A cloud-based collaborative variable frequency fan cluster energy-saving regulation method

Abstract

The present application relates to the technical field of fan control, and particularly relates to a variable frequency fan cluster energy-saving regulation and control method based on cloud cooperation. The present application comprises: constructing and training a full-capability energy-saving regulation and control model on a cloud node according to historical fan regulation and control records; based on the cluster edge node computing power of a target variable frequency fan cluster, performing model compression on the full-capability energy-saving regulation and control model to obtain a high-efficiency energy-saving regulation and control model; performing task analysis on real-time task demand information and fan operation state data to obtain a double-end task set, the double-end task set comprising a cloud task subset and an on-site task subset; performing cloud cooperative regulation and control scheme decision to obtain on-site regulation and control scheme decision results and cloud regulation and control scheme decision results; and performing energy-saving regulation and control on the target variable frequency fan cluster. The present application realizes organic cooperation between the cloud and the local site, improves the real-time performance and accuracy of regulation and control, and enhances the energy-saving and stability technical effects.

Description

Technical Field

[0001] This invention relates to the field of wind turbine control technology, and in particular to a cloud-based collaborative method for energy-saving control of variable frequency wind turbine clusters. Background Technology

[0002] Variable frequency drives (VFDs) can adjust their speed according to actual needs, thereby achieving energy-saving operation. However, traditional VFD control methods mostly rely on local control systems, with control decisions based on limited historical data and local computing power. When facing complex industrial environments and changing task requirements, they often fail to achieve optimal energy-saving effects. In recent years, the rise of cloud technology has provided new ideas for VFD control. Although cloud technology offers powerful computing capabilities and abundant data resources, relying entirely on the cloud for energy-saving control of VFDs in practical applications also presents many problems. On the one hand, data transmission between the cloud and the local fan system may experience latency and bandwidth limitations, affecting the real-time performance and response speed of control. On the other hand, the centralization of cloud computing resources may lead to reduced computing and data processing efficiency when dealing with large-scale fan clusters, making it difficult to meet the needs of real-time control. Furthermore, over-reliance on the cloud may cause the system to lose its control capabilities in the event of cloud failure or network interruption, affecting the stable operation of the fan system. Summary of the Invention

[0003] The purpose of this invention is to solve the problem that the lack of organic coordination between cloud and local decision-making in the existing technology affects the real-time performance and accuracy of regulation, and is not conducive to energy saving and stability. The invention proposes a cloud-based collaborative variable frequency fan cluster energy-saving regulation method to achieve organic coordination between cloud and local systems, thereby improving the real-time performance and accuracy of regulation, and enhancing the technical effects of energy saving and stability.

[0004] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0005] A cloud-based collaborative method for energy-saving control of variable frequency wind turbine clusters includes:

[0006] Collect historical wind turbine control records, which include control parameters and control evaluations;

[0007] Based on historical wind turbine control records, a full-capacity energy-saving control model is built and trained on cloud nodes, and then deployed on cloud nodes.

[0008] Based on the computing power of the cluster edge nodes of the target variable frequency wind turbine cluster, the full-capacity energy-saving control model is compressed to obtain a high-efficiency energy-saving control model, which is then deployed on the cluster edge nodes.

[0009] Obtain real-time task requirements and turbine operating status data for the target variable frequency wind turbine cluster;

[0010] The real-time task requirement information and wind turbine operation status data are analyzed to obtain a dual-end task set, which includes a cloud task subset and a local task subset.

[0011] Based on the dual-end task set, the full-capacity energy-saving control model and the high-efficiency energy-saving control model, cloud-based collaborative control scheme decision-making is carried out to obtain the local control scheme decision-making results and the cloud-based control scheme decision-making results.

[0012] Energy-saving control is implemented for the target variable frequency wind turbine cluster based on the decision results of the local control scheme and the decision results of the cloud-based control scheme.

