A multi-source heterogeneous data pre-acquisition and edge intelligent processing method

Abstract

The application relates to the field of power system data acquisition and edge computing, and discloses a multi-source heterogeneous data pre-acquisition and edge intelligent processing method, which comprises the following steps: using a modal division mechanism, combining the operation state of power equipment, and performing condition correction on an adaptive sampling frequency to obtain a corrected sampling frequency, so as to acquire multi-source heterogeneous data; extracting fusion core features from multi-dimensional feature vectors; substituting the fusion core features into a general fault judgment model and a time sequence prediction model to respectively output equipment fault prediction results and time sequence prediction results, and obtaining early warning information; based on the fusion core features, the early warning information and the equipment fault prediction results of each acquisition node, and combined with the Lymerzi complexity theory, the spatio-temporal coupling law is identified; and based on the spatio-temporal coupling law and the time sequence prediction results, edge end decision suggestions are provided for source-grid-load-storage collaborative regulation. The application effectively improves the power system data processing efficiency and the power grid operation and maintenance regulation level.

Description

Technical Field

[0001] This invention relates to the field of power system data acquisition and edge computing, specifically to a method for front-end acquisition and edge intelligent processing of multi-source heterogeneous data. Background Technology

[0002] The current power system is developing towards high penetration of new energy sources and integrated generation, grid, load, and storage. During operation, it generates multi-source heterogeneous data, including power operation parameters (voltage, current, power, etc.), environmental sensing data (temperature, humidity, illumination, wind speed, etc.), equipment status data (vibration, temperature rise, operating time, etc.), new energy generation data, energy storage charging and discharging data, and load power consumption data. This data is characterized by its multi-dimensionality, diverse types, high real-time requirements, and strong spatiotemporal correlation. Pre-collection and intelligent processing of this data are crucial for fault prediction, early warning, and coordinated control of the power system, directly impacting the precision of grid operation and maintenance and the efficiency of generation, grid, load, and storage linkage.

[0003] Traditional power data front-end acquisition technology, centered on data acquisition terminals (FTUs, DTUs, collectors, etc.), only achieves data collection, transmission, and simple storage, exhibiting significant technical limitations: Firstly, it has a single acquisition dimension, failing to integrate multimodal sensing data and thus unable to analyze equipment operating status from multiple dimensions; secondly, it lacks edge-end intelligent processing capabilities, with raw data directly uploaded to the cloud, resulting in wasted bandwidth resources, and centralized cloud processing suffers from latency, failing to meet the real-time control and fault early warning requirements of the power system; furthermore, existing technologies lack the ability to perform correlation analysis and time-series prediction of multi-source data, with data from each acquisition node isolated, and data from the power generation, grid, load, and storage systems not achieving collaborative sensing, making it difficult to support integrated control decisions. According to research reports in the power industry, under the traditional front-end acquisition model, the bandwidth utilization rate for power data upload is less than 40%, the false alarm rate for equipment fault early warning exceeds 35%, and the average response latency for collaborative control of the power generation, grid, load, and storage systems exceeds 500ms, making it unsuitable for the development needs of new power systems.

[0004] In the prior art, for example, Chinese Patent Publication No. 202411686606.5 discloses a power data acquisition method based on a power system, which realizes dynamic optimization of the acquisition cycle. By acquiring and sorting the transmission speed of normal data formats, and analyzing the relationship between the conversion response time of data format conversion and data capacity, the most suitable normal data format is selected for transmission based on the actual power data capacity acquired. This helps to improve the efficiency of data acquisition and transmission. However, its focus on dynamically adjusting the acquisition frequency and cycle, as well as optimizing data format transmission, has several shortcomings: First, it lacks the ability to fuse and analyze multi-source heterogeneous data, treating power data as isolated numerical values ​​and failing to achieve correlation mining of multi-modal data such as electrical parameters, environment, and equipment status, making it difficult to capture the spatiotemporal coupling patterns between data; Second, it lacks edge intelligent processing capabilities, relying entirely on the cloud or backend for analysis, without deploying lightweight AI models at the front-end acquisition end, resulting in significant delays in fault warning and time-series prediction, failing to meet real-time requirements; Third, the control loop is incomplete, only completing data acquisition and transmission without constructing an integrated architecture of acquisition, analysis, early warning, prediction, and control, failing to achieve collaborative perception and linkage decision-making among the power generation, grid, load, and storage links, and struggling to support the rapid response needs of new power systems.

[0005] For example, Chinese Patent Publication No. CN202411314580.1 discloses a power equipment fault monitoring method based on edge computing. It utilizes edge computing nodes deployed near power equipment to collect and process data from various data sources such as sensors and cameras. It uses transfer learning technology to achieve rapid adaptation and accurate identification of power equipment of different types and sizes, thereby realizing intelligent monitoring and fault early warning of power equipment. However, while it incorporates edge computing and deep learning for fault monitoring, it has several limitations: First, the multi-source data fusion dimension is singular, focusing only on sensor and camera data from the power equipment itself, without extending to data from the entire power generation, grid, load, and storage system, including renewable energy output, energy storage status, and load changes, and lacking multimodal correlation analysis from a global power system perspective; Second, edge intelligence functions are fragmented, with fault monitoring, time-series prediction, and collaborative control operating independently, failing to form a unified data processing pipeline, resulting in low resource utilization and accumulated processing latency at edge nodes; Third, it lacks collaborative control capabilities, only providing fault warnings for single-point equipment, without building a data interaction and linkage decision-making mechanism between nodes, failing to support multi-dimensional collaborative optimization of power generation, grid, load, and storage, and struggling to meet the demands of new power systems for integrated and rapid response of sensing, knowledge, and control.

[0006] For example, Chinese Patent Publication No. CN201710618593.1 discloses a method and apparatus for multi-source power data fusion processing. This method can fuse power data obtained from various data source systems according to corresponding object labeling rules, effectively simplifying the power data fusion process and improving processing efficiency. However, the centralized backend data fusion architecture has several drawbacks: First, it suffers from severe real-time limitations. Power data from all data sources must be fully uploaded to the backend relational database management system for fusion processing, without implementing edge intelligent processing at the frontend acquisition end. This results in high data transmission bandwidth consumption and large processing latency, failing to meet millisecond-level real-time response requirements. Second, it lacks AI analysis capabilities. It only performs logical relationship mapping and format fusion through object labeling rules, without introducing lightweight neural networks to achieve deep multimodal data correlation mining, fault prediction, and time series prediction, resulting in limited intelligence. Third, it lacks a collaborative control closed loop. The fusion processing results are only used for data presentation at the display layer, without constructing a cross-node collaborative perception and linkage control mechanism for source-grid-load-storage systems. This fails to support real-time optimization decisions for the power system and is ill-suited to the integrated requirements of acquisition, calculation, and control in new power systems.

[0007] Furthermore, existing technologies suffer from several technical limitations. Traditional power data front-end acquisition methods are functionally limited, possessing only basic data acquisition and forwarding capabilities. They are ill-suited to the processing needs of multi-source heterogeneous power system data and lack intelligent analysis and coordinated control capabilities. Consequently, they exhibit several shortcomings in scenarios involving source-grid-load-storage coordinated control and refined grid operation and maintenance: First, traditional acquisition equipment focuses solely on collecting power operation parameters, failing to integrate multi-modal data such as environmental perception and equipment status monitoring. This single data dimension results in a lack of comprehensive basis for equipment fault prediction and low accuracy in fault warnings. Second, the sampling frequency of multi-source heterogeneous power equipment operation data is fixed or lagging during front-end acquisition and edge processing, leading to low utilization efficiency of edge communication and storage resources. Third, they are inadequate for processing diverse data such as new energy generation, energy storage charging and discharging, and load power consumption. The heterogeneous data sources lack real-time intelligent processing capabilities. Direct uploading of raw data results in significant bandwidth consumption, and the data suffers from outliers and inconsistent formats, leading to inefficient subsequent cloud processing. Fourth, the lack of AI-driven multimodal data correlation analysis and time-series prediction capabilities makes it impossible to achieve accurate short-cycle predictions of power generation output and load demand, hindering the support for coordinated regulation and control decisions across power generation, grid, load, and storage. Fifth, data from each acquisition node is isolated, and a multi-dimensional linkage regulation and control module involving power generation, grid, load, and storage has not been constructed, resulting in a lack of data collaborative perception capabilities and limiting the precision of grid operation and control. Furthermore, the heterogeneous data from multiple sources is difficult to represent uniformly at the edge, cross-node linkage relationships are difficult to quantify, and the spatiotemporal coupling patterns between new energy output fluctuations, load mutations, energy storage responses, and grid-side power flow changes are difficult to identify in a timely and accurate manner.

[0008] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0009] To address the problems in related technologies, this invention proposes a method for multi-source heterogeneous data pre-acquisition and edge intelligent processing, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0010] Therefore, the specific technical solution adopted by the present invention is as follows:

[0011] A method for pre-collection and edge intelligent processing of multi-source heterogeneous data includes:

[0012] At each acquisition node, an adaptive sampling frequency is determined based on the data change rate. Using a modal partitioning mechanism and considering the operating status of power equipment, the adaptive sampling frequency is modified according to different operating conditions to obtain a modified sampling frequency. Based on the modified sampling frequency, multi-source heterogeneous data is acquired. This multi-source heterogeneous data is preprocessed and aggregated to obtain standardized aggregated data, which is then converted into a multi-dimensional feature vector. Fusion core features are extracted from the multi-dimensional feature vector. These fusion core features are substituted into a general fault judgment model and a time-series prediction model, outputting equipment fault prediction results and time-series prediction results, respectively. Based on the equipment fault prediction results and an early warning mechanism, early warning information is output. Based on the fusion core features, early warning information, and equipment fault prediction results of each acquisition node, a multi-node joint operation sequence is constructed. The Lemotziff complexity theory is introduced to uniformly quantify the complexity of independent node and cross-node linkage sequences, in order to identify the spatiotemporal coupling patterns between new energy output fluctuations, load mutations, and energy storage responses. Based on the spatiotemporal coupling patterns and time-series prediction results, edge-end decision-making suggestions are provided for the coordinated regulation of power generation, grid, load, and energy storage.

[0013] Furthermore, based on the data change rate, the adaptive sampling frequency is determined, including: dividing the data change rate by the data change threshold and multiplying by the sensitivity coefficient, and combining the basic sampling frequency and the device status weight factor to calculate the adaptive sampling frequency.

