was 0.12, and the overall reliability score was 0.89, which is higher than the high reliability threshold of 0.85.

[0445] Since the data has a priority level of 1 and a high reliability score, a lossless compression strategy with a low compression ratio was chosen, using an optimized Huffman coding algorithm with a compression ratio set at 22%. Meanwhile, considering the time-series correlation of the data, the difference between the current value and the previous value (110A-95A=15A) was calculated, and differential coding was used for further compression, ultimately reducing the data volume to 30% of its original size.

[0446] During the frame packaging stage, since this data has a priority of level 1, it is packaged into a separate data frame. Considering that the current BeiDou channel quality prediction level is level 3 (general), the standard frame length of 1835 bits is adjusted to 1200 bits to improve the transmission success rate. The data frame contains a complete frame header, compressed data, checksum, and metadata, using a total of 1156 bits, leaving appropriate margin.

[0447] Dynamic communication scheduling and link management:

[0448] The channel quality prediction model analyzed BeiDou signal data from the past 30 minutes, including parameters such as signal strength, signal-to-noise ratio, and transmission delay. It predicted that the channel quality would remain at level 3 for the next 15 minutes, with a confidence interval of 85%-95%, showing a stable trend. The context-aware intelligent handover protocol comprehensively evaluated the current channel quality (70 points), predicted quality (70 points), data priority (100 points), and transmission strategy resource consumption (85 points). After weighted calculation, a strategy score of 82 was obtained, leading to the decision to adopt the normal transmission mode.

[0449] During data frame transmission, the transmission status was monitored in real time. The BeiDou link transmission success rate was 92%, with zero consecutive failures and an average latency of 18 seconds, all within the normal range. Simultaneously, the 4G backup link signal strength was monitored at -85dBm, network latency at 120ms, and available bandwidth at 2.5Mbps, all in good condition. Since the main link was functioning normally, the system continued to use BeiDou for transmission.

[0450] Data reception, processing, and quality assessment:

[0451] The city dispatch center's main station successfully received the data frame, and the integrity check passed. The data was identified as employing a combined Huffman coding and differential coding compression strategy, and the corresponding decompression algorithm was executed. First, differential decoding was performed, adding the received difference of 15A to the reference value of 95A to obtain the actual current value of 110A. Then, Huffman decoding was performed to completely restore the original data structure. The consistency check between the decompressed data and the original data passed, with a restoration error of 0.

[0452] The main station assesses the transmission quality of the hydropower station. Statistics show that the station's frame loss rate was 2.1% over 24 hours, with an average transmission latency of 16.8 seconds, and the data quality is rated as excellent. Communication status feedback is generated and sent to the hydropower station's substation via the 4G network. Upon receiving the feedback, the substation updates its local database and incrementally learns and optimizes the AI ​​model based on the new performance data.

[0453] Anomaly monitoring and intelligent self-healing response:

[0454] In response to this abnormal current event, the anomaly monitoring framework immediately initiated the analysis process. The data anomaly detection module identified that the current value exceeded the normal range and generated a data anomaly signal; simultaneously, checks revealed that the synchronous voltage and power parameters were normal, leading to a preliminary judgment that it might be due to a sudden increase in load or equipment failure. An anomaly feature vector was constructed, including the anomaly type (current exceeding the limit), anomaly source (generator No. 1), anomaly index (15.8% exceeding the limit), time characteristics (high load period at 15:30), related information (normal water level, slight increase in vibration), and severity (moderate).

[0455] Four AI models collaborated to analyze the anomaly: the autoencoder model calculated a reconstruction error of 0.31, exceeding the threshold of 0.25, thus confirming it as an abnormal pattern; the LSTM model predicted that the anomaly might persist for 10-15 minutes before naturally subsiding; the graph neural network model analysis found that the anomaly only affected the generator and would not spread to other devices; the XGBoost model classified the anomaly as "load surge" with a severity level of 2.

[0456] This application issued a Level 2 early warning, automatically notifying maintenance personnel to monitor the equipment status. Simultaneously, a self-healing strategy was activated, automatically adjusting the generator's excitation parameters to control the current within a safe range and increasing the monitoring frequency of the unit. After 15 minutes, as the load decreased, the current value returned to normal, and the anomaly was automatically resolved. Throughout the entire process, the transmission and processing of critical data remained highly efficient and stable, fully demonstrating the system's intelligence and adaptive capabilities.