[0013] Furthermore, the step of constructing and training a full-capacity energy-saving control model on cloud nodes based on historical wind turbine control records includes:

[0014] Historical wind turbine control records are filtered based on preset control evaluation thresholds;

[0015] The screening results are preprocessed and uploaded to the cloud node;

[0016] Using the preprocessed results as training data, a full-capacity energy-saving regulation model is constructed and trained on cloud nodes.

[0017] Furthermore, based on the computing power of the cluster edge nodes of the target variable frequency wind turbine cluster, the full-capacity energy-saving control model is compressed to obtain a high-efficiency energy-saving control model, including:

[0018] The computing power of the cluster edge nodes is evaluated to obtain the available computing power index;

[0019] Set model compression constraint parameters based on available computing power indicators;

[0020] The full-capacity energy-saving control model is compressed based on the compression constraint parameters to obtain a high-efficiency energy-saving control model.

[0021] Based on compression constraint parameters, a high-efficiency energy-saving control model is simulated and deployed, and the simulated computing power consumption is obtained.

[0022] If the simulated computing power consumption meets the available computing power index, then a high-efficiency energy-saving control model is output.

[0023] If the simulated computing power consumption does not meet the available computing power index, the constraint parameters are adjusted and the model is compressed again until the simulated computing power consumption does not meet the available computing power index.

[0024] Furthermore, the process of parsing the real-time task requirement information and the wind turbine operating status data to obtain a dual-end task set includes:

[0025] The real-time task requirements information is decomposed to obtain a real-time task set.

[0026] Using the wind turbine ID as an index, establish a mapping relationship between real-time task sets and historical wind turbine control records;

[0027] The complexity of the control decision for each real-time task is obtained by calling historical wind turbine control records based on the mapping relationship.

[0028] Based on the complexity of control decisions and the preset real-time requirements, the real-time task set is divided into a cloud task subset and a local task subset, resulting in a dual-end task set.

[0029] Furthermore, the step of parsing the real-time task requirement information and wind turbine operating status data to obtain a dual-end task set also includes:

[0030] After obtaining the real-time task set, cluster analysis is performed on the real-time task set based on the model parameters of the variable frequency fan and the fan operating status data to generate a simplified task set;

[0031] The simplified task set will be used as the new real-time task set.

[0032] Furthermore, the cloud-based collaborative control scheme decision-making based on the dual-end task set, the full-capacity energy-saving control model, and the high-efficiency energy-saving control model yields both the local control scheme decision results and the cloud-based control scheme decision results, including:

[0033] A subset of local tasks will be transmitted to a high-efficiency energy-saving control model to make local control scheme decisions, thereby generating local control scheme decision results;

[0034] A subset of cloud tasks is synchronously uploaded to the full-capacity energy-saving control model of the cloud node to execute cloud control scheme decisions, thereby generating the cloud control scheme decision results.

[0035] Furthermore, the process of making cloud-based collaborative control scheme decisions based on dual-end task sets, a full-capacity energy-saving control model, and a high-efficiency energy-saving control model, to obtain both local and cloud-based control scheme decision results, also includes:

[0036] Based on a preset crossover ratio, the local task subset and the cloud task subset are randomly selected to obtain the uplink crossover task set and the downlink crossover task set.

[0037] The uplink cross-task set and the downlink cross-task set are added to the cloud task subset and the local task subset respectively, and cloud-based collaborative control scheme decision is made.

[0038] Based on the uplink cross-task set and the downlink cross-task set, cross-validate the decision results of the local control scheme and the decision results of the cloud control scheme.

[0039] If the cross-validation passes, the decision results of the local control scheme and the decision results of the cloud-based control scheme will be output.

[0040] If cross-validation fails, the full-capacity energy-saving control model will be retrained, and tasks will be redistributed to obtain a new dual-end task set or control scheme fusion.