[0014] Furthermore, utilizing a modal segmentation mechanism and combining the operating status of power equipment, the adaptive sampling frequency is modified according to different operating conditions to obtain the modified sampling frequency. This includes: acquiring a standardized set of power equipment operating information and processing the various power equipment operating quantities in the standardized power equipment operating information set using a sliding time window; calculating the fluctuation intensity based on the deviation of the mean within the sliding time window, and calculating the abrupt change characteristics based on the jump variables between the current value and the historical window boundary value; normalizing and weighting the fluctuation intensity and abrupt change characteristics to obtain a comprehensive fluctuation index used to characterize the dynamic change degree of power equipment operating data; binding the comprehensive fluctuation index with the corresponding data source label, time window label, and monitoring quantity label to output a set of operating change characteristics; explicitly referencing the load rate, temperature, and alarms transmitted from the standardized power equipment operating information set from the set of operating change characteristics; and based on the load rate, temperature, and alarms... The normalized value of the alarm is used to generate load risk factor, temperature rise risk factor, and alarm risk factor. These factors are then weighted and fused to obtain the equipment operation risk factor. The equipment operation risk factor is then linked to the comprehensive volatility index, outputting a set of equipment state risk characteristics. Based on metastable dwell time theory, and combined with the comprehensive volatility index judgment process and the equipment operation risk factor judgment process, the modes of the power equipment are determined, including pure steady state, metastable state, and fault state. Different sampling frequency corrections are assigned to the pure steady state, metastable state, and fault state, with the corrections increasing sequentially. Based on the duration of continuous dwell time of the power equipment in the metastable state, the sampling frequency correction for the metastable state is adjusted. Based on the sampling frequency corrections obtained under the pure steady state, metastable state, and fault state, the adaptive sampling frequency is modified according to different operating conditions to obtain the corrected sampling frequency.

[0015] Furthermore, based on the metastable residence theory and combined with the comprehensive volatility index judgment process and the equipment operation risk factor judgment process, the operating mode of the power equipment is determined, including: configuring a steady-state threshold and a fault threshold for the comprehensive volatility index, and a first risk threshold and a second risk threshold for the equipment operation risk factor, wherein the steady-state threshold is less than the fault threshold, and the first risk threshold is less than the second risk threshold; when the comprehensive volatility index is less than or equal to the steady-state threshold, the equipment operation risk factor is less than or equal to the first risk threshold, and the current power equipment does not have an alarm state, the operating mode of the power equipment is determined to be a pure steady state; when the comprehensive volatility index is between the steady-state threshold and the fault threshold, the equipment operation risk factor is between the first risk threshold and the second risk threshold, and the current power equipment does not have an alarm state, the operating mode of the power equipment is determined to be a metastable state; when the comprehensive volatility index is greater than or equal to the fault threshold, the equipment operation risk factor is greater than or equal to the second risk threshold, and the current power equipment has an alarm state, at least one of these conditions occurs, the operating mode of the power equipment is determined to be a fault state.

[0016] Furthermore, based on the current dwell time of the power equipment in a metastable state, the sampling frequency correction amount for the metastable state is adjusted, including: if the current operating mode of the power equipment is metastable and the operating mode of the previous judgment period was also metastable, the current metastable state dwell time is updated; if the current operating mode of the power equipment is metastable and the operating mode of the previous judgment period was not metastable, the current metastable state dwell time is initialized to the preset judgment period duration; if the current operating mode of the power equipment is not metastable, the metastable state dwell time is cleared to zero; the sampling frequency correction amount for the metastable state is adjusted according to the principle that the metastable state sampling frequency correction amount gradually decreases with the increase of the metastable state dwell time and is higher than the minimum metastable state correction amount.

[0017] Furthermore, the multi-source heterogeneous data is preprocessed and aggregated to obtain standardized aggregated data, which is then converted into multi-dimensional feature vectors. The core fusion features are extracted from these multi-dimensional feature vectors. This process includes: sequentially cleaning, standardizing, and aggregating the multi-source heterogeneous data to obtain standardized aggregated data; encoding the standardized multi-modal data into multi-dimensional feature vectors based on a multi-modal fusion neural network model; calculating the attention weights between the multi-dimensional feature vectors using a multi-head attention mechanism to dynamically capture correlations; generating a unified fusion feature matrix through a feature fusion layer; and extracting the core fusion features through a feature dimensionality reduction layer.

[0018] Furthermore, based on the fusion core features, early warning information, and equipment fault prediction results of each acquisition node, a multi-node joint operation sequence is constructed. The Lemzef complexity theory is introduced to uniformly quantify the complexity of independent node and cross-node linkage sequences, in order to identify the spatiotemporal coupling patterns between new energy output fluctuations, load mutations, and energy storage responses. This includes: determining cross-node linkage combinations based on the physical connection relationships, power flow transmission relationships, and regulation response relationships between the source, grid, load, and storage systems; concatenating the fusion core features, early warning information, and equipment fault prediction results within the same cross-node linkage combination and at the same analysis time according to the node order of the source, grid, load, and storage systems, and arranging them in chronological order to output the multi-node joint operation sequence of the source, grid, load, and storage systems; and analyzing the various fusion core features, early warning information, and equipment fault prediction results in the multi-node joint operation sequence. The results are determined by establishing a binary coding benchmark. The core features, early warning information, and equipment fault prediction results of each acquisition node are compared with the corresponding binary coding benchmark to form a node-independent binary sequence. The binary states of each acquisition node within the same cross-node linkage combination are concatenated according to a preset node order to form a cross-node joint binary sequence. Based on the Lemotziff complexity algorithm, the complexity of the node-independent binary sequence and the cross-node joint binary sequence are calculated to obtain the node-independent Lemotziff sequence complexity and the cross-node joint Lemotziff sequence complexity. Based on the node-independent Lemotziff sequence complexity and the cross-node joint Lemotziff sequence complexity, a time-series correspondence analysis is performed, combining the changes in new energy output, load, energy storage charging and discharging, and grid-side power flow, to identify the spatiotemporal coupling patterns.

[0019] Furthermore, based on the Lemziff complexity algorithm, the complexity of node-independent binary sequences and cross-node joint binary sequences is calculated to obtain the node-independent Lemziff sequence complexity and the cross-node joint Lemziff sequence complexity. This includes: determining the node-independent scan sequence and the cross-node joint scan sequence based on the node-independent binary sequence and the cross-node joint binary sequence; configuring the initialized Lemziff complexity calculation parameters and scanning each node-independent scan sequence and the cross-node joint scan sequence according to the time order and scan order; comparing the currently formed binary substring with the historical substring set, and if the current binary substring is not included in the historical substring set, adding the current binary substring to the historical substring set and incrementing the complexity count by one, outputting the original Lemziff complexity count value corresponding to each binary scan sequence; normalizing the original Lemziff complexity count value and extracting the node-independent Lemziff sequence complexity and the cross-node joint Lemziff sequence complexity.

[0020] Furthermore, based on the node-independent Lemotsev sequence complexity and the cross-node joint Lemotsev sequence complexity, and combined with the time-series correspondence analysis of the changes in renewable energy output, load, energy storage charging and discharging, and grid-side power flow, the spatiotemporal coupling patterns are identified. This includes: calculating operational change characteristics based on operational data from adjacent analysis windows, including the changes in renewable energy output, load, energy storage charging and discharging, and grid-side power flow; performing adjacent window difference analysis, variation amplitude statistics, and trend extraction on the node-independent Lemotsev sequence complexity and the cross-node joint Lemotsev sequence complexity to obtain complexity change characteristics; determining the temporal sequence relationship between renewable energy output fluctuations, load abrupt changes, energy storage charging and discharging responses, and grid-side power flow changes based on the complexity change characteristics and operational change characteristics; calculating the corresponding response delay based on the time interval between the trigger-side change time and the response-side change time; and identifying the spatiotemporal coupling patterns among renewable energy output fluctuations, load abrupt changes, energy storage responses, and grid-side power flow changes based on the temporal sequence relationship, response delay, the direction of change of operational change characteristics, and complexity change characteristics.

[0021] Furthermore, a method for multi-source heterogeneous data pre-acquisition and edge intelligent processing also includes: using the power balance equation to evaluate the consistency of the power system's operating state, wherein the power balance equation is: P gen +P grid +P dis =P load +P ch +P loss In the formula, P gen For new energy power generation; P grid Power supplied to the power grid; P dis P is the energy storage discharge power; load P represents the load power. ch Power for energy storage charging; P loss This refers to line loss power.

[0022] The beneficial effects of this invention are as follows:

[0023] 1. Multi-dimensional data acquisition for comprehensive perception: This invention integrates a multi-modal perception module, breaking the limitation of traditional front-end acquisition that only collects electrical parameters. It realizes the synchronous acquisition of multi-source heterogeneous data in all aspects of power operation, environment, equipment status, and power generation, grid, load and storage, providing comprehensive and complete data basis for equipment fault prediction and coordinated control, and solving the problem of single data dimension in traditional technology.

[0024] 2. Sampling frequency is processed according to different working conditions to balance monitoring accuracy and resource utilization efficiency: This invention uses a working condition correction mechanism that distinguishes between pure steady state, metastable state and fault state. When the equipment is running stably, the sampling frequency can be reduced to reduce invalid data collection and communication load; when the equipment enters metastable state, the sampling frequency can be appropriately increased to enhance the tracking ability of potential abnormal evolution process; when the equipment enters fault state, the sampling frequency can be increased to improve the accuracy of fault process capture.

[0025] 3. Intelligent processing at the edge reduces latency and bandwidth consumption: A lightweight AI model is deployed at the front-end acquisition end to complete real-time data cleaning, standardization, and aggregation. Only high-quality processing results and analysis data are uploaded to the cloud, significantly reducing the amount of raw data uploaded and reducing bandwidth resource consumption. At the same time, local processing at the edge enables low latency for fault early warning and time series prediction, meeting the needs of real-time control and fault response in the power system and solving the latency and bandwidth problems of traditional centralized cloud processing.

[0026] 4. AI multimodal fusion analysis to improve the accuracy of fault early warning: By using a lightweight multimodal fusion algorithm, the inherent correlation patterns of multi-dimensional data such as electrical parameters, environment, and equipment status are mined. Combined with the fault feature library, the accurate prediction of equipment faults is achieved. Compared with the traditional single threshold judgment early warning method, the false alarm rate and false alarm rate are greatly reduced. The integration of data collection, status monitoring and fault early warning is realized, and the level of precision of power grid operation and maintenance is improved.

