Cloud edge-based distributed new energy power generation regulation capability identification method and system

Abstract

The application provides a cloud edge end-based distributed new energy power generation regulation capacity identification method and system, and belongs to the technical field of new energy power generation monitoring. The identification method is applied to a corresponding identification system, and specifically comprises the following steps: forming a cloud edge end architecture composed of edge end devices, edge nodes and cloud platforms based on the hierarchical topology of distributed new energy power generation units; extracting first-level basic features by the edge end devices based on the multi-source historical operation data of the preprocessed distributed new energy power generation units; extracting corresponding second-level intermediate features through the edge nodes; extracting corresponding third-level deep features by the cloud platforms; training a regulation capacity identification model based on the third-level deep features and in combination with a corresponding preset regulation capacity label; and outputting a regulation capacity identification result in combination with real-time operation data of the distributed new energy power generation units. The cloud edge end architecture is used for progressive feature extraction, which improves the information carrying capacity and representation ability, so as to improve the identification accuracy.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation monitoring technology, and in particular to a method and system for identifying the regulation capacity of distributed new energy power generation based on cloud-edge-device architecture. Background Technology

[0002] The adjustability of distributed renewable energy generation, namely its ability to rapidly adjust active and reactive power under grid dispatch commands, mainly includes indicators such as active power ramp rate, maximum adjustable capacity, adjustment response time, continuous adjustment duration, tolerance range, and adjustment accuracy. These indicators directly determine the ability of distributed renewable energy to participate in grid frequency regulation, voltage regulation, and reserve services, and are key to ensuring grid frequency stability, voltage compliance, and supply-demand balance. In actual operation, it is often necessary to identify the adjustability characteristics of renewable energy generation to assist the dispatch center in formulating more reasonable power generation plans and reserve configuration schemes, thereby reducing wind and solar curtailment after renewable energy is connected to the grid and improving the grid's capacity to accommodate renewable energy.

[0003] As the scale of distributed renewable energy generation expands and its application scenarios become more complex, its operational characteristics exhibit significant multi-dimensional complexity. On the one hand, distributed renewable energy generation units are widely distributed, with small individual unit capacities and a large number of units, resulting in massive, heterogeneous, and dispersed data collection. On the other hand, its output is highly susceptible to natural environmental influences; parameters such as wind speed, solar intensity, and temperature fluctuate randomly, causing power generation to exhibit strong volatility, intermittency, and unpredictability. This randomness in distribution and strong fluctuation in output further increases the difficulty of identifying the regulation capabilities of renewable energy generation.

[0004] Existing methods for identifying the regulation capacity of new energy power generation mainly fall into two categories: centralized identification methods and distributed shallow identification methods. The former uploads the raw operating data of all distributed new energy power generation units to a cloud-based dispatch center via a communication network, where the cloud centrally processes the data, extracts features, and trains models, ultimately outputting the adjustable characteristic parameters of all units. Examples include parameter identification based on least squares and support vector machine identification methods. The latter deploys simple identification models on edge devices or regional edge nodes, performing local identification based on locally collected real-time data, without requiring the uploading of large amounts of raw data. Examples include rule-based threshold judgment methods and lightweight neural network local prediction methods.

[0005] However, the regulation capacity of distributed renewable energy generation is not determined by a single physical quantity, but rather by the coupled effects of multiple factors, including environmental parameters, equipment operating status, and grid constraints. Furthermore, distributed renewable energy generation exhibits a hierarchical distribution from local to regional levels, and then to the entire grid, with strong coupling and correlation among the regulation capacities of each level. Most current identification methods rely solely on shallow, single-dimensional features, and the feature extraction process is often parallel rather than progressive. This results in insufficient feature information carrying capacity and a tendency for feature redundancy and the masking of effective information, leading to low accuracy in identifying the regulation capacity of distributed renewable energy generation. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of related technologies in identifying the regulation capability of distributed new energy power generation, such as single feature extraction dimensions, lack of progressiveness, insufficient information carrying capacity and representation ability of extracted features, and low identification accuracy. This invention provides a cloud-edge-device-based method and system for identifying the regulation capability of distributed new energy power generation. By forming a cloud-edge-device architecture adapted to the hierarchical distribution characteristics of distributed new energy, progressive feature extraction is achieved, comprehensively capturing the multi-scale operating rules of distributed new energy power generation, improving the information carrying capacity and representation ability of features, and then, in conjunction with the regulation capability identification model, realizing the identification of the regulation capability of distributed new energy power generation, effectively improving the identification accuracy.

[0007] The objective of this invention is achieved through the following technical solution: A method for identifying the regulation capability of distributed new energy power generation based on cloud-edge-device architecture includes: Based on all distributed new energy power generation units within the target area, a cloud-edge architecture consisting of edge devices, edge nodes, and a cloud platform is formed according to the corresponding hierarchical topology. Collect multi-source historical operation data from each distributed new energy power generation unit, and construct corresponding standardized datasets by combining preprocessing. Based on the standard dataset, the edge devices extract the first-level basic features, the edge nodes extract the second-level intermediate features based on the corresponding first-level basic features, and the cloud platform extracts the third-level deep features based on the corresponding second-level intermediate features. Based on three levels of deep features, a regulatory ability identification model is trained by combining corresponding preset regulatory ability labels. Based on the real-time operation data of each distributed new energy power generation unit, and combined with the trained regulation capability identification model, the regulation capability identification results are output.

[0008] Furthermore, based on the standard dataset, the edge devices extract primary basic features, the edge nodes extract secondary intermediate features based on the corresponding primary basic features, and the cloud platform extracts tertiary deep features based on the corresponding secondary intermediate features, including: Based on the standard dataset, edge devices extract primary basic features, including time-series statistical features, environmental power mapping features, and local response features, using corresponding feature extraction algorithms. Edge nodes receive and integrate corresponding first-level basic features based on the corresponding hierarchical topology association, and extract second-level intermediate features including power synchronization features, load output matching features, and regional operating condition consistency features. Based on all the extracted secondary intermediate features, combined with power grid operation data, the cloud platform extracts tertiary deep features, including spatiotemporal distribution features, power grid constraint and adjustability correlation features, and long-term trend features.

