A wind farm group distributed data processing and collaborative control method and system

Abstract

This application relates to the field of wind farm control, specifically to a distributed data processing and collaborative control method and system for wind farm clusters. The method includes: deploying primary and backup dual edge nodes for each wind turbine to synchronously collect data; the primary node extracts key features through adaptive parsing algorithms and dimensionality reduction processing, inputs them into a local anomaly detection model for real-time monitoring and early warning; simultaneously, based on network and load status, intelligently decides to upload key data to the cloud; the backup node synchronously runs the anomaly detection model and monitors the primary node's heartbeat; the cloud calculates the overall reactive power demand of each cluster based on the entire cluster's data and grid commands, and issues commands; each primary node uses a consensus algorithm to quickly distribute reactive power to each wind turbine and executes voltage collaborative control; when the primary node fails, the backup node takes over and verifies the command signature through a lightweight consensus algorithm to ensure that the control commands are authentic, reliable, and tamper-proof, comprehensively improving the operational safety of the wind farm cluster and the grid support capability.

Description

Technical Field

[0001] This application relates to the field of wind farm control technology, specifically to a method and system for distributed data processing and collaborative control of wind farm clusters. Background Technology

[0002] As the installed capacity of wind power continues to expand, wind farm clusters are gradually transforming from isolated power sources into a key support force for the power grid, and their data processing and control modes are facing severe challenges: First, current mainstream control architectures rely on centralized remote servers or single field controllers. In remote wind farms, network links are easily disrupted by weather and terrain. Once the central node or communication link fails, the entire wind farm cluster will lose its local real-time control capabilities, and may even trigger a large-scale grid disconnection incident.

[0003] Second, wind farm clusters typically consist of wind turbines from multiple manufacturers and of various models, with different sensor configurations and communication protocols, resulting in highly heterogeneous data formats. While traditional centralized architectures can perform global data analysis, data cleaning, format conversion, and long-distance transmission introduce significant latency, making it impossible to effectively implement real-time collaborative control strategies based on global status.

[0004] Third, a single wind turbine generates tens of thousands of high-dimensional data points per second. Transmitting all the raw data back to the cloud puts enormous pressure on limited network bandwidth, resulting in a waste of storage and computing resources. Existing edge computing solutions primarily focus on data filtering, lacking integrity protection and tamper-proof mechanisms for the transmission of critical control commands.

[0005] Therefore, how to construct a distributed processing architecture for wind farm clusters that combines high reliability, low latency collaborative control, and inherent security has become a technical challenge that the industry urgently needs to solve. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a method and system for distributed data processing and collaborative control of wind farm clusters.

[0007] In a first aspect, the present invention provides a distributed data processing and collaborative control method for wind farm clusters. This method deploys mutually supporting edge nodes for each wind turbine. The primary and backup edge nodes synchronously collect multi-source heterogeneous data, and each performs data parsing and dimensionality reduction processing to obtain key feature data. Each node also runs an anomaly detection model to monitor its status and generate early warning information. The mutually supporting edge nodes monitor each other through heartbeat detection. The method includes the following steps: When the primary edge node determines that the current processing task is a cloud-based collaborative computing task based on the current network latency, its own computing load, and energy consumption constraints through an adaptive decision model, it uploads the key feature data and early warning information to the cloud. Based on the aforementioned key feature data, combined with power grid dispatch instructions and the overall network operation status, the cloud calculates the overall reactive power demand of each wind farm cluster and sends the reactive power allocation instructions to the primary edge nodes of each wind farm and copies them to the backup nodes. Each primary edge node uses a consensus algorithm to distribute the total reactive power demand to each wind turbine in the field and performs voltage coordinated control; backup edge nodes synchronously participate in the consensus algorithm iteration. When the primary edge node fails, the backup edge node seamlessly takes over all functions, including data uploading, command reception and execution, and initiates a consensus algorithm to verify subsequent control commands through multi-party signatures. The cloud automatically sends subsequent commands to the new primary edge node.

[0008] The primary edge node employs an adaptive parsing and dimensionality reduction local real-time data processing mode, reducing data transmission redundancy and network bandwidth pressure. This enables real-time monitoring and anomaly warning of wind turbine operation status, allowing for rapid detection and timely response to turbine faults. The adaptive decision-making model dynamically determines task processing methods based on network, load, and energy consumption constraints, achieving collaborative division of labor between the edge and cloud. This balances the efficiency of local real-time processing with the rationality of global optimization in the cloud, avoiding resource waste. Backup edge nodes synchronously run an anomaly detection model and achieve mutual monitoring through heartbeat detection, ensuring rapid switching and state consistency between primary and backup nodes. The cloud combines grid dispatch and overall network status calculations to determine overall reactive power demand. Edge nodes use consensus algorithms to achieve reactive power distribution and voltage coordination control within the wind farm, enabling coordinated adaptation between the wind farm cluster and the grid, improving grid voltage stability and wind power absorption capacity.

[0009] As a further limitation of the technical solution of the present invention, the data parsing is performed in real time on multi-source heterogeneous data through a node-level adaptive parsing algorithm, specifically including: Based on the computing power, storage resources, and network bandwidth of the edge nodes, the edge nodes in the wind farm cluster are divided into multiple levels; different levels of parsing templates are configured for edge nodes of different complexity, with high-level nodes performing complete data parsing and low-level nodes performing key field parsing. When collecting heterogeneous data from multiple sources, the edge nodes automatically identify the device brand and sensor type of the data source, and call the corresponding parsing template according to the preset device-template mapping table to achieve real-time hierarchical parsing of heterogeneous data without the need for unified format conversion.

[0010] By classifying edge nodes according to differences in computing, storage, and bandwidth resources and configuring parsing templates of varying complexity, this system achieves precise matching of resources with parsing tasks. This avoids processing delays caused by high-complexity parsing tasks consuming resources from low-configuration nodes, and also prevents resource waste caused by low-configuration nodes performing unnecessary full parsing. This improves the efficiency and resource utilization of edge node data parsing. By automatically identifying device brands and sensor types and calling corresponding parsing templates, the system eliminates the need for unified format conversion of multi-source heterogeneous data, simplifying the parsing process, reducing the complexity and time consumption of data preprocessing, and enabling real-time hierarchical parsing of heterogeneous data. This ensures that data from different sources and in different formats can be quickly converted into usable data, providing timely and accurate data support for subsequent anomaly detection and collaborative control.

[0011] As a further limitation of the technical solution of the present invention, the dimensionality reduction process adopts a redundancy elimination and compression strategy that integrates incremental calculation and fixed-point calculation, specifically including: For the continuous sampling data stream of the wind turbine, the change between the current sampled value and the previous sampled value is calculated. If the change exceeds a preset dynamic threshold, it is determined to be valid change data and is retained; if it does not exceed the dynamic threshold, it is determined to be redundant data and is not recorded temporarily. For periodic stationary data that continuously meet the incremental threshold condition within a preset time window, it is identified as a fixed-point sequence; the first sample value, the last sample value, and the trend indicator of the sequence are recorded, and intermediate sample data are not retained; The first and second buffers are configured to work alternately. The first buffer is used for real-time writing of the current data cycle, while the second buffer is used for data processing of the previous data cycle. Zero-wait data processing is achieved through the alternation of the two buffers.