[0027] 5. Short-cycle accurate time-series forecasting to support coordinated regulation of power generation, grid, load and storage: The model embeds a lightweight time-series forecasting algorithm to achieve minute-level / hour-level accurate forecasting of new energy power generation output and load electricity demand. The forecasting results provide high-precision data support for coordinated regulation of power generation, grid, load and storage. At the same time, the model supports real-time update training to adapt to the dynamic changes of the power system, solving the problems of low accuracy and poor adaptability of traditional forecasting methods.

[0028] 6. Source-grid-load-storage collaborative sensing to achieve multi-dimensional linkage control: Construct a multi-dimensional linkage control module for source-grid-load-storage to break the limitation of isolated data from each acquisition node, realize edge-end collaborative sensing and linkage analysis of data from all aspects of source-grid-load-storage, output preliminary control decision suggestions, realize rapid linkage of all aspects, promote integrated control of source-grid-load-storage, and improve the operating efficiency and stability of the power system.

[0029] 7. Lightweight AI model design, adapted to edge computing power: The edge intelligent processing model adopts a hybrid architecture of lightweight neural network and traditional machine learning, which is optimized for the limited computing power and storage resources at the edge. It does not rely on high-end computing equipment, is easy to deploy and promote, and is compatible with the hardware environment of various power front-end acquisition nodes, reducing the cost of technology implementation.

[0030] 8. Identification of spatiotemporal coupling patterns: By combining the time-series correspondence analysis of the changes in new energy output, load, energy storage charging and discharging, and grid-side power flow, the sequence, response delay, and direction of change among new energy fluctuations, load abrupt changes, energy storage responses, and power flow changes can be determined, thus enabling the identification of spatiotemporal coupling patterns.

[0031] 9. The advantages of this invention are mainly reflected in high prediction accuracy, efficient data processing, stability and reliability, and reduced costs. Its innovative proactive resource management paradigm provides a verifiable technical path for the intelligent operation and maintenance of industrial-grade time-series databases. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of a method for multi-source heterogeneous data pre-acquisition and edge intelligent processing according to an embodiment of the present invention;

[0034] Figure 2 This is a system block diagram according to an embodiment of the present invention;

[0035] Figure 3 This is a flowchart illustrating the process of correcting the sampling frequency in a multi-source heterogeneous data pre-acquisition and edge intelligent processing method according to an embodiment of the present invention;

[0036] Figure 4 This is a flowchart illustrating the process of obtaining spatiotemporal coupling patterns in a multi-source heterogeneous data pre-acquisition and edge intelligent processing method according to an embodiment of the present invention.

[0037] Figure 5 This is an overall architecture design diagram according to an embodiment of the present invention. Detailed Implementation

[0038] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0039] According to an embodiment of the present invention, a method for front-end acquisition and edge intelligent processing of multi-source heterogeneous data is provided.

[0040] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a method for multi-source heterogeneous data pre-acquisition and edge intelligent processing is provided, comprising:

[0041] S1. At each acquisition node, the adaptive sampling frequency is determined based on the data change rate. Using the mode division mechanism and combined with the operating status of the power equipment, the adaptive sampling frequency is modified according to the operating conditions to obtain the modified sampling frequency.

[0042] In one embodiment, determining the adaptive sampling frequency based on the data change rate includes: dividing the data change rate by a threshold for data change and multiplying by a sensitivity coefficient, and then calculating the adaptive sampling frequency by combining the base sampling frequency and the device state weighting factor.

[0043] like Figure 3As shown, in one embodiment, a modal partitioning mechanism is used, combined with the operating status of the power equipment, to perform condition-specific correction on the adaptive sampling frequency, resulting in a corrected sampling frequency. This includes: acquiring a standardized set of power equipment operating information, such as load rate, temperature rise, alarm status, and alarm level. The standardized set of power equipment operating information includes operating data after time alignment, missing value completion (e.g., linear interpolation, forward filling), outlier removal (e.g., 3σ), and normalization processing (e.g., Min-Max). Time alignment is based on the timestamp of the equipment's local high-precision clock, meaning all monitored quantities are resampled along a unified time axis (e.g., every hour), and non-hourly data is generated through linear interpolation. If the delay of a certain channel exceeds a preset tolerance window (e.g., 2 seconds), a data validity flag is triggered. During this period, the data of that channel does not participate in the calculation of the volatility index and risk factor, avoiding asynchronous data distortion in the state judgment. A sliding time window process is applied to the operating quantities of various power equipment in the standardized power equipment operation information set. Fluctuation intensity is calculated based on the deviation of the mean within the sliding time window, and abrupt change characteristics are calculated based on the jump variables between the current value and the historical window boundary value. The fluctuation intensity and abrupt change characteristics are normalized and weighted to obtain a comprehensive fluctuation index characterizing the dynamic changes in power equipment operation data. The comprehensive fluctuation index is bound to the corresponding data source label, time window label, and monitoring quantity label to output a set of operation change characteristics. Load rate, temperature, and alarms transmitted from the standardized power equipment operation information set are explicitly referenced from the set of operation change characteristics. Based on the normalized values ​​of load rate, temperature, and alarms, load risk factors, temperature rise risk factors, and alarm risk factors are generated. The load risk factors, temperature rise risk factors, and alarm risk factors are then added... Weighted fusion yields equipment operation risk factors; these factors are then linked to a comprehensive volatility index, outputting a set of equipment state risk characteristics that simultaneously reflect both real-time and dynamic operational risk levels. Based on metastable dwell time theory, and combined with the comprehensive volatility index and equipment operation risk factor judgment processes, the modes of power equipment are determined, including pure steady-state, metastable, and fault states. Different sampling frequency corrections are assigned to the pure steady-state, metastable, and fault states, with the corrections increasing sequentially. The sampling frequency correction for metastable states is adjusted based on the duration of continuous dwell time of the power equipment in the metastable state. Finally, the adaptive sampling frequency is adjusted according to the operating conditions based on the sampling frequency corrections obtained under the pure steady-state, metastable, and fault states, resulting in a corrected sampling frequency. For example, if the sampling frequency corrections for pure steady-state, metastable, and fault states are 0Hz, 2.0Hz, and 5.0Hz respectively, the corrected sampling frequency is... :

[0044] ;

[0045] In the formula, Indicates the lower limit of the sampling frequency. Indicates the upper limit of the sampling frequency. This is the modal correction value. For adaptive sampling frequency.

[0046] In one embodiment, based on the metastable dwelling theory and combined with the comprehensive volatility index judgment process and the equipment operation risk factor judgment process, the mode of the power equipment is determined, including: configuring a steady-state threshold and a fault threshold for the comprehensive volatility index, and a first risk threshold and a second risk threshold for the equipment operation risk factor, wherein the steady-state threshold is less than the fault threshold, and the first risk threshold is less than the second risk threshold; when the comprehensive volatility index is less than or equal to the steady-state threshold, the equipment operation risk factor is less than or equal to the first risk threshold, and the current power equipment does not have an alarm state, the operating mode of the power equipment is determined to be a pure steady state; when the comprehensive volatility index is between the steady-state threshold and the fault threshold, the equipment operation risk factor is between the first risk threshold and the second risk threshold (or set to the equipment operation risk factor not reaching the second risk threshold), and the current power equipment does not have an alarm state, the operating mode of the power equipment is determined to be a metastable state; when at least one of the following conditions occurs: the comprehensive volatility index is greater than or equal to the fault threshold, the equipment operation risk factor is greater than or equal to the second risk threshold, and the current power equipment has an alarm state, the operating mode of the power equipment is determined to be a fault state.

[0047] In one embodiment, the metastable state sampling frequency correction is adjusted based on the current dwell time of the power equipment in a continuous metastable state. This includes: if the current operating mode of the power equipment is metastable and the operating mode of the previous judgment period was also metastable, updating the current metastable state dwell time; if the current operating mode of the power equipment is metastable and the operating mode of the previous judgment period was not metastable, initializing the current metastable state dwell time to a preset judgment period duration, such as 60s; if the current operating mode of the power equipment is not metastable, clearing the metastable state dwell time to zero; adjusting the metastable state sampling frequency correction based on the principle that the metastable state sampling frequency correction gradually decreases with increasing metastable state dwell time and is higher than the minimum metastable state correction, where the minimum metastable state correction is, for example, 0.5Hz. The following exponential decay formula is used:

[0048] ;

[0049] in, This is the correction amount for the current metastable sampling frequency. This is the correction amount for the initial sampling frequency in metastable conditions. This is the minimum metastable correction amount. For the current metastable state residence time, The decay time constant is tuned based on both the type of power equipment and historical fault samples. For power equipment with a historical operating database, the optimal decay time constant is solved by fitting the historical metastable residence time and sample correction change curves using the least squares method. When no historical data is available, industry-standard empirical values ​​for the same type of equipment are used. The judgment cycle is configured to 30 seconds, with modal judgment completed every 30 seconds.

[0050] When performing modal classification of power equipment, dynamic characteristics reflecting changes in equipment operation are extracted from a standardized set of power equipment operation information. This involves calculating fluctuation intensity and abrupt change characteristics using a sliding time window, and then normalizing and fusing these two factors into a comprehensive fluctuation index to characterize the drastic changes in equipment operation data over time. Load rate, temperature, and alarm information are referenced from the same set of operation information to generate load risk factors, temperature rise risk factors, and alarm risk factors, respectively. These are then weighted and fused into a single equipment operation risk factor to characterize the current operational risk level of the equipment. Specifically, the load risk factor is linearly mapped based on the normalized load rate, increasing linearly in the 0-0.6 range and exponentially in the 0.6-1 range. The temperature rise risk factor is segmented with the equipment's rated temperature rise limit as the dividing line. The alarm risk factor uses discrete quantization: 0 for no alarm, 0.4 for general alarms, and 1 for emergency alarms. These three factors are then weighted using preset weights to obtain the total equipment operation risk factor.

[0051] The comprehensive volatility index and equipment operation risk factor are used together as the basis for mode classification. Combined with the presence or absence of alarm states, the equipment operation state is divided into pure steady state, metastable state and fault state. Different sampling frequency correction amounts are assigned to different modes, and a residence time mechanism is introduced for metastable state, so that the metastable state sampling frequency correction amount gradually decreases with the increase of continuous residence time, so as to realize the adaptive sampling frequency correction according to the equipment operation state.