[0009] Furthermore, the training of the regulatory ability identification model based on three levels of deep features and corresponding preset regulatory ability labels includes: A training dataset is constructed based on three levels of deep features and corresponding preset adjustment capability labels. The training dataset is input into the regulation capability identification model, which includes a feature encoding layer, a parameter prediction layer, and an error correction layer. The feature encoding layer performs associative encoding on the three-level deep features in the input training dataset to obtain the corresponding encoded feature vector; The parameter prediction layer maps the encoded feature vectors to predicted values ​​of modulating capacity; The error correction layer calculates the prediction error of the predicted value of the regulation capability based on the corresponding preset regulation capability label, and optimizes the parameters of the parameter prediction layer and the feature encoding layer based on the prediction error through the particle swarm optimization algorithm until the iteration termination condition is met.

[0010] Furthermore, the adjustment capability identification results include single-device adjustable characteristic parameters at the edge device layer, overall adjustable characteristic parameters of the cluster at the edge node layer, and overall adjustable characteristic parameters of the regional hybrid cluster at the cloud platform layer.

[0011] Furthermore, after outputting the adjustment capability identification results, the following is also performed: The cloud platform layer uses a dual-trigger mechanism of timed triggering and event triggering to dynamically iterate and update the regulation capability identification model, and then sends the model parameters of the dynamically iterated and updated regulation capability identification model to the edge node layer and the edge device layer.

[0012] Furthermore, the cloud platform layer dynamically iteratively updates the adjustment capability identification model based on a dual-trigger mechanism of timed triggering and event triggering, including: Pre-set timed triggering rules and event triggering rules, and the cloud platform layer collects and outputs monitoring data after the regulation capability identification results in real time. The monitoring data includes identification error, power grid operating conditions and power generation unit access status. When the time interval after outputting the regulation capability identification result meets the timed triggering rule, or when the monitoring data meets the corresponding event triggering rule, the newly generated third-level deep features are extracted, and the corresponding regulation capability labels are added in combination with the monitoring data. An incremental training dataset is constructed in combination with the training dataset used during model training. The regulatory capacity identification model is dynamically iteratively updated based on the incremental training dataset.

[0013] Furthermore, the cloud-edge architecture, based on all distributed new energy power generation units within the target area and arranged according to the corresponding hierarchical topology, consists of edge devices, edge nodes, and a cloud platform, including: Determine the distribution information of each distributed new energy power generation unit within the target area, and set up corresponding edge devices at each distributed new energy power generation unit to form an edge device layer; Based on the distribution information of each distributed new energy power generation unit, identify the new energy aggregation points in the target area, and set corresponding edge nodes at each new energy aggregation point to form an edge node layer; A cloud platform is set up at the power grid dispatch center corresponding to the target area to form a cloud platform layer; The cloud-edge architecture is formed by using the edge device layer as the bottom layer of the topology, the edge node layer as the middle layer of the topology, and the cloud platform layer as the top layer of the topology.

[0014] Furthermore, the multi-source historical operating data includes the active power, reactive power, bus voltage, output current, environmental parameters, and control command signals of each distributed new energy power generation unit, as well as the corresponding regional load and tie-line power data on the grid side.

[0015] A distributed new energy power generation regulation capability identification system based on cloud-edge-device architecture is used to execute any of the above identification methods, including: The edge device layer includes several edge devices set at each distributed new energy power generation unit, which are used to collect multi-source historical operation data and real-time operation data of the corresponding distributed new energy power generation unit and extract primary basic features; The edge node layer includes several edge nodes set up at the new energy aggregation point, which are used to receive the first-level basic features of each distributed new energy power generation unit under the corresponding new energy aggregation point, so as to extract the corresponding second-level intermediate features. The cloud platform layer includes a cloud platform, which receives secondary intermediate features from each edge node to extract corresponding tertiary deep features. Based on the tertiary deep features, it trains a regulation capability identification model in combination with corresponding preset regulation capability labels. Alternatively, it outputs regulation capability identification results based on the real-time operation data of each distributed new energy power generation unit and the trained regulation capability identification model.

[0016] Furthermore, the cloud platform layer also includes: The update and iteration module is used to dynamically update the regulation capability identification model based on a dual triggering mechanism of timed triggering and event triggering, and to send the model parameters of the dynamically updated regulation capability identification model to the edge node layer and the edge device layer.

[0017] The beneficial effects of this invention are: (1) By forming a cloud-edge-device architecture adapted to the hierarchical distribution characteristics of distributed new energy, a progressive extraction of primary basic features, secondary intermediate features, and tertiary deep features is carried out. This comprehensively captures the multi-scale operation patterns of distributed new energy power generation at the local, regional, and grid-wide levels, covering multi-dimensional information such as time-series statistics, unit collaboration, and grid coupling, effectively improving the information carrying capacity and representational ability of the features. Based on the optimized features, the regulation energy identification model is combined to identify the regulation capacity of distributed new energy power generation, thereby effectively improving the accuracy of the identification results.

[0018] (2) A hybrid model combining feature encoding, parameter prediction and error correction is used as the regulation ability identification model. The training and learning of the regulation ability identification model is achieved by combining the parameter fine-tuning of particle swarm optimization, which effectively reduces the identification error of regulation ability and improves the accuracy of identification results.

[0019] (3) The resulting cloud-edge-end task layering process, after the edge devices and edge nodes complete feature extraction, only uploads high-dimensional features instead of the original data, which can effectively reduce the amount of data transmission and latency, meet the real-time requirements of power grid dispatch, and the cloud platform only performs deep feature extraction, model training and application, which can effectively reduce the risk of single node computing power overload and realize the optimized allocation of computing power and communication resources.