[0012] By analyzing the relationship between sampled value changes and dynamic thresholds, redundant data with no effective changes is effectively eliminated, reducing data storage and transmission volume, and alleviating storage pressure and network transmission burden on edge nodes. Periodically stationary data is identified as fixed-point sequences, and only key information is recorded, further compressing data volume while preserving the core trend of data change. This ensures that the dimensionality-reduced data meets the needs of subsequent anomaly detection and collaborative control while minimizing redundancy. The dual-buffer alternating operation mode enables parallel execution of current data writing and previous cycle data processing, avoiding data processing delays and achieving zero-wait data processing. This ensures real-time data processing and guarantees that edge nodes can quickly output key feature data.

[0013] As a further limitation of the technical solution of the present invention, the anomaly detection model is a lightweight long short-term memory network model. Key feature data is input into this model for real-time monitoring and generation of early warning information, specifically including: The key feature data after dimensionality reduction are used to form a time series input vector. The key feature data includes at least the wind turbine vibration feature value, generator speed, wind speed, active power and blade angle. The lightweight long short-term memory network anomaly detection model performs real-time inference on the time series input vector and outputs the anomaly probability value of the current running state and the predicted anomaly type. When the anomaly probability value exceeds the preset warning threshold, a warning message is generated that includes the anomaly type, the time of the anomaly, the anomaly probability value, and the number of the wind turbine in which it occurs. Edge nodes receive model update parameters from the cloud and incrementally update the anomaly detection model so that the model can adapt to the changes in operating conditions caused by the long-term operation of the wind turbine.

[0014] Employing a locally deployed lightweight long short-term memory network anomaly detection model eliminates the need for cloud computing resources, enabling real-time anomaly reasoning, improving anomaly detection response speed, and ensuring that wind turbine anomalies can be quickly detected and generate early warning information. The early warning information contains complete key information, facilitating staff to quickly locate abnormal wind turbines, anomaly types, and anomaly severity, providing precise guidance for fault diagnosis and handling, and shortening fault handling time.

[0015] As a further limitation of the technical solution of the present invention, when the current processing task is determined to be a cloud-based collaborative computing task by an adaptive decision model based on the current network latency, its own computing load, and energy consumption constraints, the key feature data and early warning information are uploaded to the cloud, specifically including: The primary edge node is used to perceive the current environmental status in real time, and the environmental status includes at least: network status, node status, and task characteristics. The environmental state is input into an adaptive decision model deployed on an edge node, and the unloading decision for the current task is output. The unloading decision includes two types: local execution and cloud collaborative computing. When the decision result is cloud-based collaborative computing, the corresponding key feature data and early warning information are encapsulated into an upload data package, timestamps and node identifiers are added, and then uploaded to the cloud.

[0016] The adaptive decision-making model dynamically determines task processing methods based on network latency, computational load, and energy consumption constraints. This avoids network congestion and response delays caused by blindly uploading tasks to the cloud, and also prevents local nodes from undertaking tasks beyond their capabilities, leading to decreased processing efficiency and excessive energy consumption. It achieves an optimal division of labor between local processing and cloud-based collaborative computing, balancing processing efficiency, resource consumption, and energy costs. When key feature data is uploaded, timestamps and node identifiers are added, facilitating accurate traceability of data sources and collection times in the cloud, ensuring the accuracy and traceability of the overall reactive power requirements calculated by the cloud based on the data.

[0017] As a further limitation of the technical solution of this invention, the cloud calculates the total reactive power demand of each wind farm cluster and issues instructions, specifically including: The cloud receives key feature data uploaded by the main edge nodes of each wind farm. The key feature data includes at least the real-time active power, reactive power, grid connection point voltage and current phase angle of each wind farm. The cloud-based system collects dispatch instructions issued by the power grid dispatch center, including target voltage curves, reactive power limits, and power factor requirements; it also collects real-time voltage, frequency, and power flow distribution data at the power grid's point of common coupling. Based on the aforementioned key feature data, power grid dispatch instructions, and overall network operation status, the cloud platform calculates the total reactive power required for each wind farm cluster, with the optimization objective of minimizing the voltage deviation at the grid connection point, combined with the reactive power regulation capability constraints of the wind farms. The cloud platform converts the calculated total reactive power demand into reactive power allocation instructions, which include at least the target reactive power value, adjustment rate requirement, and response time window. The reactive power allocation instructions are then sent to the primary edge nodes of each wind farm and copied to the backup edge nodes.

[0018] The cloud platform integrates key characteristic data from various wind farms, grid dispatch instructions, and overall grid operation status to ensure that the calculation of overall reactive power demand takes into account both the actual operation of the wind farm cluster and the grid dispatch requirements. This avoids a disconnect between the calculated reactive power demand and the actual grid demand, achieving coordinated adaptation between the wind farm cluster and the grid. With minimizing the grid connection point voltage deviation as the optimization objective, and considering the reactive power regulation capacity constraints of the wind farms, the calculated overall reactive power demand meets the grid voltage stability requirements and conforms to the actual operating capacity of the wind farms, avoiding control failures caused by exceeding the wind farm's regulation range. Reactive power allocation instructions include specific target values, regulation rates, and response time windows, enabling edge nodes to clearly define execution standards and time requirements, ensuring the executability of reactive power allocation instructions. Instructions are issued through a secure channel to prevent tampering or theft during transmission, ensuring the security of instruction transmission and guaranteeing that each wind farm can accurately receive and execute reactive power allocation instructions.

[0019] As a further limitation of the technical solution of the present invention, the primary edge nodes of each wind farm use a consensus algorithm to distribute the total reactive power demand to each wind turbine within the farm, specifically including: The primary edge node receives reactive power allocation instructions from the cloud and parses them to obtain the overall reactive power target value that the wind farm needs to provide. Adjustment rate requirements and response time window; The primary edge node calculates the initial reactive power allocation scheme based on the current operating status and reactive power regulation capacity of each wind turbine in the field, and determines the initial reactive power output target for each wind turbine. The primary edge node initiates the consensus algorithm, and each wind turbine node exchanges its reactive power output status through the adjacent communication network, iteratively updating the reactive power allocation value until the reactive power output ratio of all wind turbines reaches a consensus; the backup edge node synchronously participates in the iterative calculation of the consensus algorithm. The primary edge node will distribute the final agreed reactive power output target to every wind turbine in the field; Each wind turbine performs reactive power output adjustment and feeds back the actual results to the primary edge node; the primary edge node monitors whether the total reactive power output of the entire site has reached the target. If a deviation exists, a secondary adjustment will be triggered.