[0052] The steady-state threshold and fault threshold of the comprehensive volatility index are set based on historical operating data statistical quantiles, expert experience, or differences in equipment type; the weights of the load risk factor, temperature rise risk factor, and alarm risk factor are determined based on the importance of the equipment, the reliability of monitoring data, or the correlation with faults. For example, the preferred steady-state threshold of the comprehensive volatility index is: The optimal fault threshold for the comprehensive volatility index is: The preferred first risk threshold for equipment operation risk factors is: The preferred second risk threshold for equipment operation risk factors is: The upper and lower limits of the sampling frequency have been revised to 0.1Hz and 10Hz, respectively, to avoid insufficient state awareness due to an excessively low sampling frequency, or waste of communication and storage resources due to an excessively high sampling frequency.

[0053] For the operating load of the j-th type of power equipment, within the sliding time window at the current time t, first calculate the window mean, and then calculate the fluctuation intensity:

[0054] ;

[0055] The mutation feature can be calculated based on the jump variable between the current value and the historical window boundary value:

[0056] ;

[0057] In the formula, The length of the sliding time window. For the first Class runtime at time The standardized value, The mean of the window. For the first Class runtime at time The standardized value.

[0058] This invention identifies dynamic changes in power equipment operating data through a comprehensive fluctuation index, avoiding delays in status identification caused by relying solely on single-point data. By incorporating equipment operation risk factors, it introduces key operational risk information such as load rate, temperature, and alarms, ensuring that sampling frequency adjustments consider not only the magnitude of data changes but also the operational risks inherent in the equipment itself. Through a condition-specific correction mechanism for pure steady-state, metastable, and fault states, the sampling frequency can be reduced during stable operation, minimizing invalid data collection and communication load; the sampling frequency can be appropriately increased when the equipment enters a metastable state, enhancing the tracking ability of potential anomaly evolution; and the sampling frequency can be increased when the equipment enters a fault state, improving the accuracy of fault process capture. The metastable dwell time decay mechanism avoids continuous high-frequency sampling when the equipment is in a state of slight fluctuation for extended periods, balancing monitoring accuracy and resource utilization efficiency.

[0059] The multimodal sensing layer of this invention serves as the foundation for front-end data acquisition. It integrates four major sensing modules: an electrical parameter acquisition module, an environmental sensing module, an equipment status monitoring module, and a multi-source power acquisition module. This enables synchronous front-end acquisition of multi-dimensional and multi-type heterogeneous data from the power system, covering the entire power generation, grid, load, and storage system, providing a comprehensive data foundation for subsequent intelligent processing. The specific data types acquired by each module are as follows:

[0060] 1. Electrical Parameter Acquisition Module: Through voltage, current, power, and frequency sensors, it collects real-time electrical parameter data from grid nodes, new energy power plants, energy storage facilities, and load-side terminals. The sampling accuracy reaches the millisecond level, and it supports AI dynamic adjustment of the sampling frequency.

[0061] 2. Environmental Sensing Module: Through temperature, humidity, vibration, light, wind speed, and air pressure sensors, it collects on-site environmental data and equipment operating environment data at each data collection node, achieving comprehensive perception of environmental status.

[0062] 3. Equipment Status Monitoring Module: Through vibration sensors, temperature rise sensors, speed sensors, and operating status detectors, it collects real-time operating status data of power equipment such as inverters, transformers, and switchgear, and captures subtle changes in equipment operation.

[0063] 4. Multi-source power acquisition module: Collects new energy power generation data (photovoltaic power output, wind power output, etc.), energy storage charging and discharging data (charging power, discharging power, remaining capacity, etc.), and load power consumption data (peak and valley load, power consumption type, load change rate, etc.), realizing the acquisition of data in the entire process of source, grid, load and storage.

[0064] The multimodal sensing layer is equipped with a data acquisition synchronization control unit. This unit utilizes timestamp synchronization technology to achieve precise synchronization of the acquisition time across the four modules, with an error controlled within 1ms, ensuring the spatiotemporal correlation of the acquired data.

[0065] ;

[0066] Among them, t i Let t be the timestamp of the i-th module. ref This is the system's base time.

[0067] Simultaneously, a dynamic sampling frequency adjustment unit is set up to receive AI decision-making instructions from the edge intelligent processing layer. Based on the data change rate, equipment operating status, and power grid control requirements, it dynamically adjusts the sampling frequency of each module, reducing invalid data collection while ensuring data integrity. This AI decision-making instruction refers to the adaptive sampling decision output generated after training with multi-source heterogeneous data, based on deep neural networks and their derivative architectures (including recurrent neural networks, graph attention networks, deep reinforcement learning, etc.). The main process is as follows: by comprehensively analyzing the data change trends uploaded by the multimodal perception layer, equipment health status, and real-time power grid control requirements, the optimal sampling strategy is output. The principle is that the intelligent decision engine constructs a multi-dimensional state space based on the data change rate of each acquisition node, equipment operating status weight factors, power grid load control priorities, and edge node communication load; constructs an action space based on the dynamic adjustment of the sampling frequency of each perception module; and constructs a multi-dimensional reward function based on data integrity benefits, communication bandwidth resource conservation, timely abnormal event capture, and equipment lifespan loss. Through continuous interactive training with the physical environment of the power system, an adaptive sampling decision strategy is formed that minimizes redundant data collection while ensuring no key data is lost. Based on the AI ​​decision-making instructions, the frequency dynamic control unit dynamically adjusts the sampling frequency of each module according to the data change rate, equipment operating status, and power grid control requirements, thereby reducing invalid data collection while ensuring data integrity.

[0068] The frequency dynamic control unit employs adaptive sampling frequency control based on the data change rate.

[0069] ;

[0070] in, The base sampling frequency; (This is the sensitivity coefficient, ranging from [0, 1]). For the rate of change of data; The threshold for data variation, ranging from [0, 1). The equipment status weight factor (range ∈ [0, 1) is used to quantify the degree of influence of the operating status of power equipment on the sampling frequency regulation.

[0071] S2. Collect multi-source heterogeneous data according to the corrected sampling frequency; preprocess and aggregate the multi-source heterogeneous data to obtain standardized aggregated data, and convert the standardized aggregated data into a multi-dimensional feature vector, and extract the fusion core features from the multi-dimensional feature vector.

[0072] In one embodiment, multi-source heterogeneous data is preprocessed and aggregated to obtain standardized aggregated data, and the standardized aggregated data is converted into multi-dimensional feature vectors. The core fusion features are extracted from the multi-dimensional feature vectors. This includes: performing data cleaning, data standardization, and data aggregation on the multi-source heterogeneous data in sequence to obtain standardized aggregated data; encoding the standardized multi-modal data into multi-dimensional feature vectors based on a multi-modal fusion neural network model; calculating the attention weights between the multi-dimensional feature vectors through a multi-head attention mechanism to dynamically capture the correlation relationship, generating a unified fusion feature matrix through a feature fusion layer, and extracting the core fusion features through a feature dimensionality reduction layer.

[0073] The edge intelligent processing layer of this invention is the core of the invention. It deploys a lightweight AI edge intelligent processing model based on a hybrid architecture design of lightweight neural networks and traditional machine learning. This model is adapted to the computing power and storage resources at the edge, and completes real-time intelligent processing of multi-source heterogeneous data at the front-end acquisition end without relying on cloud computing power, achieving low-latency processing. This layer includes four core units: a data preprocessing unit, an AI multimodal fusion analysis unit, a fault early warning unit, and a time series prediction unit. These units work together to transform raw data into high-quality analysis results.

[0074] 1. Data Preprocessing Unit: Addressing the issues of inconsistent formats, outliers, heterogeneous data dimensions, and excessive redundancy in the multi-source heterogeneous data collected by the multimodal sensing layer, the data preprocessing unit performs real-time data cleaning, standardization, and aggregation. This removes invalid data, unifies data formats, improves data quality, and reduces data volume, thereby lowering subsequent upload bandwidth usage. The specific processing flow is as follows:

[0075] (1) Data cleaning: AI outlier detection algorithm (isolated forest combined with sliding window threshold method) is used to identify and remove outliers from the collected data. For missing data, linear interpolation, K-nearest neighbor completion, and time series prediction completion are used to complete the missing data according to the data type. Redundant data collected repeatedly is deduplicated, and disordered data is reordered according to timestamp to ensure the temporal continuity of the data.

[0076] (2) Data Standardization: Construct a standardized model for multi-source heterogeneous power data, and uniformly convert data of different types, dimensions, and formats (such as analog, digital, and switching quantities) into a standardized format, map them to a unified data dimension, and generate a standardized data matrix; at the same time, label each data with tags such as acquisition node, equipment number, timestamp, and data type to achieve data traceability; Multi-source heterogeneous data standardization mapping formula:

[0077] For the i-th type of data x i Standardization maps to a unified data space:

[0078] ;

[0079] in, The dimension conversion coefficient is... The offset ensures that all data is mapped to the interval [0, 1] or [-1, 1]. , These are the maximum and minimum values ​​of the i-th data type, respectively.

[0080] Data label generation formula:

[0081] Tag=Hash(NodeID||DeviceID||Timestamp||DataType);

[0082] Among them, the Tag data tag is the unique identifier of this standardized data, used for subsequent data traceability, index retrieval, source location, and full lifecycle management, ensuring that each data can be uniquely identified and verified. Hash is a hash function that maps the concatenated multi-field string to a unique hash value of a certain length. NodeID is the acquisition node identifier, referring to the unique number of the physical or logical acquisition node where sensors are deployed in the multimodal sensing layer. DeviceID is the device identifier, referring to the unique number of the monitored power equipment or sensor. Timestamp is the precise time stamp of the data acquisition. DataType is the data type identifier, referring to the physical quantity type classification code to which this standardized data belongs.

[0083] (3) Data aggregation: Based on the source-grid-load-storage link, the geographical distribution of collection nodes, and the time dimension, the standardized data is aggregated in multiple dimensions to generate node-level, regional-level, and link-level aggregated data.

[0084] 2. Multimodal fusion analysis unit:

[0085] A lightweight Transformer-based multimodal fusion neural network model is employed to encode standardized multimodal data into multidimensional feature vectors. A multi-head attention mechanism is used to calculate attention weights between features of different modalities, dynamically capturing correlations. A unified fusion feature matrix is ​​generated through a feature fusion layer, and core features are extracted through a feature dimensionality reduction layer, providing high-information-density input for fault early warning and time-series prediction. Specifically, four types of heterogeneous data (electrical parameters, environmental perception, equipment status, and power generation-grid-load-storage data) standardized by the data preprocessing unit are input into independent lightweight encoders according to modal type. Each encoder consists of a one-dimensional convolutional layer (Conv1D) and a positional encoder: the convolutional layer extracts local temporal features of each modality, and the positional encoder injects temporal sequence information, generating a modality-specific feature vector with time-series awareness. To adapt to edge computing power limitations, the encoder uses depthwise separable convolutions instead of standard convolutions.