[0457] This comprehensive example demonstrates the complete workflow of the BeiDou power data intelligent acquisition and dynamic communication scheduling system in practical applications. From data acquisition, intelligent analysis, dynamic transmission to anomaly handling, each link works closely together to achieve intelligent and automated data transmission at remote power system sites.

[0458] See Figure 2 This invention also provides a BeiDou power data intelligent acquisition and dynamic communication scheduling system 200, comprising:

[0459] The database construction module 201 is used by the main station to perform data cleaning and feature engineering processing on the acquired historical power data, Beidou communication data, environmental and operating condition data, establish basic parameter configurations and write them into the database.

[0460] Model building module 202 is used by the main station to train a data priority classification model, a channel quality prediction model, and an intelligent compression model based on the database.

[0461] The data acquisition module 203 is used by the substation to collect power data through the edge computing unit, construct a skyline candidate set by combining the preset EPUS algorithm for data filtering, and use the data priority classification model from the main station to intelligently classify and label the power data after data filtering, and output a data stream with priority labels.

[0462] The data processing module 204 is used to perform a reliability assessment on the data stream using a preset Gaussian mixture model classification framework, obtain a reliability score, and select a compression strategy based on the reliability score, the data type identifier corresponding to the data stream, and the priority label, and package the data stream into a data frame according to the compression strategy.

[0463] The data frame sending module 205 is used by the substation to determine a transmission strategy based on the data frame and the channel quality prediction result, combined with a preset context-aware intelligent handover protocol, and to send the data frame to the master station according to the transmission strategy. The transmission strategy is used to indicate the transmission timing selection, multi-link intelligent handover, and parameter adaptive adjustment. The channel quality prediction result is obtained by the channel quality prediction model from the master station after predicting the current BeiDou signal strength.

[0464] Data restoration module 206 is used for the master station to receive data frames from the sub-station, perform integrity verification, data decompression and reassembly on the data frames, and feed back the communication status to the sub-station, so that the sub-station updates the database, the data priority classification model, the channel quality prediction model and the intelligent compression model according to the communication status;

[0465] The anomaly monitoring module 207 is used by the main station to monitor the operating status of itself, the substation and the edge computing unit, to provide early warning of abnormal states based on a pre-trained anomaly detection model, and to trigger corresponding self-healing strategies. The abnormal states include data anomalies, communication anomalies and device anomalies.

[0466] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0467] It should be noted that those skilled in the art can make various modifications and variations to this invention without departing from the spirit and scope of this invention. If such modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include such modifications and variations.

[0468] This disclosure also provides a computer-readable storage medium storing a computer program. When a processor executes the computer program, it performs the steps of the BeiDou power data intelligent acquisition and dynamic communication scheduling method described in the above-described method embodiments. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0469] Furthermore, this disclosure also provides a computer program product storing a computer program. When the computer program is run by a processor, it executes the steps of a Beidou power data intelligent acquisition and dynamic communication scheduling method provided in any of the above embodiments of this disclosure. For details, please refer to the above method embodiments, which will not be repeated here.