[0020] The application of consensus algorithms enables wind turbine nodes to exchange status and iteratively update reactive power allocation values ​​through adjacent communication, ensuring that the reactive power output ratio of all wind turbines is consistent. This achieves a balanced distribution of reactive power demand within the wind farm, avoiding local overload or resource idleness and improving the collaborative control effect within the wind farm. The primary edge nodes monitor the total reactive power output of the entire wind farm and trigger secondary adjustments to promptly correct reactive power output deviations. This ensures that the wind farm accurately meets the overall reactive power target required by the cloud, guarantees the collaborative stability between the wind farm and the power grid, and improves voltage control accuracy.

[0021] As a further limitation of the technical solution of the present invention, the consensus algorithm adopts the average consensus protocol, and the specific iterative formula is as follows:

[0022] in, For the first Typhoon machine in reactive power output value at the next iteration In order to be with the first A group of wind turbines adjacent to each other. This is the iteration step size, which takes a value between 0 and 1; The iteration termination condition is: ,in This is the preset convergence accuracy threshold.

[0023] This ensures the accuracy of the allocation results. The protocol is computationally simple, has low communication overhead, and fast convergence speed, enabling it to complete reactive power allocation across the entire field within milliseconds, meeting the real-time requirements of grid voltage support.

[0024] As a further limitation of the technical solution of this invention, when the primary edge node fails, the backup edge node takes over all functions and ensures instruction security through a lightweight consensus algorithm, specifically including: A millisecond-level heartbeat detection mechanism is established between the primary edge node and the backup edge node. The two parties periodically send heartbeat messages to each other. The heartbeat messages include node identifier, timestamp, and current working status. If the backup edge node does not receive a heartbeat response from the primary edge node within N consecutive heartbeat cycles, the primary edge node is deemed to have failed. The backup edge node automatically switches to the new primary edge node, taking over all the functions of the original primary edge node, including data uploading, command reception and execution; The newly appointed master edge node establishes a consensus group with the edge nodes of other wind turbines in the field, initializes the lightweight consensus protocol, and prepares to conduct multi-party verification of the control commands to be issued. After the new master uses the edge node to generate control commands, it uses its own private key to sign the command content and generates a digital signature to be appended to the command. After receiving the instruction, other edge nodes in the consensus group use the public key of the new master edge node to verify the validity of the signature, confirming that the source of the instruction is genuine and that the content has not been tampered with. Once the verification is successful, each wind turbine executes the control command; if the verification fails, execution is refused and an anomaly is reported to the cloud.

[0025] Backup edge nodes automatically switch over and take over the network address, connectivity, and control tasks of the original primary node, achieving seamless switching and avoiding control chaos and data loss during the switchover process. The application of lightweight consensus algorithms and digital signature verification mechanisms eliminates the need for complex computational processes, balancing security and efficiency. Multi-node verification and anomaly reporting mechanisms within the consensus group further enhance the reliability of verification, avoiding vulnerabilities in single-node verification and ensuring the secure and accurate execution of control commands.

[0026] Secondly, the present invention provides a distributed data processing and collaborative control system for wind farm clusters, comprising: Multiple dual-edge node units, each deployed on a wind turbine, include a primary edge node and a backup edge node that serve as each other's primary and backup edge nodes, used to collect multi-source heterogeneous data and process it locally; The cloud-based control platform communicates with the primary edge nodes of each wind farm to calculate the overall reactive power demand based on key feature data and grid dispatch instructions and to issue instructions. The primary edge node, backup edge node, and cloud control platform are configured to perform the method described in the first aspect.

[0027] As can be seen from the above technical solutions, this application has the following advantages: By deploying primary and backup dual edge nodes for each wind turbine, a highly reliable local autonomous architecture is constructed. The primary node is responsible for real-time data processing and control, while the backup node runs synchronously and monitors in real time, achieving seamless fault switching and completely eliminating the risk of single-point failure. Through adaptive parsing and dimensionality reduction processing, massive heterogeneous data is extracted into key feature data, preserving core information while significantly reducing the transmission burden. Based on the adaptive decision-making model of network status and node load, the task execution position is intelligently determined, achieving a dynamic balance between local real-time control and cloud-based collaborative optimization. The cloud calculates reactive power requirements based on the entire field data, and the edge nodes use a consensus algorithm to quickly distribute them to each wind turbine, solving the problems of large latency and slow response in centralized control. A lightweight consensus algorithm verifies the signatures of control commands, ensuring that the commands are authentic, trustworthy, and tamper-proof, thus guaranteeing the security of the control system from the source.

[0028] This method achieves highly reliable, low-latency, and inherently safe collaborative control of wind farm clusters, thereby improving grid support capabilities and power generation efficiency. Attached Figure Description

[0029] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention.

[0031] Figure 2 A schematic diagram of the process for implementing voltage coordinated control. Detailed Implementation

[0032] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0034] like Figure 1 As shown, this embodiment of the invention provides a distributed data processing and collaborative control method for wind farm clusters, including: S1. Deploy mutually supporting edge nodes for each wind turbine. The primary edge node and the backup edge node synchronously collect multi-source heterogeneous data, and each performs data parsing and dimensionality reduction processing to obtain key feature data. Each runs an anomaly detection model to monitor the status and generate early warning information. The mutually supporting edge nodes monitor each other through heartbeat detection. S2. When the primary edge node determines that the current processing task is a cloud-based collaborative computing task based on the current network latency, its own computing load, and energy consumption constraints through an adaptive decision model, it uploads the key feature data and early warning information to the cloud. S3. Based on the key feature data, combined with the power grid dispatch instructions and the overall network operation status, the cloud calculates the overall reactive power demand of each wind farm cluster, and sends the reactive power allocation instructions to the primary edge nodes of each wind farm and copies them to the backup nodes. S4. Each primary edge node uses a consensus algorithm to distribute the overall reactive power demand to each wind turbine in the field and performs voltage coordinated control; backup edge nodes synchronously participate in the consensus algorithm iteration. S5. When the primary edge node fails, the backup edge node seamlessly takes over all functions, including data uploading, command reception and execution, and initiates a consensus algorithm to verify subsequent control commands through multi-party signatures. The cloud automatically sends subsequent commands to the new primary edge node.

[0035] In this embodiment of the invention, in S1, the data parsing is performed in real time on multi-source heterogeneous data using an adaptive parsing algorithm based on node hierarchy, specifically including: S11. Based on the computing power, storage resources and network bandwidth of the edge nodes, divide the edge nodes in the wind farm group into multiple levels; configure parsing templates of different complexity for edge nodes of different levels, with high-level nodes performing complete data parsing and low-level nodes performing key field parsing. S12. When collecting heterogeneous data from multiple sources, the edge node automatically identifies the device brand and sensor type of the data source, and calls the corresponding parsing template according to the preset device-template mapping table to achieve real-time hierarchical parsing of heterogeneous data without the need for unified format conversion.