[0086] Since the sampling frequencies, time window lengths, and feature dimensions of the four types of data vary, cross-modal temporal alignment is required. A linear projection layer is used to map the feature vectors of each modality to a unified d-dimensional latent space, and temporal interpolation or downsampling is used to ensure that the lengths of the modal sequences are consistent, generating a standardized multimodal feature sequence matrix X∈R. M ×T×d Where M=4 is the number of modes and T is the uniform timing length.

[0087] Multimodal feature sequences are input into a lightweight multi-head attention module. Unlike traditional Transformers, this module employs a low-rank attention mechanism, reducing the O(T) time complexity of standard attention. 2 The complexity is reduced to O(T), significantly reducing the computational overhead at the edge. Specific steps:

[0088] (1) Input Z into a lightweight multi-head attention module. Low-rank approximate attention is used to reduce complexity. For each attention head h (h=1, 2, ..., H), where H is a hyperparameter, calculate the query, key, and value matrix:

[0089] ; ; ;

[0090] The query matrix is ​​W. h Q ∈R d×d Key matrix W h K ∈R d×d Value matrix W h K ∈R d×dEach attention head independently computes the attention weight distribution across and within modes, using low-rank approximate attention. calculate:

[0091] ;

[0092] Where r≪T represents the reduced rank.

[0093] (2) Multi-head attention output h Parallel computing:

[0094] ;

[0095] (3) The attention output is fed into the feedforward network (FFN) after being normalized by layer normalization (LayerNorm) and connected to the residual.

[0096] S3. Substitute the core features into the general fault judgment model and the time series prediction model, and output the equipment fault prediction results and the time series prediction results respectively; based on the equipment fault prediction results, combined with the early warning mechanism, output the early warning information.

[0097] The fault early warning unit of this invention: Based on the core features extracted by the AI ​​multimodal fusion analysis unit, the fault early warning unit constructs a multi-dimensional predictive model for equipment faults, realizing accurate prediction and real-time early warning of faults in power equipment (inverters, transformers, lines, etc.), achieving integrated status monitoring, fault prediction, and early warning dissemination. The specific process is as follows:

[0098] Fault Feature Database Construction: Based on fault cases, operational data, and experimental data, a fault feature database is constructed for common faults such as inverter anomalies, line losses, transformer temperature rise, and switchgear faults.

[0099] Fault prediction: The core features are input into the fault prediction model (an existing general fault judgment model, such as a lightweight CNN and LSTM hybrid model), compared with the fault feature library, and combined with equipment operating trends, environmental changes, and electrical parameter fluctuations, to output the probability, type, and location of the fault. The following are brief steps for fault prediction using a lightweight CNN and LSTM hybrid model of a general fault judgment model:

[0100] Feature access: The core features of multimodal fusion are constructed into a temporal input tensor according to the standard model format, and the prior candidate set is obtained by simultaneously searching the fault feature library.

[0101] CNN Local Extraction: Extracting local spatial feature patterns of electrical parameter fluctuations and abnormal equipment status through convolutional layers.

[0102] LSTM Temporal Modeling: Based on CNN output, it utilizes a long short-term memory network to model the long-term degradation trend of device state and the parameter coupling evolution law.

[0103] Fusion decision output: Combining the LSTM hidden state with prior information from the fault feature library, the fault probability distribution, fault type classification, and device location are output through a fully connected layer.

[0104] Dynamic threshold early warning: Based on preset probability thresholds, a three-level early warning judgment is made, and dynamic corrections are made in combination with real-time environmental changes and electrical parameter fluctuations to generate the final early warning level and handling suggestions.

[0105] Tiered early warning system: A three-tiered early warning mechanism (Level 1 / Level 2 / Level 3) is set up, corresponding to different early warning methods (local audio and visual alerts, platform messages, and SMS), while also providing fault handling suggestions. Early warning information is pushed to local operation and maintenance terminals and the cloud monitoring platform in real time.

[0106] The timing prediction unit of the present invention:

[0107] By embedding a lightweight time-series prediction algorithm (existing general fault diagnosis models, such as improved models based on TCN and GRU), and using integrated core features as input, we can achieve short-cycle accurate prediction (minute / hour level) of new energy power generation output and load electricity demand.

[0108] The model captures the spatiotemporal correlation patterns of multimodal data and combines historical operating data, environmental forecast data, and load change trends to achieve predictions. It supports real-time training updates, dynamically optimizing prediction parameters based on the latest collected data. Specifically, addressing the randomness and volatility of renewable energy power generation and the periodicity and variability of load demand, the model captures the spatiotemporal correlation patterns of multimodal data and combines historical operating data, environmental forecast data (solar radiation, wind speed), and load change trends to achieve short-cycle predictions of power generation output and load demand. Simultaneously, the model supports real-time training updates, dynamically optimizing prediction parameters based on the latest collected data analysis results to improve prediction accuracy and adaptability.

[0109] like Figure 4 As shown in Figure S4, based on the fusion core features, early warning information and equipment fault prediction results of each acquisition node, a multi-node joint operation sequence is constructed. The Lemotzov complexity theory is introduced to unify and quantify the complexity of independent node and cross-node linkage sequences in order to identify the spatiotemporal coupling law between new energy output fluctuations, load mutations and energy storage response.

[0110] In one embodiment, based on the fusion core features, early warning information, and equipment fault prediction results of each acquisition node, a multi-node joint operation sequence is constructed. The Lemzef complexity theory is introduced to uniformly quantify the complexity of independent node and cross-node linkage sequences, in order to identify the spatiotemporal coupling patterns between new energy output fluctuations, load mutations, and energy storage responses. This includes: determining cross-node linkage combinations based on the physical connection relationships, power flow transmission relationships, and regulation response relationships between the source, grid, load, and storage; concatenating the fusion core features, early warning information, and equipment fault prediction results within the same cross-node linkage combination and at the same analysis time according to the node order of the source, grid, load, and storage, and arranging them in chronological order to output the multi-node joint operation sequence of the source, grid, load, and storage; determining the binary encoding benchmark for various fusion core features, early warning information, and equipment fault prediction results in the multi-node joint operation sequence; and... The core features, early warning information, and equipment fault prediction results of each acquisition node are compared with the corresponding binary encoding benchmark to form a node-independent binary sequence. The binary states of each acquisition node within the same cross-node linkage combination are concatenated according to a preset node order to form a cross-node joint binary sequence, such as according to node number, equipment topology location, communication address, or pre-configured priority. Based on the Lemotziff complexity algorithm, the complexity of the node-independent binary sequence and the cross-node joint binary sequence is calculated to obtain the node-independent Lemotziff sequence complexity and the cross-node joint Lemotziff sequence complexity. Based on the node-independent Lemotziff sequence complexity and the cross-node joint Lemotziff sequence complexity, a time-series correspondence analysis is performed in conjunction with the changes in new energy output, load, energy storage charging and discharging, and grid-side power flow to identify the spatiotemporal coupling pattern.

[0111] In this invention, physical connection relationship refers to whether there is an actual electrical connection channel between nodes, such as a photovoltaic power station connected to a feeder, an energy storage power station connected in parallel to the same bus, and an industrial load being powered by the transformer; power flow transmission relationship refers to the actual transmission direction and influence path of power in these connection channels, such as an increase in photovoltaic output affecting the power flow of the line, and affecting the bus voltage and the power supply status of the downstream load; control response relationship refers to whether, and how quickly, another node will make a control or status response after a change in one node, such as the energy storage immediately discharging to support the load after a sudden increase, or the inverter actively limiting power generation after the power flow of the line is too large.

[0112] For example, if a photovoltaic power station is connected to a feeder, and the park load and energy storage power station are connected to the downstream bus of the feeder, and the photovoltaic output decline during historical operation will cause changes in the feeder power flow and trigger the energy storage discharge, then the photovoltaic node, feeder node, park load node and energy storage node can be identified as a cross-node linkage combination.

[0113] The binary encoding benchmark can be determined by using methods such as historical averages, operating limits, quantile thresholds, expert-set thresholds, prediction deviation thresholds, or warning level thresholds. For continuous operating characteristics, such as power, voltage, frequency, temperature, and load rate, a threshold comparison method is used for encoding. For warning information, if the warning level is greater than or equal to the preset warning level, it is encoded as 1; otherwise, it is encoded as 0. For equipment fault prediction results, if the fault probability is greater than or equal to the preset value, it is encoded as 1; otherwise, it is encoded as 0.

[0114] In one embodiment, based on the Lemoziff complexity algorithm, the complexity of node-independent binary sequences and cross-node joint binary sequences is calculated to obtain the node-independent Lemoziff sequence complexity and the cross-node joint Lemoziff sequence complexity. This includes: determining the node-independent scan sequence and the cross-node joint scan sequence based on the node-independent binary sequence and the cross-node joint binary sequence; configuring the initialized Lemoziff complexity calculation parameters and scanning each node-independent scan sequence and the cross-node joint scan sequence according to the time order and scan order; comparing the currently formed binary substring with the historical substring set, and if the current binary substring is not included in the historical substring set, adding the current binary substring to the historical substring set and incrementing the complexity count by one, outputting the original Lemoziff complexity count value corresponding to each binary scan sequence; normalizing the original Lemoziff complexity count value and extracting the node-independent Lemoziff sequence complexity and the cross-node joint Lemoziff sequence complexity.

[0115] For example, the node-independent binary sequence formed by a node over 8 consecutive analysis times is: If the sequence is not explicitly defined, it can be directly treated as an independent node scan sequence. The historical substring set is initialized to empty with a complexity count of 0, and the sequence is scanned from left to right in chronological order. During scanning, the current binary substring is formed sequentially. For example, if substring 0 is obtained first, since 0 does not exist in the historical substring set, it is added to the historical substring set, and the complexity count is incremented by 1. Substring 1 is obtained by continuing the scan; if 1 does not exist in the historical substring set, it is added to the set, and the count is incremented by 1. When a subsequently formed substring already exists in the historical substring set, the substring is expanded backward until a new, previously unseen substring is formed, such as 00 or 11, which is then added to the set, and the count is incremented by 1. This yields the Lemotzius original complexity count for the independent scan sequence of the photovoltaic node. For cross-node joint sequences, assuming that at the same time, state 101 is obtained by splicing together photovoltaic, load, and energy storage in that order, multiple consecutive times form a cross-node joint binary sequence. Then, the scanning is performed in the same order of time and node concatenation. By counting the number of newly added binary substrings during the scanning process, the Lemoziv primitive complexity count of the cross-node joint scanning sequence is output, which is used to represent the complexity of the node combination linkage state.