[0470] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium, which can be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0471] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and apparatuses described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0472] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0473] In addition, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0474] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0475] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for intelligent acquisition and dynamic communication scheduling of Beidou power data, characterized in that, The method comprises the following steps: The master station performs data cleaning and feature engineering processing on the obtained historical power data, Beidou communication data, environment and working condition data, establishes and writes basic parameter configurations into a database, wherein the basic parameter configurations include model triggering threshold, channel quality threshold, power data acquisition frequency default value, Beidou transmission default frequency level, backup link switching condition and feature engineering parameter; The master station trains a data priority classification model, a channel quality prediction model and an intelligent compression model according to the database; The substation collects power data through an edge computing unit, constructs a skyline candidate set by combining a preset edge-assisted parallel uncertain skyline (EPUS) algorithm to perform data screening, and uses the data priority classification model from the master station to intelligently classify and mark the power data after data screening, and outputs a data stream with a priority label; wherein the step of constructing a skyline candidate set by combining a preset edge-assisted parallel uncertain skyline (EPUS) algorithm to perform data screening comprises the following steps: regarding the power data as data points in a multi-dimensional space, constructing an importance index space, the dimension index of the importance index space including data change rate, deviation from rated value degree and pre-defined importance of data type; using a preset EPUS algorithm and the importance index space to calculate the importance index of each data point, and comparing the importance index of the data point with the data in the current skyline candidate set based on the importance index, and when the importance index of the data point exceeds the data in the current skyline candidate set, adding the data point to the skyline candidate set and updating; The substation uses a classification framework of a preset Gaussian mixture model to evaluate the reliability of the data stream, obtains a reliability score, and selects a compression strategy according to the reliability score, the data type identifier corresponding to the data stream and the priority label, and packs the data stream into a data frame according to the compression strategy; The substation determines a transmission strategy by combining a preset context-aware intelligent switching protocol according to the data frame and the channel quality prediction result, and sends the data frame to the master station according to the transmission strategy, wherein the transmission strategy is used to indicate transmission opportunity selection, multi-link intelligent switching and parameter adaptive adjustment, and the channel quality prediction result is obtained by predicting the current Beidou signal strength by the channel quality prediction model from the master station; The master station receives the data frame from the substation, performs integrity check, data decompression and reorganization on the data frame, and feeds back the communication state to the substation, so that the substation updates the database, the data priority classification model, the channel quality prediction model and the intelligent compression model according to the communication state; The master station monitors the running state of itself, the substation and the edge computing unit, performs early warning on abnormal state based on a pre-trained abnormality detection model, and triggers a corresponding self-healing strategy, wherein the abnormal state includes data abnormality, communication abnormality and device abnormality.

2. The method of claim 1, wherein, The substation collects power data through the edge computing unit, constructs a skyline candidate set for data screening in combination with a preset EPUS algorithm, and uses the data priority classification model from the master station to intelligently classify and mark the power data after screening, and outputs a data stream with priority labels, specifically including: The edge computing unit detects RS-485 and Ethernet interface connected to the electric meter and the measurement and control device to collect power data, records the metadata generated in the process of collecting the power data, and forms an original data stream. The metadata includes timestamp, device identification, associated working condition information and data type identification. The power data includes voltage, current, power and energy; The data classified in the skyline candidate set after EPUS screening is converted into a feature vector; The data priority classification model is used to classify the feature vector to generate a priority label and a classification confidence score; Metadata labels are added to the classified data to obtain the data stream. The metadata labels include the priority label, the data type identification, the timestamp and the classification confidence score.

3. The method of claim 2, wherein, The substation uses a preset Gaussian mixture model classification framework to evaluate the reliability of the data stream to obtain a reliability score, and selects a compression strategy according to the reliability score, the data type identification corresponding to the data stream and the priority label, and packs the data stream into a data frame according to the compression strategy, specifically including: A feature space representation is established for different types of power data, numerical features and context features are extracted based on the feature space representation, and a GMM model is trained based on the numerical features and the context features to obtain a Gaussian mixture model; The Gaussian mixture model is used to calculate the probability density value of each data stream belonging to the classified priority and the probability density value belonging to each priority category. The index evaluation value for the reliability index is calculated by comparing the probability density values, and the reliability score is composed of the index evaluation value, wherein the reliability index includes the intra-class reliability index, the inter-class difference index and the abnormal index; When it is detected that the priority label of the data stream is within a preset low priority range, but the reliability score is higher than a preset high reliability threshold, the priority corresponding to the priority label is raised; When it is detected that the priority label of the data stream is within a preset high priority range, but the reliability score is lower than a preset low reliability threshold, the context information corresponding to the data stream is attached to the data stream; According to at least one of the priority label, the reliability score and the data type identification, a compression strategy is selected, and the data stream is compressed according to the compression strategy to obtain compressed data; Based on the frame structure, data size and priority label of the Beidou short message, the compressed data is packed into a data frame.

4. The method of claim 3, wherein, The compression strategy is used to indicate a compression rate and / or a compression mode of the data stream; the compression strategy is selected according to at least one of the priority label, the reliability score and the data type identifier, and the data stream is compressed according to the compression strategy to obtain compressed data, and specifically includes: The compression strategy is determined according to the priority label and the reliability score, wherein, for the priority label being a first priority, when the reliability score is greater than a preset first reliability threshold, a specified first compression rate is used to compress the data stream, or when the reliability score is not greater than the first reliability threshold, the data stream and simplified data determined by the data stream are used to generate compressed data; for the priority label being a second priority, a specified second compression rate is used to compress the data stream, wherein the first priority is higher than the second priority, and the first compression rate is less than the second compression rate; For power data with time sequence correlation, the data difference value is differentially encoded by combining the current data change trend and the preset load prediction model, to obtain compressed data, wherein the data difference value is determined by the difference between the power data of the current timestamp and the power data of the previous timestamp; The compressed data is packaged into data frames based on the frame structure, data size and the priority label of the Beidou short message, including: For compressed data with the first priority label, each compressed data is packaged into a data frame separately; For compressed data with the second priority label, the compressed data is combined and packaged according to the number of data points of the frame structure to form a data frame; For compressed data with a priority label lower than the second priority, multiple compressed data are spliced in the same data frame.