[0036] In this embodiment of the invention, the edge nodes of different levels include: The primary node is deployed on the main line or key wind turbine, and is configured with a complete field parsing template to perform full parsing of all collected data; Secondary nodes, deployed on ordinary wind turbines, are configured with key field parsing templates to parse only preset core operating parameters; The division of primary and secondary nodes is dynamically adjusted based on the topology of the wind farm cluster and the importance of the wind turbines.

[0037] In this embodiment of the invention, in S1, the dimensionality reduction process employs a redundancy elimination and compression strategy that integrates incremental computation and fixed-point computation, specifically including: S13. For the continuous sampling data stream of the wind turbine, calculate the change between the current sampled value and the previous sampled value. If the change exceeds a preset dynamic threshold, it is determined to be valid change data and retained; if it does not exceed the dynamic threshold, it is determined to be redundant data and not recorded. S14. For periodic stationary data that continuously meets the incremental threshold condition within a preset time window, identify it as a fixed-point sequence; record the first sample value, the last sample value, and the trend indicator of the sequence, and do not retain intermediate sample point data; S15. Set the first buffer and the second buffer to work alternately. The first buffer is used for real-time writing of the current data cycle, while the second buffer is used for data processing of the previous data cycle. Zero-wait data processing is achieved by rotating the two buffers.

[0038] It should be noted here that the dynamic threshold is adaptively adjusted according to the real-time operating conditions of the wind turbine: When the wind turbine is in steady-state operation, a first threshold is used to reduce the amount of data transmission. When the wind turbine is operating under varying conditions or when abnormal signs are detected, it automatically switches to the second threshold to collect more intensive data for fault diagnosis, where the first threshold is higher than the second threshold.

[0039] The trend indicators include three types: continuous rise, continuous fall, and periodic fluctuation, which are used to restore the data's change patterns even after data compression. In some embodiments, in S2, the anomaly detection model is a lightweight long short-term memory network model. Key feature data is input into this model for real-time monitoring and to generate early warning information. Specifically, this includes: S21. Construct a time series input vector from the key feature data after dimensionality reduction processing. The key feature data includes at least the wind turbine vibration feature value, generator speed, wind speed, active power and blade angle. S22. A lightweight long short-term memory network anomaly detection model deployed at edge nodes performs real-time inference on the time series input vector, outputting the anomaly probability value of the current running state and the predicted anomaly type; the anomaly type includes: The types of failures of key components of the wind turbine include at least gearbox failure, generator bearing failure, and blade imbalance failure. Abnormal operating conditions include at least sudden wind speed changes, sharp power drops, excessive speed operation, and vibration exceeding limits; Specifically, for anomalies such as sudden wind speed changes and fault warnings, the model can issue warning information at a predetermined time before the anomaly occurs or issue warning information in advance.

[0040] S23. When the anomaly probability value exceeds a preset warning threshold, a warning message is generated containing the anomaly type, anomaly occurrence time, anomaly probability value, and the wind turbine number. The warning message is stored locally and then reported to the cloud. Here, reporting the warning message to the cloud specifically includes: The primary edge node adopts a tiered reporting strategy based on the current network status and alert level: For Level 1 warnings (emergency faults), an immediate upload and request for cloud intervention will be triggered. For Level 2 warnings (potential fault trends), they are uploaded along with periodic data packets; For Level 3 early warnings (fluctuations in normal operating status), data is stored locally only and not uploaded to the cloud.

[0041] S24. The edge node receives the model update parameters sent from the cloud and performs incremental updates on the local anomaly detection model so that the model can adapt to the changes in operating conditions caused by the long-term operation of the wind turbine.

[0042] In S22, the lightweight long short-term memory network anomaly detection model is trained through the following steps: S22a. Historical operating data and fault samples from multiple wind farms are collected in the cloud to construct a training dataset. A deep LSTM network is used as the teacher model and pre-trained in the cloud to obtain a high-precision fault diagnosis teacher model. The training dataset includes: SCADA operating data under normal operating conditions, used to train the model's ability to represent normal states; fault sample data, including at least five typical anomalies: gearbox faults, generator bearing faults, blade imbalance faults, sudden wind speed changes, and sharp power drops. The fault sample data is expanded through data augmentation techniques, including adding noise, time series pruning, and amplitude scaling, to improve the model's generalization ability.

[0043] S22b. Using knowledge distillation technology, the knowledge of the teacher model is transferred to a lightweight student model. The student model is a shallow LSTM network or a bidirectional gated recurrent unit network. The number of model parameters is compressed to less than 10% of that of the teacher model, and the inference latency is less than 50 milliseconds. S22c and the cloud will use the distilled lightweight anomaly detection model as the initial model and distribute it to the primary edge nodes and backup edge nodes of each wind farm. S22d: The edge node receives model update parameters from the cloud and uses local real-time operating data to incrementally learn the anomaly detection model, enabling the model to adapt to the personalized operating conditions of a single wind turbine.

[0044] In S22b, knowledge distillation is used to transfer the knowledge from the teacher model to the lightweight student model, specifically including: The training samples are input into the teacher model to obtain the soft labels output by the teacher model. The soft labels contain probability distribution information of each category and reflect the confidence level of the teacher model in judging the samples. The same training sample is input into the student model, and the KL divergence between the student model output and the soft label is calculated as the distillation loss, so that the student model can learn the generalization ability of the teacher model. Simultaneously, the cross-entropy loss between the student model output and the real fault label is calculated as the supervision loss; We combine the distillation loss and the supervision loss using a weighted summation method, and optimize the student model parameters through backpropagation. The weight of the distillation loss is higher than that of the supervision loss, so that the student model can inherit the diagnostic accuracy of the teacher model while maintaining its lightweight nature.

[0045] Furthermore, the edge node receives model update parameters from the cloud and uses local real-time running data to incrementally learn the anomaly detection model, specifically including: The edge node caches the normal operation data of this wind turbine for the past 30 days and abnormal signs data before the failure occurred. Incremental updates are triggered when the number of locally cached fault samples reaches a preset number, or when the model diagnostic accuracy continuously declines beyond a preset threshold. Edge nodes use cached local data to fine-tune the anomaly detection model in a few rounds, update the model parameters, and adapt the model to the personalized vibration characteristics and operating conditions of the wind turbine. The edge nodes report the updated model parameters and version information to the cloud. The cloud records the model version of each wind turbine and synchronizes it to the backup edge nodes when needed.

[0046] In some embodiments, when S2 determines that the current processing task is a cloud-based collaborative computing task based on the current network latency, its own computing load, and energy consumption constraints using an adaptive decision model, the key feature data and early warning information are uploaded to the cloud. Specifically, this includes: S25. The primary edge node perceives the current environmental status in real time, and the environmental status includes at least: network status, node status, and task characteristics. S26. Input the environmental state into the adaptive decision model deployed on the edge node, and output the unloading decision of the current task. The unloading decision includes two types: local execution and cloud collaborative computing. S27. When the decision result is cloud-based collaborative computing, the corresponding key feature data is encapsulated into an upload data package, a timestamp and node identifier are added, and then uploaded to the cloud.