[0020] (4) The incremental iterative update of the regulation capacity identification model is realized based on the dual triggering mechanism of timing and event, so as to quickly adapt to the scenarios of new energy fluctuations, grid status changes and new power generation unit access, and improve the adaptability of application scenarios. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a process of the present invention; Figure 2 This is a schematic diagram of the structure of an identification system according to an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] Example: A method for identifying the regulation capability of distributed new energy power generation based on cloud-edge-device architecture, such as Figure 1 As shown, it includes: Based on all distributed new energy power generation units within the target area, a cloud-edge architecture consisting of edge devices, edge nodes, and a cloud platform is formed according to the corresponding hierarchical topology. Collect multi-source historical operation data from each distributed new energy power generation unit, and construct corresponding standardized datasets by combining preprocessing. Based on the standard dataset, the edge devices extract the first-level basic features, the edge nodes extract the second-level intermediate features based on the corresponding first-level basic features, and the cloud platform extracts the third-level deep features based on the corresponding second-level intermediate features. Based on three levels of deep features, a regulatory ability identification model is trained by combining corresponding preset regulatory ability labels. Based on the real-time operation data of each distributed new energy power generation unit, and combined with the trained regulation capability identification model, the regulation capability identification results are output.

[0024] By forming a cloud-edge-device architecture adapted to the hierarchical distribution characteristics of distributed new energy, progressive extraction of primary basic features, secondary intermediate features, and tertiary deep features is performed sequentially. This comprehensively captures the multi-scale operation patterns of distributed new energy power generation at the local, regional, and grid-wide levels. It can cover multi-dimensional information such as time-series statistics, unit collaboration, and grid coupling, effectively improving the information carrying capacity and representational capacity of features used for regulation capacity identification, thereby improving the accuracy of subsequent regulation capacity identification.

[0025] The regulation capability of distributed renewable energy generation is the result of the coupling of multiple dimensions, including the characteristics of the equipment itself, regional operational coordination, and grid constraints. Its operational patterns and adjustability cannot be fully represented by a single local dimension, but rather manifested in three interconnected and layered dimensions: local, regional, and grid-wide. Furthermore, the grid's need to identify regulation capability covers multiple levels, including single-equipment control, regional cluster scheduling, and grid-wide ancillary service configuration. Simultaneously, distributed renewable energy generation exhibits a natural hierarchical distribution from local single-equipment to regional aggregation point clusters, and then to the entire grid. The output and regulation behaviors at each level are closely coordinated and constrained. A single-dimensional characteristic can only reflect the operational status of a certain level and cannot fully characterize the multi-dimensional patterns of regulation capability formation and change. Therefore, a cloud-edge architecture is pre-constructed based on the actual hierarchical topology of distributed new energy power generation units. The edge corresponds to a single device, the edge corresponds to a regional aggregation cluster, and the cloud corresponds to the whole network scheduling. This is to match the multi-scale feature mining paths of local, regional, and whole network, so that the edge can extract basic features nearby, the edge can integrate regional collaborative features, and the cloud can integrate the whole network constraint features. Each layer's computing tasks, data sources, and feature outputs have their corresponding physical carriers and transmission paths, which not only ensures that multi-scale features can be completely extracted, but also avoids the communication and computing power pressure caused by directly uploading massive amounts of raw data.

[0026] Specifically, the cloud-edge architecture, based on all distributed new energy power generation units within the target area and arranged according to corresponding hierarchical topologies, consists of edge devices, edge nodes, and a cloud platform, including: Determine the distribution information of each distributed new energy power generation unit within the target area, and set up corresponding edge devices at each distributed new energy power generation unit to form an edge device layer; Based on the distribution information of each distributed new energy power generation unit, identify the new energy aggregation points in the target area, and set corresponding edge nodes at each new energy aggregation point to form an edge node layer; A cloud platform is set up at the power grid dispatch center corresponding to the target area to form a cloud platform layer; The cloud-edge architecture is formed by using the edge device layer as the bottom layer of the topology, the edge node layer as the middle layer of the topology, and the cloud platform layer as the top layer of the topology.

[0027] The distribution information includes the actual location, quantity, and connection relationship of each distributed renewable energy power generation unit. Based on the distribution information, corresponding edge devices are set up at the actual location of each distributed renewable energy power generation unit, thus forming the bottom edge device layer. This layer is used to collect the raw operating data of the corresponding distributed renewable energy power generation unit nearby and complete the basic feature extraction, realizing on-site processing at the device level.

[0028] Based on the spatial layout and access methods of each distributed renewable energy generation unit, and according to geographical concentration, electrical proximity access relationships, and line aggregation paths, multiple distributed renewable energy generation units that are spatially adjacent and connected to the grid through the same outgoing line or the same busbar are divided into a regional group. The common access node or line aggregation node corresponding to the regional group is regarded as the renewable energy aggregation point of the region. After identifying the renewable energy aggregation points of each regional group in the target area, edge nodes are deployed at each renewable energy aggregation point to form an intermediate edge node layer, which is responsible for receiving and integrating data from multiple edge devices within its jurisdiction and extracting regional collaborative features.

[0029] Then, a cloud platform is set up in the corresponding power grid dispatch center to form the top cloud platform layer, which is responsible for global deep feature extraction, model training, dispatch decision-making and model iterative update.

[0030] Finally, with the edge device layer as the bottom layer, the edge node layer as the middle layer, and the cloud platform layer as the top layer, a three-level cloud-edge-device topology architecture is formed according to hierarchical subordination and data flow. This allows the entire process of identifying regulation capabilities to match the hierarchical characteristics of distributed new energy, thus supporting progressive feature extraction and high-precision regulation capability identification.

[0031] The acquisition of corresponding multi-source historical operating data is achieved by the edge devices installed in each distributed new energy power generation unit. The multi-source historical operating data includes the active power, reactive power, bus voltage, output current, environmental parameters and control command signals of each distributed new energy power generation unit, as well as the regional load and tie-line power data of the corresponding grid side.