5. The method of claim 4, wherein, The sub-station determines a transmission strategy according to the data frame and the channel quality prediction result, in combination with a preset context-aware intelligent switching protocol, specifically including: The Beidou signal strength in the previous specified time period is input into the channel quality prediction model to predict the channel quality change in the subsequent specified time period, and the channel quality prediction result is output, including a channel quality prediction level, a confidence interval and a change trend; Based on the channel quality score at the current time, the channel quality prediction level in the subsequent specified time period, the data priority, the resource consumption of the candidate transmission strategy, the current channel quality weight, the predicted channel quality weight, the data priority weight and the strategy resource consumption weight, the strategy score of the candidate transmission strategy is obtained in combination with a preset context-aware intelligent switching protocol. determining a transmission strategy based on the candidate transmission strategies according to the channel quality prediction level and the strategy score, wherein the transmission strategy is used to indicate normal transmission when the channel quality prediction level is greater than a preset high-quality threshold, or to reduce the transmission frequency of the data frames with the priority label in the low priority range when the channel quality prediction level is greater than a preset poor-quality threshold but less than the high-quality threshold, or to suspend transmission of the data frames with the priority label in the low priority range when the channel quality prediction level is less than the poor-quality threshold, and the high-quality threshold is greater than the poor-quality threshold.

6. The method of claim 5, wherein, The data frames are sent to the master station according to the transmission strategy, and specifically include: Monitoring the Beidou transmission state, which includes the transmission success rate, the number of consecutive failures, the average time delay, the availability of backup links, the 4G signal strength, the network delay, and the available bandwidth; When the number of consecutive failures is greater than a preset failure threshold, the transmission success rate falls below a preset success threshold within a specified time, or the channel quality prediction level is lower than a preset quality threshold, at least one available backup link is determined based on the data priority, the timeliness requirement, the current state of each backup link, the energy consumption, and the communication cost, and the data frames with the first priority are switched to the backup link; When no available backup link is detected, the corresponding transmission parameters are adjusted according to a preset optimal transmission path configuration, and the data frames are transmitted according to the transmission parameters, wherein the transmission parameters include the sending power parameter that can be adjusted by the substation in a non-Beidou communication link and the data frame application layer length generated according to the compressed data amount.

7. The method of claim 6, wherein, The master station receives the data frames from the substation, performs integrity check, data decompression and reorganization on the data frames, and feeds back the communication state to the substation, so that the substation updates the database, the data priority classification model, the channel quality prediction model, and the intelligent compression model according to the communication state, and specifically includes: Receiving data frames from the substation and verifying the data integrity of the data frames, generating a retransmission request based on the damaged data frames and returning to the substation until the data integrity verification is passed, wherein the retransmission request is used to instruct the substation to transmit the damaged data frames; Using a decompression strategy matched with the compression strategy adopted by the substation, decompressing and restoring the data frames that pass the verification, reorganizing and recovering the complete data set through the frame number information, and reversing the calculation to restore the original data for the time series data of differential encoding, to obtain complete restored data; According to the restored data, the data transmission quality of each substation is evaluated, and the communication state is generated and fed back to the corresponding substation, so that the substation updates the database based on the communication state, and updates the data priority classification model, the channel quality prediction model, and the intelligent compression model based on the updated database, wherein the transmission evaluation result includes the frame loss rate, the time delay, and the data quality evaluation.