[0047] It should be further noted that the adaptive decision-making model described in S26 is trained through the following steps: S26a. Collect historical operating data from multiple wind farms in the cloud, including network latency records, edge node computing load logs, energy consumption data, and task processing records; clean and normalize the data to construct a training sample set, where each training sample contains state features and a corresponding optimal offloading decision label. S26b. In a cloud-based offline environment, a deep reinforcement learning algorithm is used to pre-train the adaptive decision-making model, enabling the model to learn the mapping relationship from the environment state to the unloading decision; the deep reinforcement learning algorithm uses a deep Q-network or a policy gradient algorithm; in this step, the cloud-based offline pre-training step specifically includes: Environment modeling: A simulator for the edge computing environment is built in the cloud. The simulator simulates the dynamic changes in network latency, the fluctuation of computing load on edge nodes, energy consumption constraints, and the randomness of task arrival. A reward function is defined to evaluate the merits of each offloading decision. Agent initialization: Initialize the deep Q network as the decision agent. The network structure includes an input layer, two hidden layers, and an output layer. The dimension of the input layer corresponds to the dimension of the state features, and the dimension of the output layer corresponds to the number of unloading actions. Experience replay: The agent interacts with the simulation environment and stores the experience (state, action, reward, next state) of each interaction into the experience replay pool; Parameter update: Randomly sample a small batch of experience from the experience replay pool, calculate the mean square error between the target Q value and the current Q value as the loss function, and update the network parameters through gradient backpropagation; Iterative training: Repeat the interaction-replay-update process until the model converges or reaches the preset number of training rounds.

[0048] The reward function is designed as follows:

[0049] in, and These are the estimated completion times for tasks executed locally and in the cloud, respectively. and These are the estimated energy consumptions for local execution and cloud execution, respectively. This is an indicator function for successful task completion (1 for success, 0 for failure). Penalties for violations of task timeliness; For the weighting coefficients, satisfying ,and and The value is dynamically adjusted based on the importance of the task; for critical control tasks, A larger value is used to encourage successful task completion.

[0050] S26c. Use the validation set to evaluate the pre-trained model and calculate the model's decision accuracy and task completion efficiency in various scenarios. When the evaluation metrics reach the preset threshold, freeze the model and prepare it for distribution. S26d and the cloud will process the pre-trained adaptive decision model in a lightweight manner and distribute it to the primary edge nodes and backup edge nodes of each wind farm. In the actual operating environment, S26e and edge nodes perform online fine-tuning of the model based on real-time feedback data to adapt the model to the network characteristics and computational load patterns of the wind farm.

[0051] The optimal unloading decision label in the training sample set is generated using one of the following methods: Method 1: Expert experience annotation - Wind power experts manually annotate task offloading decisions in typical scenarios based on historical operation data to form the labeled data required for supervised learning; Method 2: Offline optimization calculation - Establish a mathematical optimization model for task unloading, with the optimization objective of minimizing task completion time and energy consumption, and use genetic algorithm or particle swarm optimization algorithm to solve the optimal unloading decision for each state; Method 3: Hybrid labeling – Initial labels are first generated through offline optimization calculations, and then boundary cases are manually corrected by experts, combining the advantages of both methods.

[0052] In S26e, the edge online fine-tuning steps specifically include: Edge nodes cache the actual operating data of this wind farm for the past 30 days, including the actual completion time, energy consumption and task success rate after each task unloading decision; Online fine-tuning is triggered when the number of locally cached samples reaches a preset threshold, or when the model's decision accuracy continuously decreases by more than a preset percentage. Edge nodes use cached local data to perform incremental training on the model in a few rounds, using a small learning rate to avoid catastrophic forgetting; Edge nodes report the fine-tuned model parameters and version information to the cloud. The cloud records the model version of each wind farm and synchronizes it to the corresponding backup edge nodes when needed. When the performance of the fine-tuned model deteriorates on the validation set, it is automatically rolled back to the model version before fine-tuning to ensure decision stability.

[0053] In S26d, the lightweighting process specifically includes: Remove connections with weights close to zero in deep Q-networks to reduce the number of model parameters; Convert the floating-point representation of model parameters to an 8-bit or 16-bit integer representation to reduce model storage and computational overhead; A smaller, lightweight student model is trained and deployed on edge nodes using a pre-trained large model as the teacher model. In S2, the backup edge node synchronously runs the anomaly detection model, maintaining a state consistent with the primary edge node, specifically including: Backup edge nodes receive the same multi-source heterogeneous data as the primary edge nodes in real time and run the same anomaly detection model. Backup edge nodes synchronize model status with primary edge nodes via a heartbeat channel, including model parameter versions, recent detection results, and warning records; When the primary edge node is working normally, the backup edge node only runs the model but does not output detection results; when the primary edge node fails, the backup edge node takes over the anomaly detection task and starts outputting detection results and warning information.

[0054] After the primary edge node completes the incremental update, the updated model parameters are synchronized to the backup edge node to ensure that the models of the primary and backup nodes remain consistent.

[0055] In some embodiments, in S3, the cloud calculates the total reactive power demand of each wind farm cluster and issues instructions, specifically including: S31. The cloud receives key feature data uploaded by the main edge nodes of each wind farm. The key feature data includes at least the real-time active power, reactive power, grid connection point voltage and current phase angle of each wind farm. S32. The cloud collects dispatch instructions issued by the power grid dispatch center, including target voltage curves, reactive power limits, and power factor requirements; at the same time, it collects real-time voltage, frequency, and power flow distribution data of the power grid's point of common coupling. S33. Based on the aforementioned key feature data, power grid dispatch instructions, and overall network operation status, the cloud platform calculates the total reactive power required for each wind farm cluster, with the optimization objective of minimizing the voltage deviation at the grid connection point, combined with the reactive power regulation capability constraints of the wind farm. S34. The cloud converts the calculated total reactive power demand into a reactive power allocation instruction, which includes at least the target reactive power value, the adjustment rate requirement, and the response time window; the reactive power allocation instruction is sent to the primary edge nodes of each wind farm and copied to the backup edge nodes.

[0056] In this embodiment of the invention, step S33, calculating the total reactive power, includes: S331. Calculate the deviation between the current voltage at the grid connection point of each wind farm and the target voltage in the dispatch instruction. ; S332. Based on the power grid topology and power flow distribution, calculate the sensitivity coefficient matrix K of the reactive power change of each wind farm to the voltage at the grid connection point. S333. Establish an optimization model with the objective of minimizing voltage deviation, where the objective function is: The constraints include the upper limit of reactive power regulation capacity for each wind farm. and lower limit , regulation rate limits and power factor requirements; among which, For the first Reactive power adjustment of a wind farm This represents the average reactive power adjustment of all wind farms. These are the balance weighting coefficients; S334. Solve the optimization model using a quadratic programming algorithm or a particle swarm optimization algorithm to obtain the total reactive power adjustment required by each wind farm cluster. .