[0116] It should be noted that the Lemotzov complexity calculation adopts the sequence parsing strategy of the standard LZ76 algorithm: starting from the first character of the sequence, the currently unparsed part is divided into the shortest new substrings in turn, that is, each time it expands to the right from the current position until the substring formed appears for the first time in the set of parsed substrings; if the entire remaining sequence is a new substring, it is counted as a whole; the scanning is strictly performed in chronological order. For cross-node joint binary sequences, each bit corresponds to a time and node unit, and the splicing order is fixed as (time 1: node A, node B, node C; time 2: node A, node B, node C; ...), ensuring that the spatiotemporal structure is explicitly preserved in the sequence.

[0117] In one embodiment, based on node-independent Lemotsiv sequence complexity and cross-node joint Lemotsiv sequence complexity, and combined with the changes in renewable energy output, load, energy storage charging and discharging, and grid-side power flow, a time-series correspondence analysis is performed to identify spatiotemporal coupling patterns. This includes: calculating operational change characteristics based on operational data from adjacent analysis windows, including changes in renewable energy output, load, energy storage charging and discharging, and grid-side power flow; performing adjacent window difference analysis, amplitude statistics, and trend extraction on node-independent and cross-node joint Lemotsiv sequence complexities to obtain complexity change characteristics; determining the temporal sequence relationship between renewable energy output fluctuations, load abrupt changes, energy storage charging and discharging responses, and grid-side power flow changes based on complexity change characteristics and operational change characteristics; calculating the corresponding response delay based on the time interval between the triggering (significant) change moment and the response (significant) change moment; and identifying the spatiotemporal coupling patterns between renewable energy output fluctuations, load abrupt changes, energy storage responses, and grid-side power flow changes based on the temporal sequence relationship, response delay, direction of change of operational change characteristics, and complexity change characteristics.

[0118] It should be noted that this invention uses the order in which significant changes occur first, the order in which significant changes occur subsequently, the time interval between the two, whether the direction of change conforms to physical mechanisms, and whether the complexity increases or decreases synchronously as the criteria for determining the coupling relationship. Specifically, within the continuous analysis window, the significant change times of the renewable energy output changes, load changes, energy storage charging and discharging changes, and grid-side power flow changes are first determined. If the renewable energy output fluctuation or load abrupt change occurs before the energy storage charging and discharging changes and grid-side power flow changes, the side that changes first can be determined as the trigger side, and the side that changes subsequently can be determined as the response side. The time interval between the significant change time of the trigger side and the significant change time of the response side is calculated to obtain the response delay; at the same time, it is determined whether the direction of change conforms to the source-grid-load-storage operation mechanism. For example, a decrease in renewable energy output usually causes an increase in grid-side compensation power flow or an increase in energy storage discharge, a sudden increase in load usually causes an increase in line power flow and triggers an energy storage discharge response, and an increase in renewable energy output may cause an increase in energy storage charging or power transmission power flow. By combining the changes in the independent Lemoziff sequence complexity of nodes and the joint Lemoziff sequence complexity across nodes, if after a triggering event occurs, the independent complexity of the relevant nodes and the joint complexity of the linkage combination increase synchronously within adjacent windows, and the timing relationship, response delay, and direction of change repeat in multiple operating segments, then it can be determined that there is a stable spatiotemporal coupling law between the fluctuation of new energy output, sudden load change, energy storage response, and grid-side power flow change.

[0119] The criteria for determining significant changes include, for example, when changes in renewable energy output, load, energy storage charging / discharging, or grid-side power flow exceed a preset threshold. This preset threshold is not set to a fixed value but is normalized based on equipment capacity, historical fluctuation levels, sampling period, and scheduling sensitivity. It typically combines a percentage threshold of rated capacity with a historical statistical threshold. In this invention, analysis windows are set at 5 minutes, 15 minutes, or 1 hour, with preset step sizes to ensure that complexity calculations and operational change calculations correspond on the time scale.

[0120] For the threshold for determining moments of significant change, a dynamic adaptive approach can be adopted, such as weighted normalization fusion based on the equipment's rated capacity and the standard deviation of that quantity over the same period in the past 7 days.

[0121] This invention encodes core features, early warning information, and equipment fault prediction results into a unified binary sequence and incorporates Lemzek complexity for calculation. This transforms multi-source heterogeneous operational data into a comparable complexity index, avoiding the difficulty of directly correlating different data types, dimensions, and nodes. By simultaneously calculating the node-independent Lemzek sequence complexity and the cross-node joint Lemzek sequence complexity, it can identify both the degree of change in the operational state of a single acquisition node and the degree of change in the coordinated state between multiple nodes, thus improving the characterization of the source-grid-load-storage coordinated operation status. Furthermore, by combining the time-series correlation analysis of changes in new energy output, load, energy storage charging and discharging, and grid-side power flow, this invention can determine the sequence, response delay, and direction of change among new energy fluctuations, load abrupt changes, energy storage responses, and power flow changes, thereby identifying spatiotemporal coupling patterns. Compared to analysis methods based solely on single node monitoring values ​​or single power changes, this invention can more comprehensively reflect the dynamic linkage between sources, grid, load, and storage, which is beneficial for improving the accuracy of new energy absorption analysis, load disturbance identification, energy storage regulation effect evaluation, and grid operation risk prediction. For example, a change in photovoltaic output exceeding a threshold is recorded as 1, otherwise as 0; a sudden increase in load exceeding a threshold is recorded as 1, otherwise as 0; and energy storage switching from standby to discharge is recorded as 1, otherwise as 0. This forms a node-independent binary sequence and a cross-node joint binary sequence. The Lemozoff complexity algorithm is then used to scan the above sequences to obtain the node-independent complexity of photovoltaic nodes, load nodes, and energy storage nodes, as well as the cross-node joint complexity of the source-grid-load-storage combination. If, within a certain period, a sudden drop in photovoltaic output occurs, the power flow of the line changes in the opposite direction, and after an analysis window, the energy storage discharge power increases, and the cross-node joint Lemozoff sequence complexity increases significantly and repeats in multiple similar scenarios, then the spatiotemporal coupling law of the decline in new energy output, the change in grid-side power flow, and the energy storage discharge response can be identified.

[0122] The collaborative sensing and control layer of this invention breaks through the limitations of isolated data from traditional front-end acquisition nodes. It constructs a multi-dimensional linkage control module for power generation, grid, load, and storage, integrating edge-processed data from grid nodes, new energy power plants, energy storage facilities, and load-side terminals. This enables edge-end collaborative sensing and linkage analysis of data across power generation, grid, load, and storage. The module includes a data interaction gateway, a collaborative sensing and analysis unit, and a linkage control decision-making unit.

[0123] 1. Data Interaction Gateway: Enables low-latency wireless data interaction between various front-end acquisition nodes. It adopts power-specific narrowband IoT (NB-IoT) combined with 5G lightweight communication technology to ensure the real-time performance and stability of data transmission.

[0124] 2. Collaborative Sensing and Analysis Unit: This unit performs cross-node and cross-link collaborative analysis on the fused analysis data, fault early warning information, and time-series prediction results from each node. It uncovers the linkage patterns between the power generation, grid, load, and storage systems, and generates a comprehensive power system operation status analysis report. The specific sub-steps of the Collaborative Sensing and Analysis Unit are as follows:

[0125] Data aggregation and alignment: The system receives fusion features, early warning information and prediction results from each node through a data interaction gateway, and performs cross-node time-series alignment and data cleaning based on a unified timestamp.

[0126] Cross-node correlation mining: Analyze the energy flow coupling relationship between the source, grid, load and storage links, and identify the linkage pattern and spatiotemporal correlation characteristics between new energy output fluctuations, load changes and energy storage response.

[0127] Comprehensive status assessment: Combining the power balance equation, assess the consistency of the operating status of all links, and detect the risk of power imbalance and abnormal transmission paths among power sources, grids, loads and storage.

[0128] Report generation and output: Summarize the operational status, potential risk points and linkage patterns of each link, generate a full-link power system operation status analysis report, and push it to the linkage control decision-making unit.

[0129] S5. Based on the spatiotemporal coupling law and time series prediction results, provide edge-end decision-making suggestions for the coordinated regulation of source, grid, load and storage.

[0130] In the joint control decision-making unit: based on the results of collaborative perception analysis and time-series forecast data, preliminary decision-making suggestions are provided at the edge for the coordinated control of power generation, grid, load and storage, such as adjusting the output of new energy power generation, optimizing the charging and discharging strategy of energy storage, and guiding the demand response on the load side. The decision-making suggestions are pushed to each control terminal in real time to realize the rapid joint control of power generation, grid, load and storage. At the same time, the decision-making suggestions are uploaded to the cloud platform to provide a reference for the global control in the cloud.

[0131] In one embodiment, a method for multi-source heterogeneous data pre-acquisition and edge intelligent processing further includes: evaluating the consistency of the operating state of a power system using a power balance equation, wherein the power balance equation is: P gen +P grid +P dis =P load +P ch +P loss In the formula, P gen For new energy power generation; P grid Power supplied to the power grid; P dis P is the energy storage discharge power; load P represents the load power. ch Power for energy storage charging; P loss This refers to line loss power.

[0132] The data interaction layer of this invention enables on-demand data interaction between the edge terminal and the cloud monitoring platform, local operation and maintenance terminals, and various control terminals. It only uploads pre-processed high-quality data, AI analysis results, fault warning information, time-series prediction data, and collaborative control decision suggestions, significantly reducing bandwidth consumption. It supports cloud command downloads (adjustment of collection frequency, optimization of algorithm parameters, modification of warning thresholds), achieving bidirectional interaction and collaborative optimization between the edge terminal and the cloud. A data caching unit is set up to cache edge-processed data during network interruptions and re-transmit it upon recovery, ensuring data integrity.

[0133] like Figure 2 As shown, according to another embodiment of the present invention, a multi-source heterogeneous data pre-acquisition and edge intelligent processing system is provided, comprising:

[0134] The sampling frequency correction module is used to determine the adaptive sampling frequency at each acquisition node based on the data change rate; by using the mode division mechanism and combining the operating status of the power equipment, the adaptive sampling frequency is corrected according to different operating conditions to obtain the corrected sampling frequency.