[0032] Distributed renewable energy power generation units are widely distributed, with multiple sources of operational data in varying formats, and are prone to outliers, missing values, and differences in units of measurement. Directly using raw data for feature extraction and model training would lead to chaotic feature representation, large model learning bias, and unstable identification results. On the other hand, setting different feature extraction algorithms at each edge device would be too labor-intensive, and unit of measurement differences would still exist during subsequent edge node processing. Therefore, the collected multi-source historical operational data is first preprocessed to form a standardized dataset, and then the corresponding edge devices extract features from the preprocessed data.

[0033] The preprocessing operations include data cleaning, data normalization, time-series alignment, and time-frequency conversion. First, outliers in the corresponding multi-source historical data are removed using the 3σ criterion. Short-term missing data with a duration of no more than 5 seconds is filled using linear interpolation, while long-term missing data with a duration of more than 5 seconds is filled using LSTM interpolation. A Min-Max normalization algorithm is then used to map all data to the [0, 1] interval. Next, the corresponding multi-source historical data is aligned to a unified preset time-series axis based on timestamps. Then, a short-time Fourier transform is used, combined with preset window functions, window lengths, and overlap rates, to convert the aligned time-series data into time-frequency data, completing the preprocessing of the multi-source historical data.

[0034] The standardized dataset contains time-frequency data obtained from each edge device after corresponding preprocessing. However, when performing feature extraction, each edge device only performs corresponding feature processing on the time-frequency data of the corresponding distributed new energy power generation unit.

[0035] Based on the standardized dataset, in order to match the hierarchical distribution and coupled operation characteristics of distributed new energy power generation units, and relying on the established cloud-edge-device three-level architecture, progressive, hierarchical, and multi-scale feature mining is realized to fully capture the inherent laws of new energy regulation capabilities under the coupling effect of multiple factors, and avoid the problems of feature redundancy and insufficient effective information caused by single-dimensional and parallel extraction.

[0036] The process, based on a standard dataset, involves edge devices extracting primary basic features, edge nodes extracting secondary intermediate features based on the corresponding primary basic features, and the cloud platform extracting tertiary deep features based on the corresponding secondary intermediate features. This includes: Based on the standard dataset, edge devices extract primary basic features, including time-series statistical features, environmental power mapping features, and local response features, using corresponding feature extraction algorithms. Edge nodes receive and integrate corresponding first-level basic features based on the corresponding hierarchical topology association, and extract second-level intermediate features including power synchronization features, load output matching features, and regional operating condition consistency features. Based on all the extracted secondary intermediate features, combined with power grid operation data, the cloud platform extracts tertiary deep features, including spatiotemporal distribution features, power grid constraint and adjustability correlation features, and long-term trend features.

[0037] Based on a standardized dataset, each edge device extracts corresponding primary basic features from the standardized dataset using lightweight feature extraction algorithms, such as improved one-dimensional convolutional neural networks. These primary basic features include temporal volumetric features, environmental power mapping features, and local response features. The temporal statistical features include power mean, variance, peak value, valley value, peak-valley difference, instantaneous rate of change, and cumulative change of environmental parameters. The environmental power mapping features include the slope of the wind speed and active power curves, the fitting coefficient between light intensity and active power, and the attenuation coefficient of power output due to temperature. The local response features include the delay time from control command issuance to power response, the maximum rate of change of power per unit time, and the overshoot and steady-state error during the response process.

[0038] After each edge device completes the extraction of primary basic features, it uploads these features to the corresponding edge node under its jurisdiction. Based on hierarchical topology relationships, each edge node receives the primary basic features uploaded by all edge devices within its jurisdiction and extracts secondary intermediate features reflecting the collaborative operation characteristics of the regional cluster through a feature fusion network with an attention mechanism.

[0039] The secondary intermediate characteristics include power synchronicity characteristics, load-output matching characteristics, and regional operating condition consistency characteristics. The power synchronicity characteristics include the Pearson correlation coefficient matrix of the power change rate of all power generation units within the region, the time-series phase difference, and the proportion of synchronicity fluctuations in the basic characteristics. The load-output matching characteristics include the deviation sequence between the total regional load and the total output of new energy sources, the standard deviation of the deviation sequence per unit time, the maximum peak value of the deviation sequence, and the load-output complementarity coefficient. The regional operating condition consistency characteristics include the difference in environmental parameters among units within the region, the output fluctuation coordination coefficient, and the overall regional operating condition stability index.

[0040] After each edge node completes the extraction of secondary intermediate features, they are uploaded to the cloud platform. The cloud platform receives and integrates all the secondary intermediate features uploaded by the edge nodes, and combines them with the power grid operation data received by the cloud platform. A combined architecture of Transformer network and deep residual network is used to extract tertiary deep features. These tertiary deep features include spatiotemporal distribution features, power grid constraint adjustability correlation features, and long-term trend features. The spatiotemporal distribution features include the temporal coupling coefficient of power output in different regions, regional processing spatial clustering results, and spatiotemporal thermodynamic features. The power grid constraint and adjustability correlation features include the matching degree between tie-line transmission capacity and renewable energy regulation capacity, the correlation coefficient between voltage stability and reactive power regulation, frequency deviation, and active power regulation response sensitivity. The long-term trend features include intraday power output fluctuation patterns, intraweek power output fluctuation patterns, intra-quarter power output fluctuation patterns, adjustability adaptation coefficients under different seasonal operating conditions, and the drift amount and drift rate of adjustability parameters over a long period.