[0057] In S34, the steps for generating reactive power allocation instructions include: S341, Adjust the total reactive power amount Convert the control instructions into a standard format, including the instruction type, target value, execution timestamp, and instruction validity period; S342. Add priority tags to the instructions according to the urgency of the grid voltage fluctuations. The priority tags include emergency regulation, normal regulation and preventive regulation. S343. The cloud uses a private key to digitally sign the instruction content to ensure the authenticity and integrity of the instruction source; S344. Based on the communication addresses of each wind farm, the signed instructions are sent out in parallel to the primary edge nodes of all wind farms, and the sending log is recorded in the cloud.

[0058] In this embodiment of the invention, the cloud also performs an instruction execution monitoring step: After the cloud issues the command, it continuously receives the command execution status returned by the main edge nodes of each wind farm. The execution status includes received, executing, completed and executed. For wind farms that fail to return to the execution status or fail to return to the execution status within the specified time, the cloud initiates a resend mechanism to resend the instructions to the edge nodes of the wind farm. If three consecutive retransmissions fail, the cloud will mark the wind farm as having a communication anomaly and adjust the reactive power allocation scheme of other wind farms to make up for the shortfall of the wind farm.

[0059] It should be further noted that the key feature data underwent data quality verification before being uploaded to the cloud: The cloud performs integrity checks on the received key feature data, and discards data packets with missing fields exceeding the threshold and requests retransmission. The cloud performs data validation to identify and remove data points that exceed physical limits or have abnormal rates of change. The cloud-based system employs multi-source data cross-validation, comparing multiple data sources from the same wind farm to identify and correct abnormal data.

[0060] In this embodiment of the invention, the backup edge node synchronously receives reactive power allocation instructions from the cloud during normal operation, but does not perform instruction distribution and monitoring functions; when the primary edge node fails, the backup edge node takes over instruction reception, execution status feedback, and communication tasks with the cloud.

[0061] In some embodiments, in S4, the primary edge node of each wind farm receives the reactive power allocation command, uses a consensus algorithm to allocate the total reactive power demand to each wind turbine in the farm, and performs voltage coordination control, such as... Figure 2 As shown, it specifically includes: S41. The primary edge node receives the reactive power allocation command sent from the cloud and parses it to obtain the overall reactive power target value, adjustment rate requirement and response time window that the wind farm needs to provide. S42. The primary edge node calculates the initial reactive power allocation scheme based on the current operating status and reactive power regulation capacity of each wind turbine in the field, and determines the initial reactive power output target for each wind turbine. S43. The primary edge node starts the consensus algorithm. Each wind turbine node exchanges its reactive power output status through the adjacent communication network and iteratively updates the reactive power allocation value until the reactive power output ratio of all wind turbines reaches a consensus. S44. The primary edge node will distribute the final agreed reactive power output target to every wind turbine in the field. S45. Each wind turbine performs reactive power output adjustment and feeds back the actual results to the primary edge node. The primary edge node monitors whether the total reactive power output of the entire field reaches the overall reactive power target value. If there is a deviation, a secondary adjustment is triggered.

[0062] It should be noted that in S43, the consensus algorithm adopts the average consensus protocol, and the specific iterative formula is as follows:

[0063] in, For the first Typhoon machine in reactive power output value at the next iteration In order to be with the first A group of wind turbines adjacent to each other. This is the iteration step size, which takes a value between 0 and 1; The iteration termination condition is: ,in This is the preset convergence accuracy threshold.

[0064] In some embodiments, the consensus algorithm considers the reactive power regulation capability constraints of the wind turbine and introduces a projection operation: After each iteration update, for A feasibility study will be conducted. If the problem exceeds the reactive power regulation capacity of the wind turbine, […]. If ], then project it onto the boundary values:

[0065] At the same time, to ensure that the total reactive power output of the entire site meets the requirements Requirement: Introduce a total quantity correction term.

[0066] in, This is the total correction factor. This represents the total number of wind turbines on site.

[0067] In further S43, the adjacent communication network is constructed based on the actual topology within the wind farm: An adjacency matrix is ​​established based on the electrical distance and communication link quality between wind turbines, with wind turbines that are close in electrical distance and have reliable communication being neighbor nodes. When some communication links are interrupted, the primary edge node dynamically adjusts the neighbor relationship and enables the backup communication path to ensure the connectivity of the consensus algorithm. For isolated wind turbine nodes, the primary edge node communicates directly with them, acting as a virtual neighbor to complete iterative calculations.

[0068] In step S42, the reactive power output target for each wind turbine is determined based on a comprehensive calculation of the following factors: Current active power output of the wind turbine Determine the available reactive power capacity range based on the wind turbine power curve. ]; Wind turbine converter capacity The constraints are ; Grid connection voltage of the wind turbine When the voltage approaches its upper limit, the reactive power generation capacity is limited; when it approaches its lower limit, the reactive power absorption capacity is limited. For maximum reactive power absorption capacity, This represents the maximum reactive power generation capacity. Initial allocation value Allocation based on the proportion of available reactive power capacity of each wind turbine: ; For the first Available reactive power capacity of typhoon generators.

[0069] In this embodiment of the invention, the monitoring step specifically includes: The primary edge node collects the actual reactive power output of each wind turbine in real time. Calculate the total actual reactive power of the entire field. ; Calculate the deviation between the actual total reactive power and the target value. ; If |ΔQ| exceeds the preset deviation threshold, and the current time is still within the response time window, then a secondary adjustment is initiated: ΔQ is used as the new allocation target, and the consistency iteration steps are re-executed. If the response time window has closed, the deviation value will be recorded and reported to the cloud, which will then compensate for it in the next scheduling cycle.

[0070] It should be further noted that backup edge nodes participate in the consistency iteration calculation synchronously under normal operating conditions, but do not issue commands externally. Backup edge nodes receive the same reactive power allocation instructions as primary edge nodes, run the same consensus algorithm, and synchronously update their own iterative states. Backup edge nodes synchronize intermediate results and the final allocation scheme with primary edge nodes via a heartbeat channel; When the primary edge node fails, the backup edge node continues to execute the consensus algorithm in the current synchronized iterative state and distributes the final allocation scheme to each wind turbine.