[0041] Furthermore, the energy-saving control of the target variable frequency wind turbine cluster based on the decision results of the local control scheme and the decision results of the cloud-based control scheme includes:

[0042] The decision-making results of local control plans will be integrated with those of cloud-based control plans.

[0043] The integrated control commands are sent to the corresponding ID of the fan in the target variable frequency fan cluster;

[0044] Each wind turbine adjusts its own operating parameters according to the control instructions it receives.

[0045] Furthermore, it also includes:

[0046] Based on a preset constraint period, the decision results of the local control scheme and the corresponding local tasks for the current constraint period are randomly extracted and uploaded to the cloud node.

[0047] Verification and control decisions are made by cloud nodes based on a full-capacity energy-saving control model;

[0048] Based on the preset decision verification constraints, the decision verification results are compared with the decision results of the local control scheme to obtain the decision verification results;

[0049] Optimize the corresponding high-efficiency energy-saving control model based on the decision verification results.

[0050] Furthermore, optimizing the corresponding high-efficiency energy-saving control model based on the decision verification results includes:

[0051] Calculate the decision residuals based on the decision verification results, and construct a reinforcement training dataset based on the decision residuals;

[0052] The high-efficiency energy-saving control model is trained using a reinforcement training dataset;

[0053] The high-efficiency energy-saving control model after reinforcement learning is optimized, verified, and iterated until the optimization and verification results meet the decision verification constraints.

[0054] Update and deploy high-efficiency energy-saving control models.

[0055] Compared with existing technologies, the significant advantages of this invention are as follows: 1. This invention places complex global optimization calculations in the cloud while keeping real-time control requiring rapid response local, achieving both high precision in macro-level decision-making and low latency in micro-level control, fundamentally improving the overall control quality. 2. By delegating some tasks to the cloud and edge for parallel processing and comparing the results, a built-in fault tolerance and verification mechanism is provided for the system, greatly enhancing the credibility of the output decisions and the operational stability of the system. 3. Through rigorous compression and verification processes, it ensures that advanced energy-saving algorithms can accurately adapt to local hardware with different capabilities, enabling various field devices to obtain decision-making capabilities stably and efficiently. 4. It possesses periodic self-verification and model update capabilities, continuously adapting to changes in operating conditions and maintaining optimal energy-saving status over the long term. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a cloud-based collaborative method for energy-saving control of variable frequency fan clusters according to the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0058] like Figure 1 As shown, this invention provides a cloud-based collaborative method for energy-saving control of variable frequency wind turbine clusters, the specific steps of which include:

[0059] S1: Collect historical wind turbine control records, including control parameters and control evaluations.

[0060] Specifically, the wind turbine control record refers to the historical control behavior data generated and stored by the local controller or cloud control platform during the operation of the variable frequency wind turbine. This record includes control parameters and control evaluation. Control parameters refer to the control command parameters applied to the wind turbine within a specific time period, such as target frequency, speed setpoint, output power, operating mode (energy saving / high power), PID control coefficient, etc. Control evaluation refers to the feedback information on the effect of the control behavior after execution, used to evaluate the energy-saving effect, operational stability, and user satisfaction of this control, typically including indicators such as energy consumption change, temperature rise change, noise level, and system response time.

[0061] Optionally, control records can be stored in time series or clustered by scenario tags (such as high load, low load, nighttime operation, etc.) and synchronized to the cloud for unified management and analysis.

[0062] The above process is used to collect and record structured data on the entire process of wind turbine regulation behavior, thereby providing a data foundation for subsequent regulation optimization and coordination strategies. This helps to more accurately learn the operating rules and regulation effects of wind turbines under different operating conditions, thereby improving decision-making capabilities.

[0063] S2: Based on historical wind turbine control records, a full-capacity energy-saving control model is built and trained on cloud nodes, and the model is compressed by combining the computing power of the cluster edge nodes of the target variable frequency wind turbine cluster to obtain a high-efficiency energy-saving control model.