[0041] The extracted three-level deep features can reflect the multi-dimensional relationship between the operation status of distributed new energy power generation units and the power grid in the target area, but they cannot directly output the quantitative adjustable characteristic parameters required for power grid dispatch, such as ramp rate, maximum regulation capacity, and response time. Therefore, based on the three-level deep features, the regulation capacity identification model is further trained by combining the measured preset regulation capacity labels, so that the model can learn and explore the complex internal coupling law between the three-level deep features and the actual adjustable characteristics, and realize the transformation from abstract features to specific quantitative identification results.

[0042] The step of training a regulatory ability identification model based on three levels of deep features and corresponding preset regulatory ability labels includes: A training dataset is constructed based on three levels of deep features and corresponding preset adjustment capability labels. The training dataset is input into the regulation capability identification model, which includes a feature encoding layer, a parameter prediction layer, and an error correction layer. The feature encoding layer performs associative encoding on the three-level deep features in the input training dataset to obtain the corresponding encoded feature vector; The parameter prediction layer maps the encoded feature vectors to predicted values ​​of modulating capacity; The error correction layer calculates the prediction error of the predicted value of the regulation capability based on the corresponding preset regulation capability label, and optimizes the parameters of the parameter prediction layer and the feature encoding layer based on the prediction error through the particle swarm optimization algorithm until the iteration termination condition is met.

[0043] The three-level deep features are used as input samples for model training. Then, the adjustable characteristic parameters such as the upper and lower limits of the ramp rate, the maximum active power regulation capacity, the maximum reactive power regulation capacity, and the regulation response time measured by the power grid are used as corresponding preset regulation capability labels to achieve a one-to-one correspondence between high-dimensional features and quantitative labels. At the same time, the labeled dataset is divided into training set, validation set and test set according to preset ratio to form training dataset.

[0044] Since the adjustable characteristics of distributed new energy are determined by multiple scale factors including equipment, region, and the entire network, and there is a high degree of nonlinear correlation between them and the three-level deep features, the transformation from features to parameters cannot be achieved through manual calculation or simple linear algorithms. Therefore, this embodiment specifically uses a hybrid identification model composed of a feature encoding layer, a parameter prediction layer, and an error correction layer as the adjustment identification model.

[0045] The feature encoding layer employs a Transformer encoder to correlate and encode the input three-level deep features, identifying hidden coupling patterns between features of different dimensions. This transforms the three-level deep features into encoded feature vectors that retain only the correlation information, reducing the computational complexity of the subsequent parameter prediction layer and allowing the model to focus on the discriminative value of the features. The parameter prediction layer takes the encoded feature vectors as input and constructs a non-linear mapping relationship between features and adjustable characteristics through a four-layer fully connected network. The hidden layers use the GELU activation function to ensure fitting ability to complex correlations, and the output layer uses the Sigmoid activation function to map the prediction results to a predetermined physical interval. Finally, it outputs a predicted value of the adjustable capability that is completely consistent with the preset adjustable capability label dimension, achieving the initial transformation from abstract high-dimensional features to quantitative adjustable characteristic parameters that can be directly used in power grid dispatching.

[0046] The error correction layer compares the predicted regulation capacity output by the parameter prediction layer with the corresponding preset regulation capacity label of the actual power grid measurement. It calculates the quantitative prediction error between the two through evaluation indicators such as mean square error, and uses this error as the objective function. With minimizing the objective function as the direction, the network parameters of the feature encoding layer and parameter prediction layer are globally iteratively optimized using the particle swarm optimization algorithm. During the optimization process, forward inference and parameter adjustment of the model are continuously carried out until the preset iteration termination conditions are met, namely, the prediction error of the validation set changes less than or equal to the set threshold for multiple consecutive iterations, the maximum number of iterations is reached, or the identification accuracy of the test set meets the power grid dispatching requirements. At this time, the optimization stops and the optimal parameters of the current model are saved, completing the offline training of the regulation capacity identification model.

[0047] Based on the three-level progressive feature extraction method used during model training, corresponding real-time three-level deep features are extracted from the real-time operation data of each distributed new energy power generation unit, and then input into the regulation capability identification model after training to output the regulation capability identification result.

[0048] The adjustment capability identification results include the single-device adjustable characteristic parameters of the edge device layer, the overall adjustable characteristic parameters of the cluster at the edge node layer, and the overall adjustable characteristic parameters of the regional hybrid cluster at the cloud platform layer.

[0049] However, considering that the three-level progressive feature extraction process used in model training mainly relies on bottom-up unidirectional transmission, with feature extraction and uploading between the edge, edge nodes, and cloud at fixed levels, lacking cross-level feature feedback and complementarity, it may not fully explore some weakly correlated and implicitly coupled operational patterns. Furthermore, with relatively fixed feature types and extraction methods, some features may become redundant or insufficiently representative in scenarios such as drastic fluctuations in new energy output, sudden changes in grid operating conditions, and abnormal equipment operation. Therefore, in addition to setting up a progressive three-level feature extraction system, a bidirectional feature interaction channel is further added between edge devices, edge nodes, and the cloud platform, relying on the standardized communication protocol of the original cloud-edge architecture.

[0050] The bidirectional feature interaction channel includes a downlink feedback channel and an uplink supplementary channel, both established between the cloud platform and edge nodes, and between edge nodes and edge devices. The downlink feedback channel employs a delayed, iterative distribution logic. During the initial feature extraction phase, the cloud does not distribute secondary intermediate feature fragments. Only after the edge nodes complete the first round of secondary intermediate feature extraction and upload it to the cloud platform, the cloud platform further lightweights the extracted tertiary deep features, such as grid constraints and adjustability-related features and temporal distribution features, and splits them according to the jurisdiction of each edge node. These splits are then distributed to the corresponding edge nodes via the downlink feedback channel at predetermined time intervals. Upon receiving the feedback, the edge nodes further decompose the extracted secondary intermediate features to obtain global constraint features and regional coordination features related to the adjustment capabilities of individual devices within their jurisdiction. These features are then distributed to the corresponding edge devices via the corresponding downlink feedback channel.