[0071] In some embodiments, when the primary edge node fails, the backup edge node takes over the control task and verifies the control commands using a lightweight consensus algorithm to ensure the authenticity and immutability of the transmitted commands. Specifically, this includes: A millisecond-level heartbeat detection mechanism is established between the primary edge node and the backup edge node. The two parties periodically send heartbeat messages to each other. The heartbeat messages include node identifier, timestamp, and current working status. When a backup edge node does not receive a heartbeat response from the primary edge node within N consecutive heartbeat cycles, the primary edge node is deemed to have failed. N is a value of 3-5. The backup edge node automatically switches to become the new primary edge node, taking over the Internet Protocol address, network connection and control tasks of the original primary edge node, and begins to perform data acquisition, processing and command issuance functions; The newly appointed master edge node establishes a consensus group with the edge nodes of other wind turbines in the field, initializes the lightweight consensus protocol, and prepares to conduct multi-party verification of the control commands to be issued. After the new master uses the edge node to generate control commands, it uses its own private key to sign the command content and generates a digital signature to be appended to the command. After receiving the instruction, other edge nodes in the consensus group use the public key of the new master edge node to verify the validity of the signature, confirming that the source of the instruction is genuine and that the content has not been tampered with. Once the verification is successful, each wind turbine executes the control command; if the verification fails, execution is refused and an anomaly is reported to the cloud.

[0072] In this embodiment of the invention, the lightweight consensus algorithm adopts a lightweight variant of the practical Byzantine fault-tolerant algorithm, specifically including: The consensus group consists of all primary edge nodes and backup edge nodes in the field, with a total number of nodes of 3f+1, where f is the number of allowed Byzantine nodes; The newly appointed master node uses an edge node as the master node to initiate a command proposal and broadcasts the signed command to all replica nodes; After receiving the proposal, the replica node verifies the validity of the signature and broadcasts a confirmation message to all other nodes; Once the master node receives valid confirmation messages from at least f+1 different replica nodes, it determines that the command has reached a consensus and broadcasts the commit message to the entire network. After receiving the commit message, each node executes the control command.

[0073] Furthermore, the consensus protocol has been optimized for lightweight design in wind farm control scenarios: A batch consensus mechanism is adopted to package multiple control commands into a batch for one-time consensus, thereby reducing the number of consensus interactions; An optimistic execution strategy is adopted. For urgent instructions with high timeliness requirements, execution is carried out first and consensus is reached later. If subsequent consensus fails, a rollback is performed through compensation instructions. Consensus messages are transmitted using UDP multicast to reduce network overhead, and message sequence numbers and acknowledgment retransmission mechanisms are added to ensure reliability.

[0074] It should also be noted that before the backup edge node takes over the control task, state synchronization must be completed: The backup edge node loads the most recent system status snapshot synchronized with the primary edge node from local storage. The status snapshot includes the current reactive power allocation scheme, the operating status of each wind turbine, and the queue of unfinished control commands. The backup edge node broadcasts a status query request to each wind turbine in the field to obtain the current real-time status of each wind turbine and compares and corrects it with the local snapshot; Once the state synchronization is complete, the backup edge node will begin issuing new control commands to ensure the continuity and consistency of the control process.

[0075] In some embodiments, the heartbeat message also carries status summary information: In each heartbeat cycle, the primary edge node appends a summary of the current key state data (such as the hash value of the most recent control command, model version number, and task queue length) to the heartbeat message. After receiving a heartbeat message, the backup edge node compares the summary information with its local status. If an inconsistency is found, it actively requests complete status data from the primary edge node for synchronization. By using heartbeats to carry digests, continuous incremental synchronization of the primary and backup node states is achieved, reducing the state synchronization overhead during sudden failures.

[0076] In this embodiment of the invention, after the backup edge node takes over the control task, it performs fault reporting and recovery: The newly appointed primary edge node reports the primary edge node failure event to the cloud, including the failure time, detection method, and takeover completion time. Failure events are recorded in the cloud and alarms are generated in the operation and maintenance system to notify on-site personnel to troubleshoot and repair the fault; After the original primary edge node is repaired and brought back online, it is added to the system as a new backup edge node, establishes heartbeat monitoring with the new primary edge node, and restores the primary / backup redundancy state.

[0077] The master / slave switching time is less than a preset threshold to ensure that the wind turbine can maintain stable operation in the event of a power grid failure or communication interruption, and to avoid grid disconnection accidents caused by data processing delays.

[0078] This invention also provides a distributed data processing and collaborative control system for wind farm clusters, comprising: Multiple dual-edge node units, each deployed on a wind turbine, include a primary edge node and a backup edge node that serve as each other's primary and backup edge nodes, used to collect multi-source heterogeneous data and process it locally; The cloud-based control platform communicates with the primary edge nodes of each wind farm to calculate the overall reactive power demand based on key feature data and grid dispatch instructions and to issue instructions. The primary edge node, backup edge node, and cloud control platform are configured to perform the methods described in the above embodiments.

[0079] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium such as a USB flash drive, mobile hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or other media capable of storing program code. It includes several instructions to cause a computer terminal (which may be a personal computer, server, or a second terminal, network terminal, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0080] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0081] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0082] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0083] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for distributed data processing and collaborative control of wind farm clusters, characterized in that, Each wind turbine is equipped with mutually supporting edge nodes. The primary and backup edge nodes synchronously collect multi-source heterogeneous data, and each performs data parsing and dimensionality reduction processing to obtain key feature data. Each node runs an anomaly detection model to monitor the status and generate early warning information. The mutually supporting edge nodes monitor each other through heartbeat detection. The method includes the following steps: When the primary edge node determines that the current processing task is a cloud-based collaborative computing task based on the current network latency, its own computing load, and energy consumption constraints through an adaptive decision model, it uploads the key feature data and early warning information to the cloud. Based on the aforementioned key feature data, combined with power grid dispatch instructions and the overall network operation status, the cloud calculates the overall reactive power demand of each wind farm cluster and sends the reactive power allocation instructions to the primary edge nodes of each wind farm and copies them to the backup nodes. Each primary edge node uses a consensus algorithm to distribute the total reactive power demand to each wind turbine in the field and performs voltage coordinated control; backup edge nodes synchronously participate in the consensus algorithm iteration. When the primary edge node fails, the backup edge node seamlessly takes over all functions, including data uploading, command reception and execution, and initiates a consensus algorithm to verify subsequent control commands through multi-party signatures. The cloud automatically sends subsequent commands to the new primary edge node.

2. The distributed data processing and collaborative control method for wind farm clusters according to claim 1, characterized in that, The data parsing is performed in real time on multi-source heterogeneous data using a node-level adaptive parsing algorithm, specifically including: Based on the computing power, storage resources, and network bandwidth of the edge nodes, the edge nodes in the wind farm cluster are divided into multiple levels; different levels of parsing templates are configured for edge nodes of different complexity, with high-level nodes performing complete data parsing and low-level nodes performing key field parsing. When collecting heterogeneous data from multiple sources, the edge nodes automatically identify the device brand and sensor type of the data source, and call the corresponding parsing template according to the preset device-template mapping table to achieve real-time hierarchical parsing of heterogeneous data without the need for unified format conversion.