8. The method of claim 7, wherein, The main station monitors the running state of itself, the sub-station and the edge computing unit, early warns the abnormal state based on the pre-trained abnormality detection model, and triggers the corresponding self-healing strategy, specifically including: An abnormality monitoring framework is established, wherein the abnormality states included in the abnormality monitoring framework include data abnormality, communication abnormality and device abnormality; When it is detected that the collected power data exceeds the preset normal range, a data abnormality signal is sent out; When it is detected that the Beidou signal terminal and / or the bandwidth of the backup link is lower than the preset bandwidth threshold, a communication abnormality signal is sent out; When it is detected that the processor load of the sub-station exceeds the preset load threshold and / or the storage space of the sub-station is lower than the preset space threshold, a device abnormality signal is sent out; Based on the data abnormality signal, the communication abnormality signal and the device abnormality signal, a corresponding abnormal feature vector is generated, which includes abnormal type, abnormal source, abnormal index, time feature, correlation information and severity; Using the pre-established reconstruction error model based on autoencoder, the time series prediction model based on LSTM, the correlation analysis model based on graph neural network and the multi-feature classification model based on XGBoost, the abnormal feature vector is classified by abnormal type, evaluated by severity and analyzed by root cause, and the abnormal classification result and severity level are outputted. According to the abnormal classification result and the severity level, a multi-level early warning mechanism is constructed, and a warning is sent out based on the early warning mechanism.

9. The method of claim 8, wherein, The main station triggers the self-healing strategy by the following steps, specifically including: A self-healing strategy library is established, which includes data abnormality self-healing strategy, communication abnormality self-healing strategy and device abnormality self-healing strategy; Based on the self-healing strategy library, according to the abnormal classification result, the severity level, the system state, the environmental condition, the historical experience, the influence evaluation and the resource constraint, a corresponding self-healing strategy is selected, and a self-healing operation instruction is outputted according to the self-healing strategy; According to the self-healing operation instruction, the self-healing operation is performed, and the self-healing operation instruction includes data marking repair, link parameter adjustment switching, service restart and configuration rollback.

10. A Beidou power data intelligent acquisition and dynamic communication scheduling system, characterized in that, It includes: A database construction module is used for the main station to perform data cleaning and feature engineering processing on the acquired historical power data, Beidou communication data, environment and working condition data, to determine the basic parameter configuration and write it into the database, wherein the basic parameter configuration includes model trigger threshold, channel quality threshold, power data acquisition frequency default value, Beidou transmission default frequency level, backup link switching condition and feature engineering parameter; A model construction module is used for the main station to train the data priority classification model, the channel quality prediction model and the intelligent compression model according to the database. The data acquisition module is configured to collect power data by the edge computing unit, construct a skyline candidate set by combining a preset EPUS algorithm to perform data screening, and perform intelligent classification marking on the power data after data screening by using the data priority classification model from the master station, and output a data stream with a priority label; wherein the data screening by combining the preset edge-assisted parallel uncertain skyline EPUS algorithm to construct the skyline candidate set comprises: regarding the power data as data points in a multidimensional space, constructing an importance index space, and the dimension index of the importance index space includes a data change rate, a deviation from a rated value, and a predefined importance of a data type; calculating an importance index of each data point by using the preset EPUS algorithm and the importance index space, and comparing the data point with data in the current skyline candidate set based on the importance index; and when the importance index of the data point exceeds the data in the current skyline candidate set, adding the data point to the skyline candidate set and updating the skyline candidate set. The data processing module is configured to perform reliability evaluation on the data stream by using a classification framework of a preset Gaussian mixture model, obtain a reliability score, select a compression strategy according to the reliability score, a data type identifier corresponding to the data stream, and the priority label, and pack the data stream into a data frame according to the compression strategy. The data frame sending module is configured to determine a transmission strategy according to the data frame and a channel quality prediction result, in combination with a preset context-aware intelligent switching protocol, and send the data frame to the master station according to the transmission strategy, wherein the transmission strategy is used to indicate transmission time selection, multi-link intelligent switching, and parameter adaptive adjustment, and the channel quality prediction result is obtained by predicting a current Beidou signal strength by using the channel quality prediction model from the master station. The data restoration module is configured to receive the data frame from the substation, perform integrity checking, data decompression and recombination on the data frame, and feed back a communication state to the substation, so that the substation updates the database, the data priority classification model, the channel quality prediction model, and the intelligent compression model according to the communication state. The anomaly monitoring module is configured to monitor the running state of the master station, the substation, and the edge computing unit, perform early warning on an abnormal state based on a pre-trained anomaly detection model, and trigger a corresponding self-healing strategy, wherein the abnormal state includes data anomaly, communication anomaly, and device anomaly.

Images