[0135] The feature acquisition module is used to collect multi-source heterogeneous data according to the corrected sampling frequency; preprocess and aggregate the multi-source heterogeneous data to obtain standardized aggregated data; convert the standardized aggregated data into multi-dimensional feature vectors; and extract the fusion core features from the multi-dimensional feature vectors.

[0136] The information acquisition module is used to input the core features into the general fault judgment model and the time series prediction model, and output the equipment fault prediction results and the time series prediction results respectively; based on the equipment fault prediction results, combined with the early warning mechanism, early warning information is output.

[0137] The coupling law acquisition module is used to construct a multi-node joint operation sequence based on the fusion core features, early warning information and equipment fault prediction results of each acquisition node. It introduces the Lemotzov complexity theory to uniformly quantify the complexity of independent node and cross-node linkage sequences in order to identify the spatiotemporal coupling law between new energy output fluctuations, load mutations and energy storage response.

[0138] The edge-end decision suggestion module is used to provide edge-end decision suggestions for coordinated regulation of source, grid, load and storage based on spatiotemporal coupling laws and time series prediction results.

[0139] This invention relates to power system data acquisition and edge computing, applicable to multi-source heterogeneous data processing in power scenarios such as new energy power plants, energy storage facilities, grid nodes, and load-side terminals. It is a core technology for achieving coordinated regulation and control of power generation, grid, load, and storage, and intelligent operation and maintenance of the power grid. This invention constructs a lightweight edge intelligent processing model by integrating artificial intelligence technology and multimodal perception modules to achieve pre-processing of multi-source heterogeneous data, multimodal correlation analysis, accurate fault early warning, and coordinated perception of power generation, grid, load, and storage, effectively improving the efficiency of power system data processing and the level of power grid operation and maintenance regulation.

[0140] The purpose of this invention is to provide an AI-driven solution for multi-source heterogeneous data pre-collection and edge intelligent processing. It integrates artificial intelligence technology to construct a lightweight edge intelligent processing model, combining multi-modal perception modules such as electrical parameter acquisition, environmental perception, and equipment status monitoring to achieve pre-collection of multi-dimensional data. Utilizing AI multi-modal fusion algorithms and edge computing capabilities, it performs real-time cleaning, standardization, and aggregation of multi-source heterogeneous data at the pre-collection end, reducing bandwidth consumption and improving data quality. Through multi-modal data correlation analysis, it achieves accurate prediction and real-time early warning of equipment faults, realizing the integration of data acquisition, status monitoring, and fault early warning. Embedding time-series prediction algorithms enables accurate short-cycle prediction of power generation output and load demand. Simultaneously, it constructs a multi-dimensional linkage control module for power generation, grid, load, and storage, achieving collaborative perception of data from all links. This provides low-latency, high-precision data support for coordinated control of power generation, grid, load, and storage, breaking the limitations of traditional pre-collection methods that only involve data acquisition and forwarding. It promotes the deep integration of AI and power pre-collection, improving the data processing efficiency of the power system and the level of precision in grid operation and maintenance.

[0141] This invention is based on a four-layer architecture consisting of a multimodal perception layer, an edge intelligent processing layer, a collaborative perception and control layer, and a data interaction layer. It integrates power system data acquisition, intelligent processing, fault early warning, time-series prediction, and collaborative control. The overall architecture design is as follows: Figure 5 As shown.

[0142] The multimodal perception layer completes comprehensive front-end acquisition of heterogeneous data from multiple sources, including power, environment, equipment, new energy, energy storage, and load. The edge intelligent processing layer, as the core layer, uses lightweight AI models to achieve real-time data cleaning, standardization, aggregation, multimodal correlation analysis, fault early warning, and time-series prediction. The collaborative perception and control layer constructs a multi-dimensional linkage control module for power generation, grid, load, and storage, enabling collaborative perception and analysis of data from all stages. The data interaction layer enables low-latency data interaction between the edge and the cloud, and between various acquisition nodes, while simultaneously uploading high-quality processed data on demand, reducing bandwidth consumption. These layers work collaboratively to achieve integrated capabilities for data acquisition, intelligent processing, fault early warning, time-series prediction, and collaborative control, comprehensively improving the intelligence level of power front-end data acquisition.

[0143] The core of this invention's AI-driven multi-source heterogeneous data pre-acquisition and edge intelligent processing solution lies in the deep integration of artificial intelligence technology with power pre-acquisition, constructing a lightweight edge intelligent processing model, integrating multimodal perception modules to achieve comprehensive pre-acquisition of multi-source heterogeneous data; performing real-time data cleaning, standardization, and aggregation processing at the edge to reduce bandwidth consumption; utilizing AI multimodal fusion algorithms to mine the correlation patterns of multi-dimensional data to achieve accurate prediction and real-time early warning of equipment faults; embedding lightweight time-series prediction algorithms to achieve accurate short-cycle prediction of power generation output and load demand; and constructing a multi-dimensional linkage control module for power generation, grid, load, and storage to achieve collaborative perception and linkage analysis of data from all links. Key points include: the integration and synchronous acquisition method of multimodal perception modules, the lightweight AI multimodal fusion analysis model adapted to the edge, the pre-acquisition edge intelligent processing flow of multi-source heterogeneous power data, the equipment fault classification and early warning method based on multimodal data, and the edge collaborative perception mechanism for multi-dimensional linkage control of power generation, grid, load, and storage.

[0144] In summary, this invention integrates a multi-modal perception module and synchronous acquisition control mechanism for power parameter acquisition, environmental perception, equipment status monitoring, and multi-source power acquisition through an AI-driven multi-source heterogeneous power data front-end acquisition method; a method and process for real-time cleaning, standardization, and aggregation processing of multi-source heterogeneous power data at the edge; a lightweight AI multi-modal fusion analysis model adapted to edge computing power and a multi-source power data correlation analysis method; a scheme for accurate prediction and graded early warning of power equipment faults based on multi-modal fusion data; a lightweight time-series prediction algorithm embedded at the edge and a short-cycle prediction method for new energy power generation output and load demand; a multi-dimensional linkage control module for power generation, grid, load, and storage and an edge-end collaborative perception and linkage analysis mechanism for data in each link; and an on-demand data interaction method between the edge and cloud for multi-source heterogeneous power data, enabling the targeted uploading of high-quality data and bidirectional interaction of cloud commands.

[0145] This invention provides a method for front-end acquisition and edge intelligent processing of multi-source heterogeneous data. It integrates artificial intelligence technology to construct a lightweight edge intelligent processing model, combining multi-modal perception modules such as electrical parameter acquisition, environmental perception, and equipment status monitoring to achieve front-end acquisition of multi-dimensional data. Utilizing AI multi-modal fusion algorithms and edge computing capabilities, it performs real-time cleaning, standardization, and aggregation of multi-source heterogeneous data at the front-end acquisition end, reducing bandwidth consumption and improving data quality. Through multi-modal data correlation analysis, it achieves accurate prediction and real-time early warning of equipment faults, realizing the integration of data acquisition, status monitoring, and fault early warning. Embedded time-series prediction algorithms enable accurate short-cycle prediction of power generation output and load demand. Simultaneously, it constructs a multi-dimensional linkage control module for power generation, grid, load, and storage, achieving collaborative perception of data from all links. This provides low-latency, high-precision data support for coordinated control of power generation, grid, load, and storage, breaking the limitations of traditional front-end acquisition which only involves data acquisition and forwarding. It promotes the deep integration of AI and power front-end acquisition, improving the data processing efficiency of the power system and the level of refined grid operation and maintenance.

[0146] in:

[0147] Alternatives to lightweight AI models: If edge computing power is extremely limited, lightweight multimodal fusion neural networks can be replaced with improved combinations of traditional machine learning algorithms (such as XGBoost, random forest, and association rule mining). By combining algorithms, association analysis of multimodal data can be achieved. Although the feature mining depth is slightly reduced, it can still achieve device fault prediction and time series prediction, meeting the basic edge intelligent processing needs.

[0148] Alternative communication technologies: For collaborative sensing data interaction between various data acquisition nodes, if deployment conditions are limited, power line carrier communication technology can be used to achieve data transmission using power lines. This eliminates the need for additional communication lines and still ensures the real-time performance and stability of data interaction, enabling collaborative sensing of power sources, grids, loads, and storage.

[0149] Alternatives to data preprocessing algorithms: In the outlier detection stage, if the Isolation Forest algorithm is complex to deploy at the edge, it can be replaced by the improved 3σ rule and sliding window method based on statistical features. By combining the temporal characteristics of the data for outlier identification, efficient data cleaning can still be achieved and data quality can be guaranteed.

[0150] Alternatives to time series prediction algorithms: If the deployment of hybrid models is difficult, they can be replaced with improved deep learning algorithms and long short-term memory methods, optimized for the characteristics of new energy power generation and load power consumption, which can still achieve accurate predictions for short cycles and provide data support for coordinated regulation.

[0151] All of the above alternative solutions can achieve the core purpose of this invention under different hardware and deployment environments, realize the front-end acquisition and edge intelligent processing of multi-source heterogeneous data, fault early warning, time series prediction and source-grid-load-storage collaborative perception, and can be flexibly selected according to the computing power, communication and deployment conditions of the actual application scenario.

[0152] In this invention, multi-source heterogeneous data refers to datasets originating from different acquisition nodes and sensing modules, exhibiting differences in data type, dimension, format, and dimensionality. This includes analog power quantities, digital equipment status quantities, environmental sensing switch quantities, and time-series data of new energy power generation. Edge intelligent processing refers to deploying artificial intelligence models and computing resources at the edge (front-end acquisition node) near the data acquisition source to achieve local intelligent processing of data, without relying on centralized computing power in the cloud, and featuring low latency and high real-time performance. Multimodal fusion analysis refers to combining multimodal data of different types and dimensions (such as electrical parameters, environmental parameters, etc.) into a single dataset. This invention deeply integrates data from various sources, including equipment status data, to uncover the inherent correlations between data points and achieve more comprehensive and accurate data analysis. Lightweight neural networks, designed for edge computing and storage resources, are improved models by pruning, quantizing, and lightweighting traditional neural networks. This significantly reduces computing and storage consumption while maintaining model performance, making them suitable for edge deployment. Time-series forecasting predicts future trends based on historical patterns of time-series data and relevant influencing factors. In this invention, short-cycle time-series forecasting is primarily used for new energy power generation output and load electricity demand.