3. The distributed data processing and collaborative control method for wind farm clusters according to claim 1, characterized in that, The dimensionality reduction process employs a redundancy elimination and compression strategy that integrates incremental computation and fixed-point computation, specifically including: For the continuous sampling data stream of the wind turbine, the change between the current sampled value and the previous sampled value is calculated. If the change exceeds a preset dynamic threshold, it is determined to be valid change data and is retained; if it does not exceed the dynamic threshold, it is determined to be redundant data and is not recorded temporarily. For periodic stationary data that continuously meet the incremental threshold condition within a preset time window, it is identified as a fixed-point sequence; the first sample value, the last sample value, and the trend indicator of the sequence are recorded, and intermediate sample data are not retained; The first and second buffers are configured to work alternately. The first buffer is used for real-time writing of the current data cycle, while the second buffer is used for data processing of the previous data cycle. Zero-wait data processing is achieved through the alternation of the two buffers.

4. The distributed data processing and collaborative control method for wind farm clusters according to claim 1, characterized in that, The anomaly detection model is a lightweight long short-term memory network model. Key feature data is input into this model for real-time monitoring and to generate early warning information. Specifically, it includes: The key feature data after dimensionality reduction are used to form a time series input vector. The key feature data includes at least the wind turbine vibration feature value, generator speed, wind speed, active power and blade angle. The lightweight long short-term memory network anomaly detection model performs real-time inference on the time series input vector and outputs the anomaly probability value of the current running state and the predicted anomaly type. When the anomaly probability value exceeds the preset warning threshold, a warning message is generated that includes the anomaly type, the time of the anomaly, the anomaly probability value, and the number of the wind turbine in which it occurs. Edge nodes receive model update parameters from the cloud and incrementally update the anomaly detection model so that the model can adapt to the changes in operating conditions caused by the long-term operation of the wind turbine.

5. The distributed data processing and collaborative control method for wind farm clusters according to claim 1, characterized in that, Based on current network latency, its own computing load, and energy consumption constraints, when an adaptive decision model determines that the current processing task is a cloud-based collaborative computing task, the key feature data and early warning information are uploaded to the cloud, specifically including: The primary edge node is used to perceive the current environmental status in real time, and the environmental status includes at least: network status, node status, and task characteristics. The environmental state is input into an adaptive decision model deployed on an edge node, and the unloading decision for the current task is output. The unloading decision includes two types: local execution and cloud collaborative computing. When the decision result is cloud-based collaborative computing, the corresponding key feature data and early warning information are encapsulated into an upload data package, timestamps and node identifiers are added, and then uploaded to the cloud.

6. The distributed data processing and collaborative control method for wind farm clusters according to claim 1, characterized in that, The cloud calculates the total reactive power demand of each wind farm cluster and issues instructions, specifically including: The cloud receives key feature data uploaded by the main edge nodes of each wind farm. The key feature data includes at least the real-time active power, reactive power, grid connection point voltage and current phase angle of each wind farm. The cloud-based system collects dispatch instructions issued by the power grid dispatch center, including target voltage curves, reactive power limits, and power factor requirements; it also collects real-time voltage, frequency, and power flow distribution data at the power grid's point of common coupling. Based on the aforementioned key feature data, power grid dispatch instructions, and overall network operation status, the cloud platform calculates the total reactive power required for each wind farm cluster, with the optimization objective of minimizing the voltage deviation at the grid connection point, combined with the reactive power regulation capability constraints of the wind farms. The cloud platform converts the calculated total reactive power demand into reactive power allocation instructions, which include at least the target reactive power value, adjustment rate requirement, and response time window. The reactive power allocation instructions are then sent to the primary edge nodes of each wind farm and copied to the backup edge nodes.

7. The distributed data processing and collaborative control method for wind farm clusters according to claim 1, characterized in that, The primary edge nodes of each wind farm use a consensus algorithm to distribute the total reactive power demand to each wind turbine within the farm, specifically including: The primary edge node receives reactive power allocation instructions from the cloud and parses them to obtain the overall reactive power target value that the wind farm needs to provide. Adjustment rate requirements and response time window; The primary edge node calculates the initial reactive power allocation scheme based on the current operating status and reactive power regulation capacity of each wind turbine in the field, and determines the initial reactive power output target for each wind turbine. The primary edge node initiates the consensus algorithm, and each wind turbine node exchanges its reactive power output status through the adjacent communication network, iteratively updating the reactive power allocation value until the reactive power output ratio of all wind turbines reaches a consensus; the backup edge node synchronously participates in the iterative calculation of the consensus algorithm. The primary edge node will distribute the final agreed reactive power output target to every wind turbine in the field; Each wind turbine performs reactive power output adjustment and feeds back the actual results to the primary edge node; the primary edge node monitors whether the total reactive power output of the entire site has reached the target. If a deviation exists, a secondary adjustment will be triggered.

8. The distributed data processing and collaborative control method for wind farm clusters according to claim 7, characterized in that, The consensus algorithm adopts the average consensus protocol, and the specific iterative formula is as follows: in, For the first Typhoon machine in reactive power output value at the next iteration In order to be with the first A group of wind turbines adjacent to each other. The iteration step size, which takes a value between 0 and 1; The iteration termination condition is: ,in This is the preset convergence accuracy threshold.

9. The distributed data processing and collaborative control method for wind farm clusters according to claim 1, characterized in that, When the primary edge node fails, the backup edge node takes over all functions and ensures instruction security through a lightweight consensus algorithm, specifically including: A millisecond-level heartbeat detection mechanism is established between the primary edge node and the backup edge node. The two parties periodically send heartbeat messages to each other. The heartbeat messages include node identifier, timestamp, and current working status. If the backup edge node does not receive a heartbeat response from the primary edge node within N consecutive heartbeat cycles, the primary edge node is deemed to have failed. The backup edge node automatically switches to the new primary edge node, taking over all the functions of the original primary edge node, including data uploading, command reception and execution; The newly appointed master edge node establishes a consensus group with the edge nodes of other wind turbines in the field, initializes the lightweight consensus protocol, and prepares to conduct multi-party verification of the control commands to be issued. After the new master uses the edge node to generate control commands, it uses its own private key to sign the command content and generates a digital signature to be appended to the command. After receiving the instruction, other edge nodes in the consensus group use the public key of the new master edge node to verify the validity of the signature, confirming that the source of the instruction is genuine and that the content has not been tampered with. Once the verification is successful, each wind turbine executes the control command; if the verification fails, execution is refused and an anomaly is reported to the cloud.

10. A distributed data processing and collaborative control system for wind farm clusters, characterized in that, include: Multiple dual-edge node units, each deployed on a wind turbine, include a primary edge node and a backup edge node that serve as each other's primary and backup edge nodes. These units are used to synchronously collect heterogeneous data from multiple sources and independently perform data parsing, dimensionality reduction processing, anomaly detection, and early warning information generation. The cloud-based control platform communicates with the primary edge nodes of each wind farm to calculate the overall reactive power demand based on key feature data and grid dispatch instructions, and issues instructions while simultaneously copying the instructions to the corresponding backup edge nodes. The primary edge node, backup edge node, and cloud control platform are configured to perform the method described in any one of claims 1-9.

Images