[0051] The uplink supplementary channel is used by edge devices to mark unexplained anomalous features, such as environmental and power mapping anomalies and local response delay anomalies, after extracting the primary basic features. These are then uploaded to the edge nodes via the uplink supplementary channel. The edge nodes then integrate the anomalous features from their subordinate edge devices, combine them with their own secondary intermediate features, and form a regional-level anomalous feature package. This package is then uploaded to the cloud platform via the corresponding uplink supplementary channel, where it is integrated into the tertiary deep features to improve the feature information.

[0052] When outputting regulation capability identification results based on real-time operating data of each distributed new energy power generation unit and the trained regulation capability identification model, feature extraction is carried out continuously according to the extraction cycle. In the first round of feature extraction, the edge device only extracts the first-level basic features of the current cycle without fusion and directly uploads them to the edge node to complete the feature extraction initialization. After the first round of feature extraction is completed and the edge node sends out relevant feature fragments, the edge device receives the regional collaborative feature fragments of the previous extraction cycle, such as regional operating condition consistency features and power synchronization feature fragments, sent by the edge node in each real-time extraction cycle. These fragments are then fused with the first-level basic features extracted in the current round. A weighted summation fusion algorithm is used to output the fused edge-optimized features, which are then uploaded to the edge node for real-time supplementation of the second-level intermediate features of the edge node in the current round. Similarly, the edge node only receives the global constraint features sent by the cloud and then combines them with its own extracted second-level intermediate features and the fused optimized features uploaded by the edge devices under its jurisdiction. A feature splicing and attention fusion algorithm is used to output the fused edge-optimized features, which are then uploaded to the cloud platform. Finally, the cloud platform receives the fused edge optimization features uploaded by all edge nodes, combines them with its own extracted level 3 deep features and the regional anomaly feature packages uploaded by each edge node, and outputs the final level 3 deep features used as input to the training and adjustment capability identification model.

[0053] Meanwhile, considering that the three-level deep features used in offline training are general features extracted based on fixed rules and for the average scenario of all working conditions, they cannot specifically highlight the key influencing factors at a certain moment. In actual real-time operation, the intensity of environmental fluctuations, the tightness of power grid constraints, and the operating status of equipment are constantly changing, and the features that play a dominant role in the adjustment capability will also change accordingly. If fixed feature extraction weights are used throughout the process, it will lead to an excessively high proportion of irrelevant features and a weakening of key features, ultimately making it impossible for the three-level deep features to accurately reflect the current real adjustable potential.

[0054] Therefore, in order to enable the three-level deep features input to the model to dynamically adapt to the current real-time operating conditions, thereby improving the accuracy and robustness of online identification, an adaptive weighting mechanism is further used to dynamically weight the first-level basic features and the second-level intermediate features at the edge and peripheral levels. This strengthens the more critical and representative feature information under the current operating conditions, while weakening redundant and secondary feature information. Then, through a hierarchical extraction process, this weighting is transmitted and solidified into the final three-level deep features, so that the three-level deep features fed into the model always focus on the most decisive operating rules and constraints at present.

[0055] Specifically, pre-defined feature relevance, operating condition adaptability, and feature redundancy indicators serve as dynamic filtering indicators. In each real-time feature extraction cycle, edge devices extract primary basic features based on pre-processed real-time data. Immediately after extraction, dynamic feature filtering is performed, eliminating redundant or invalid features under the current operating condition according to preset dynamic filtering indicators. Simultaneously, the weights of the primary basic features are dynamically adjusted using a random forest algorithm combined with the current and previous cycle's real-time operating conditions. Edge nodes receive the optimized primary basic features from their subordinate edge devices, extract secondary intermediate features, and simultaneously perform dynamic filtering and weight adjustment. They also receive the previous cycle's tertiary deep feature fragments from the cloud, fusing and optimizing them with the currently extracted secondary intermediate features to ensure that the information composition of the secondary features is tilted towards the current operating condition. The cloud platform receives the optimized secondary intermediate features from all edge nodes and extracts tertiary deep features by combining them with real-time power grid operation data. At this point, the weight adjustments of the primary and secondary intermediate features have been propagated hierarchically, altering the internal information distribution of the tertiary deep features. For example, when power grid interconnection lines are under strain, the weight of the load output matching feature in the secondary features is increased, resulting in a corresponding increase in the proportion of information related to grid constraints and adjustability in the tertiary deep features. The cloud platform then performs final dynamic filtering and weight calibration on the extracted tertiary deep features, eliminating globally redundant features and strengthening core related features, outputting optimized tertiary deep features adapted to the current real-time operating conditions.

[0056] The optimized three-level deep features are then input into the regulation capability identification model that has been deployed to the cloud, enabling the output of real-time regulation capability identification results.

[0057] The offline-trained model is built solely based on historical data and can only adapt to the operating scenario during training. However, in actual operation, gradual or abrupt changes such as the addition or removal of distributed renewable energy generation units, adjustments to the grid topology, changes in environmental patterns, and equipment performance degradation will gradually cause the mapping relationship between the original features and the regulation capability to become mismatched. Simply relying on adaptive adjustment of pre-existing features cannot fundamentally correct the mapping bias of the model itself, leading to a continuous decline in identification accuracy in the long run. Therefore, after outputting the regulation capability identification result, the following is also performed: The cloud platform layer uses a dual-trigger mechanism of timed triggering and event triggering to dynamically iterate and update the regulation capability identification model, and then sends the model parameters of the dynamically iterated and updated regulation capability identification model to the edge node layer and the edge device layer.

[0058] The dual-trigger mechanism, which combines timed triggering and event triggering, can achieve periodic lightweight iteration by slowly drifting the operating conditions through timed triggering to maintain the basic adaptability of the model. At the same time, it can immediately update the model mapping relationship through event triggering in sudden scenarios such as the access and exit of power generation units, drastic changes in grid constraints, and identification errors exceeding the threshold, so as to ensure the identification reliability under extreme operating conditions.