[0153] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multi-source heterogeneous data pre-acquisition and edge intelligent processing, characterized in that, include: At each acquisition node, an adaptive sampling frequency is determined based on the data change rate; By utilizing the mode division mechanism and combining the operating status of power equipment, the adaptive sampling frequency is modified according to different operating conditions to obtain the modified sampling frequency; Based on the corrected sampling frequency, multi-source heterogeneous data is collected; the multi-source heterogeneous data is preprocessed and aggregated to obtain standardized aggregated data, and the standardized aggregated data is converted into a multi-dimensional feature vector. The core fusion features are extracted from the multi-dimensional feature vector. By integrating the core features into the general fault diagnosis model and the time series prediction model, the equipment fault prediction results and the time series prediction results are output respectively. Based on the equipment failure prediction results and combined with the early warning mechanism, early warning information is output; Based on the core features of each acquisition node, early warning information and equipment fault prediction results, a multi-node joint operation sequence is constructed. The Lemotzov complexity theory is introduced to unify and quantify the complexity of independent node and cross-node linkage sequences in order to identify the spatiotemporal coupling law between new energy output fluctuations, load mutations and energy storage response. Based on the spatiotemporal coupling law and time series prediction results, edge-end decision-making suggestions are provided for the coordinated regulation of source, grid, load and storage. The method of using a mode division mechanism, combined with the operating status of power equipment, to perform condition-specific correction on the adaptive sampling frequency to obtain the corrected sampling frequency includes: Obtain a standardized set of power equipment operation information, and perform sliding time window processing on the operation quantities of various types of power equipment in the standardized power equipment operation information set; The fluctuation intensity is calculated based on the degree of deviation of the mean within the sliding time window, and the abrupt change characteristics are calculated based on the jump variables between the current value and the historical window boundary values. By normalizing and weighting the fluctuation intensity and abrupt change characteristics, a comprehensive fluctuation index is obtained to characterize the degree of dynamic change in power equipment operation data. The comprehensive volatility index is bound to the corresponding data source label, time window label, and monitoring quantity label to output a set of operational change characteristics; The load rate, temperature, and alarms are explicitly referenced from the set of standardized power equipment operation information from the set of operational change characteristics; based on the normalized values ​​of load rate, temperature, and alarms, load risk factors, temperature rise risk factors, and alarm risk factors are generated. The load risk factor, temperature rise risk factor, and alarm risk factor are weighted and integrated to obtain the equipment operation risk factor; The equipment operation risk factors are linked to the comprehensive volatility index, and the output is a set of equipment status risk characteristics; Based on the metastable residence theory, and combined with the comprehensive volatility index judgment process and the equipment operation risk factor judgment process, the modes of power equipment are determined, including pure steady state, metastable state and fault state. Different sampling frequency corrections are assigned to the pure steady state, metastable state, and fault state, and the sampling frequency corrections increase sequentially from the pure steady state to the metastable state and the fault state. Based on the current dwell time of power equipment in a continuous metastable state, the sampling frequency correction amount for the metastable state is adjusted; Based on the sampling frequency correction values ​​obtained under pure steady-state, metastable, and fault states, the adaptive sampling frequency is corrected according to different operating conditions to obtain the corrected sampling frequency.

2. The method for multi-source heterogeneous data pre-acquisition and edge intelligent processing according to claim 1, characterized in that, The determination of the adaptive sampling frequency based on the data change rate includes: The adaptive sampling frequency is calculated by dividing the data change rate by the data change threshold and multiplying by the sensitivity coefficient, combined with the base sampling frequency and the device status weight factor.

3. The method for multi-source heterogeneous data pre-acquisition and edge intelligent processing according to claim 1, characterized in that, The determination of the modes of power equipment based on metastable residence theory, combined with the comprehensive volatility index judgment process and the equipment operation risk factor judgment process, includes: Configure the steady-state threshold and fault threshold of the comprehensive volatility index, as well as the first risk threshold and the second risk threshold of the equipment operation risk factor, wherein the steady-state threshold is less than the fault threshold and the first risk threshold is less than the second risk threshold; When the comprehensive volatility index is less than or equal to the steady-state threshold, the equipment operation risk factor is less than or equal to the first risk threshold, and there is no alarm state of the current power equipment, the operating mode of the power equipment is determined to be pure steady state. When the comprehensive volatility index is between the steady-state threshold and the fault threshold, the equipment operation risk factor is between the first risk threshold and the second risk threshold, and there is no alarm state of the current power equipment, the operating mode of the power equipment is determined to be metastable. When at least one of the following conditions occurs: the comprehensive volatility index is greater than or equal to the fault threshold, the equipment operation risk factor is greater than or equal to the second risk threshold, or the current power equipment is in an alarm state, the operating mode of the power equipment is determined to be a fault state.

4. The method for multi-source heterogeneous data pre-acquisition and edge intelligent processing according to claim 1, characterized in that, The adjustment of the metastable sampling frequency correction based on the current dwell time of the power equipment in a continuous metastable state includes: If the current operating mode of the power equipment is metastable, and the operating mode of the previous judgment period was also metastable, update the current metastable dwell time; If the current operating mode of the power equipment is metastable and the operating mode of the previous judgment period is not metastable, the current metastable dwell time is initialized to the preset judgment period duration. If the current operating mode of the power equipment is not metastable, the metastable dwell time will be reset to zero. The sampling frequency correction for metastable states is adjusted based on the principle that the metastable state sampling frequency correction gradually decreases with the increase of the metastable state residence time and is higher than the minimum metastable state correction.

5. The method for multi-source heterogeneous data pre-acquisition and edge intelligent processing according to claim 1, characterized in that, The process involves preprocessing and aggregating multi-source heterogeneous data to obtain standardized aggregated data, then converting the standardized aggregated data into a multi-dimensional feature vector. From this multi-dimensional feature vector, core fusion features are extracted, including: Multi-source heterogeneous data is sequentially cleaned, standardized, and aggregated to obtain standardized aggregated data. Based on a multimodal fusion neural network model, standardized multimodal data is encoded into multidimensional feature vectors; The attention weights between multi-dimensional feature vectors are calculated through a multi-head attention mechanism to dynamically capture the correlation. A unified fused feature matrix is ​​generated through a feature fusion layer, and the core fused features are extracted through a feature dimensionality reduction layer.

6. The method for multi-source heterogeneous data pre-acquisition and edge intelligent processing according to claim 1, characterized in that, Based on the fusion of core features, early warning information, and equipment fault prediction results from each acquisition node, a multi-node joint operation sequence is constructed. This sequence incorporates Lemzef complexity theory to uniformly quantify the complexity of independent node and cross-node linkage sequences, thereby identifying the spatiotemporal coupling patterns between new energy output fluctuations, load abrupt changes, and energy storage response. This includes: Based on the physical connection relationship, power flow transmission relationship, and regulation response relationship among the source, grid, load, and storage, cross-node linkage combinations are determined. The core features, early warning information, and equipment fault prediction results of the same cross-node linkage combination and at the same analysis time are spliced ​​together according to the node order of source, grid, load and storage, and arranged in chronological order to output the multi-node joint operation sequence of source, grid, load and storage. For various fusion core features, early warning information and equipment fault prediction results in the multi-node joint operation sequence, a binary coding benchmark is determined; the fusion core features, early warning information and equipment fault prediction results of each acquisition node are compared with the corresponding binary coding benchmark to form a node-independent binary sequence. The binary states of each acquisition node within the same cross-node linkage combination are concatenated according to a preset node order to form a cross-node joint binary sequence; Based on the Lemozoff complexity algorithm, the complexity of node-independent binary sequences and cross-node joint binary sequences is calculated to obtain the complexity of node-independent Lemozoff sequences and cross-node joint Lemozoff sequences. Based on node-independent Lemotsev sequence complexity and cross-node joint Lemotsev sequence complexity, time-series correspondence analysis is performed in conjunction with changes in new energy output, load, energy storage charging and discharging, and grid-side power flow to identify spatiotemporal coupling patterns.

7. The method for multi-source heterogeneous data pre-acquisition and edge intelligent processing according to claim 6, characterized in that, The Lermeziff complexity algorithm is used to calculate the complexity of node-independent binary sequences and cross-node joint binary sequences, yielding the complexity of node-independent Lermeziff sequences and the complexity of cross-node joint Lermeziff sequences, including: The node-independent binary sequence and the cross-node joint binary sequence are used to determine the node-independent scan sequence and the cross-node joint scan sequence. Configure the initial Lermeziff complexity calculation parameters, and scan each node's independent scan sequence and cross-node joint scan sequence according to the time order and scan order; compare the currently formed binary substring with the historical substring set, and if the current binary substring is not included in the historical substring set, add the current binary substring to the historical substring set and increment the complexity count by one, and output the original Lermeziff complexity count value corresponding to each binary scan sequence; The original Lemoziff complexity count is normalized, and the node-independent Lemoziff sequence complexity and cross-node joint Lemoziff sequence complexity are extracted.

8. The method for multi-source heterogeneous data pre-acquisition and edge intelligent processing according to claim 6, characterized in that, The method employs time-series correspondence analysis based on node-independent and cross-node joint Lemotsev sequence complexity, combined with changes in new energy output, load, energy storage charging and discharging, and grid-side power flow, to identify spatiotemporal coupling patterns, including: Based on the operational data of adjacent analysis windows, the operational change characteristics are calculated, including the changes in new energy output, load, energy storage charging and discharging, and grid-side power flow. The complexity variation characteristics are obtained by performing adjacent window difference, variation amplitude statistics and variation trend extraction on the node-independent Lemozoff sequence complexity and cross-node joint Lemozoff sequence complexity. Based on the characteristics of complexity and operation changes, the temporal sequence relationship between new energy output fluctuations, load abrupt changes, energy storage charging and discharging responses, and grid-side power flow changes is determined; the corresponding response delay is calculated based on the time interval between the triggering-side change time and the response-side change time. Based on the temporal sequence, response delay, direction of change of operational characteristics, and complexity change characteristics, the spatiotemporal coupling law between new energy output fluctuations, load abrupt changes, energy storage response, and grid-side power flow changes is identified.

9. The method for multi-source heterogeneous data pre-acquisition and edge intelligent processing according to claim 1, characterized in that, It also includes: using the power balance equation to assess the consistency of the operating state of the power system, wherein the power balance equation is: P gen +P grid +P dis =P load +P ch +P loss ; In the formula, P gen For new energy power generation; P grid Power supplied to the power grid; P dis P is the energy storage discharge power; load P represents the load power. ch Power for energy storage charging; P loss This refers to line loss power.

Images