[0059] The cloud platform layer, based on a dual-trigger mechanism of timed triggering and event triggering, dynamically iterates and updates the adjustment capability identification model, including: Pre-set timed triggering rules and event triggering rules, and the cloud platform layer collects and outputs monitoring data after the regulation capability identification results in real time. The monitoring data includes identification error, power grid operating conditions and power generation unit access status. When the time interval after outputting the regulation capability identification result meets the timed triggering rule, or when the monitoring data meets the corresponding event triggering rule, the newly generated third-level deep features are extracted, and the corresponding regulation capability labels are added in combination with the monitoring data. An incremental training dataset is constructed in combination with the training dataset used during model training. The regulatory capacity identification model is dynamically iteratively updated based on the incremental training dataset.

[0060] The timed triggering rules are used to deal with slow drift-type operating condition changes such as environmental characteristics and equipment performance, while the event triggering rules are for sudden or drastic change scenarios such as identification error exceeding the standard, grid topology adjustment, and the addition or removal of power generation units. At the same time, after each output of regulation capability identification results, the cloud platform layer will continuously collect multi-dimensional monitoring data, including identification error, real-time grid operating conditions, and access status of distributed new energy power generation units, and use it as the basis for determining whether to start model iteration.

[0061] When the time interval after the output identification result reaches the period set by the timed triggering rule, or when the detected identification error exceeds the preset threshold, the power grid operating conditions change, or the power generation unit access status is adjusted, etc., which meet the event triggering rule, the cloud platform layer will immediately extract the newly generated third-level deep features in the corresponding time period, add matching regulation capability labels based on the parameters reflecting the real regulation capability in the monitoring data, and then combine the newly added feature and label data with the original training dataset used in the initial training stage of the model to construct a lightweight incremental training dataset, without having to call up the full historical data again, thus reducing the amount of iterative computation.

[0062] Based on this incremental training dataset, the regulation capability identification model is dynamically iteratively updated using an incremental learning approach. While retaining the mature feature mapping rules learned at the bottom layer of the model, only the top-level parameters of the model are optimized and adjusted, enabling the model to quickly adapt to new operating conditions and data distributions. After iteration, an optimized model adapted to the current power grid operating state is obtained, which not only ensures the real-time and efficient updating of the model, but also continuously conforms to the dynamic changes in distributed new energy generation and power grid operation, achieving long-term stable and high-precision regulation capability identification.

[0063] After model iteration is completed in the cloud, the updated model parameters are synchronously distributed to the edge node layer and edge device layer. This ensures that the real-time feature extraction, dynamic filtering, and adaptive weight allocation logic of the edge and edge devices are consistent with the updated model in terms of encoding rules, feature preferences, and input requirements. This ensures that the three-level deep features finally fed into the model match the learning patterns of the iterated model, avoiding the mismatch between features and the model caused by the updated model but the previous feature extraction rules not being synchronized. This ensures that high-precision and high-reliability adjustment capability identification results can be stably output throughout the entire lifecycle.

[0064] Another aspect of this embodiment also provides a distributed new energy power generation regulation capability identification system based on cloud-edge-device architecture, such as... Figure 2 As shown, it includes: The edge device layer includes several edge devices set at each distributed new energy power generation unit, which are used to collect multi-source historical operation data and real-time operation data of the corresponding distributed new energy power generation unit and extract primary basic features; The edge node layer includes several edge nodes set up at the new energy aggregation point, which are used to receive the first-level basic features of each distributed new energy power generation unit under the corresponding new energy aggregation point, so as to extract the corresponding second-level intermediate features. The cloud platform layer includes a cloud platform, which receives secondary intermediate features from each edge node to extract corresponding tertiary deep features. Based on the tertiary deep features, it trains a regulation capability identification model in combination with corresponding preset regulation capability labels. Alternatively, it outputs regulation capability identification results based on the real-time operation data of each distributed new energy power generation unit and the trained regulation capability identification model.

[0065] The cloud platform layer also includes: The update and iteration module is used to dynamically update the regulation capability identification model based on a dual triggering mechanism of timed triggering and event triggering, and to send the model parameters of the dynamically updated regulation capability identification model to the edge node layer and the edge device layer.

[0066] The edge devices are data processors with data processing capabilities, which can be deployed in the inverters of photovoltaic power generation units or in the main control system of wind turbines to realize local data processing of distributed new energy power generation units. The edge nodes are data processing servers with data processing capabilities, located at the new energy aggregation point, and can collect and process data uploaded by each distributed new energy power generation unit under their jurisdiction. The cloud platform consists of servers, computers, etc., with data processing units, located at the power grid dispatch center cloud platform, and can receive and process data uploaded by each edge node under its jurisdiction.

[0067] Furthermore, the edge devices and edge nodes are equipped with local processing modules that store model parameters of the regulation capability identification model issued by the cloud platform. In the event of a communication failure, the regulation capability can also be identified locally to ensure the normal operation of the power grid dispatch.

[0068] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Other variations and modifications are possible without departing from the technical solutions described in the claims.

Claims

1. A method for identifying the regulation capability of distributed new energy power generation based on cloud-edge-device architecture, characterized in that, include: Based on all distributed new energy power generation units within the target area, a cloud-edge architecture consisting of edge devices, edge nodes, and a cloud platform is formed according to the corresponding hierarchical topology. Collect multi-source historical operation data from each distributed new energy power generation unit, and construct corresponding standardized datasets by combining preprocessing. Based on the standard dataset, the edge devices extract the first-level basic features, the edge nodes extract the second-level intermediate features based on the corresponding first-level basic features, and the cloud platform extracts the third-level deep features based on the corresponding second-level intermediate features. Based on three levels of deep features, a regulatory ability identification model is trained by combining corresponding preset regulatory ability labels. Based on the real-time operation data of each distributed new energy power generation unit, and combined with the trained regulation capability identification model, the regulation capability identification results are output.

2. The method for identifying the regulation capability of distributed new energy power generation based on cloud-edge-device as described in claim 1, characterized in that, Based on a standard dataset, the edge devices extract primary basic features, the edge nodes extract secondary intermediate features based on the corresponding primary basic features, and the cloud platform extracts tertiary deep features based on the corresponding secondary intermediate features, including: Based on the standard dataset, edge devices extract primary basic features, including time-series statistical features, environmental power mapping features, and local response features, using corresponding feature extraction algorithms. Edge nodes receive and integrate corresponding first-level basic features based on the corresponding hierarchical topology association, and extract second-level intermediate features including power synchronization features, load output matching features, and regional operating condition consistency features. Based on all the extracted secondary intermediate features, combined with power grid operation data, the cloud platform extracts tertiary deep features including spatiotemporal distribution features, power grid constraint adjustability correlation features, and long-cycle trend features.

3. The method for identifying the regulation capability of distributed new energy power generation based on cloud-edge-device as described in claim 1, characterized in that, The training of the regulatory ability identification model based on three levels of deep features and corresponding preset regulatory ability labels includes: A training dataset is constructed based on three levels of deep features and corresponding preset adjustment capability labels. The training dataset is input into the regulation capability identification model, which includes a feature encoding layer, a parameter prediction layer, and an error correction layer. The feature encoding layer performs associative encoding on the three-level deep features in the input training dataset to obtain the corresponding encoded feature vector; The parameter prediction layer maps the encoded feature vectors to predicted values ​​of modulating capacity; The error correction layer calculates the prediction error of the predicted value of the regulation capability based on the corresponding preset regulation capability label, and optimizes the parameters of the parameter prediction layer and the feature encoding layer based on the prediction error through the particle swarm optimization algorithm until the iteration termination condition is met.

4. The method for identifying the regulation capability of distributed new energy power generation based on cloud-edge-device as described in claim 1, characterized in that, The adjustment capability identification results include the single-device adjustable characteristic parameters of the edge device layer, the overall adjustable characteristic parameters of the cluster at the edge node layer, and the overall adjustable characteristic parameters of the regional hybrid cluster at the cloud platform layer.

5. The method for identifying the regulation capability of distributed new energy power generation based on cloud-edge-device as described in claim 1, characterized in that, After outputting the adjustment capability identification results, the following is also executed: The cloud platform layer uses a dual-trigger mechanism of timed triggering and event triggering to dynamically iterate and update the regulation capability identification model, and then sends the model parameters of the dynamically iterated and updated regulation capability identification model to the edge node layer and the edge device layer.

6. The method for identifying the regulation capability of distributed new energy power generation based on cloud-edge-device as described in claim 5, characterized in that, The cloud platform layer uses a dual-trigger mechanism of timed triggering and event triggering to dynamically iterate and update the adjustment capability identification model, including: Pre-set timed triggering rules and event triggering rules, and the cloud platform layer collects and outputs monitoring data after the regulation capability identification results in real time. The monitoring data includes identification error, power grid operating conditions and power generation unit access status. When the time interval after outputting the regulation capability identification result meets the timed triggering rule, or when the monitoring data meets the corresponding event triggering rule, the newly generated third-level deep features are extracted, and the corresponding regulation capability labels are added in combination with the monitoring data. An incremental training dataset is constructed in combination with the training dataset used during model training. The regulatory capacity identification model is dynamically iteratively updated based on the incremental training dataset.

7. The method for identifying the regulation capability of distributed new energy power generation based on cloud-edge-device as described in claim 1, characterized in that, The aforementioned cloud-edge architecture, based on all distributed new energy power generation units within the target area and arranged according to corresponding hierarchical topology, consists of edge devices, edge nodes, and a cloud platform, including: Determine the distribution information of each distributed new energy power generation unit within the target area, and set up corresponding edge devices at each distributed new energy power generation unit to form an edge device layer; Based on the distribution information of each distributed new energy power generation unit, identify the new energy aggregation points in the target area, and set corresponding edge nodes at each new energy aggregation point to form an edge node layer; A cloud platform is set up at the power grid dispatch center corresponding to the target area to form a cloud platform layer; The cloud-edge architecture is formed by using the edge device layer as the bottom layer of the topology, the edge node layer as the middle layer of the topology, and the cloud platform layer as the top layer of the topology.

8. The method for identifying the regulation capability of distributed new energy power generation based on cloud-edge-device as described in claim 1, characterized in that, The multi-source historical operating data includes the active power, reactive power, bus voltage, output current, environmental parameters, and control command signals of each distributed new energy power generation unit, as well as the corresponding regional load and tie-line power data on the grid side.

9. A distributed new energy power generation regulation capability identification system based on cloud-edge-device architecture, used to execute the identification method according to any one of claims 1 to 8, characterized in that, include: The edge device layer includes several edge devices set at each distributed new energy power generation unit, which are used to collect multi-source historical operation data and real-time operation data of the corresponding distributed new energy power generation unit and extract primary basic features; The edge node layer includes several edge nodes set up at the new energy aggregation point, which are used to receive the first-level basic features of each distributed new energy power generation unit under the corresponding new energy aggregation point, so as to extract the corresponding second-level intermediate features. The cloud platform layer includes a cloud platform, which receives secondary intermediate features from each edge node to extract corresponding tertiary deep features. Based on the tertiary deep features, it trains a regulation capability identification model in combination with corresponding preset regulation capability labels. Alternatively, it outputs regulation capability identification results based on the real-time operation data of each distributed new energy power generation unit and the trained regulation capability identification model.

10. The distributed new energy power generation regulation capability identification system based on cloud-edge-device according to claim 9, characterized in that, The cloud platform layer also includes: The update and iteration module is used to dynamically update the regulation capability identification model based on a dual triggering mechanism of timed triggering and event triggering, and to send the model parameters of the dynamically updated regulation capability identification model to the edge node layer and the edge device layer.

Images