[0064] Specifically, cloud nodes refer to computing units within cloud servers, possessing powerful computing and data storage capabilities to support the construction and training of complex models. The full-capability energy-saving control model is a highly complex model trained using historical data through algorithms such as deep learning, reinforcement learning, or other machine learning methods, possessing cross-scenario energy-saving control capabilities. This model has strong generalization ability, adapting to various operating scenarios and load conditions to achieve energy-saving optimization control of wind turbine clusters.

[0065] Specifically, the computing power of the cluster edge nodes refers to the computing capabilities of edge computing devices in a variable frequency wind turbine cluster. These devices are deployed close to the wind turbines and can quickly process local data and execute real-time control tasks. The high-efficiency energy-saving control model, on the other hand, is a model processed using model compression technology. It is designed to adapt to the limited computing power of edge nodes while retaining as much of the energy-saving control performance of the full-capacity model as possible, so that it can be deployed at the edge for real-time operation.

[0066] In some embodiments, step S2 includes:

[0067] Historical wind turbine control records are filtered based on preset control evaluation thresholds, the filtering results are preprocessed, and the preprocessed results are uploaded to the cloud node.

[0068] The cloud nodes use the preprocessed results as training data to build and train a full-capacity energy-saving control model.

[0069] Specifically, the control evaluation threshold is a set of standard parameters used to assess the effectiveness and representativeness of historical wind turbine control records. This threshold may include, but is not limited to: energy consumption reduction of not less than a certain percentage (e.g., ≥5%); system temperature fluctuation not exceeding the set range (e.g., ±1℃); user satisfaction rating higher than the set level (e.g., ≥3 stars); and wind turbine operation stability indicators (e.g., no alarms, no frequent start-stops) meeting the requirements.

[0070] Specifically, the edge nodes first perform a preliminary screening of the historical wind turbine control records stored locally, removing invalid or abnormal samples based on preset control evaluation thresholds. For example, if energy consumption decreases by less than 3% after a certain control, the record is considered inefficient control and will not be included in subsequent training. Then, the collected historical wind turbine control records are preprocessed, such as through standardization, normalization, and feature construction, to eliminate dimensional differences between different parameters and improve data quality and the efficiency and accuracy of subsequent model training. The preprocessed data is then packaged, encrypted, and uploaded to the cloud nodes to form a structured training dataset.

[0071] Next, based on the preprocessed data, a full-capacity energy-saving control model is constructed using machine learning algorithms (such as neural networks and support vector machines). The model parameters are continuously optimized through multiple iterations of training, enabling the model to accurately predict the energy-saving control strategies of wind turbines under different operating conditions.

[0072] Through the above process, the system can achieve high-quality screening and structured utilization of historical regulatory data, significantly improving the effectiveness and generalization ability of model training. Compared with directly using the entire historical data for training, this method can effectively avoid the interference of inefficient or abnormal regulatory behaviors on model performance and provide a solid foundation for subsequent model compression and edge deployment.

[0073] In some embodiments, step S2 further includes:

[0074] The computing power of the cluster edge nodes is assessed to obtain available computing power metrics.

[0075] Model compression constraint parameters are set based on available computing power indicators.

[0076] The full-capacity energy-saving control model is compressed based on the compression constraint parameters to obtain a high-efficiency energy-saving control model.

[0077] The high-efficiency energy-saving control model is simulated and deployed based on compression constraint parameters, and the simulated computing power consumption is obtained. If the simulated computing power consumption meets the available computing power index, the high-efficiency energy-saving control model is output. If the simulated computing power consumption does not meet the available computing power index, the constraint parameters are adjusted and the model is compressed again until the simulated computing power consumption is not higher than the available computing power index.

[0078] Specifically, the computing power of a wind turbine cluster edge node refers to the computing capabilities of edge computing devices (deployed near the wind turbines for local data processing and control) within the wind turbine cluster. It determines the model complexity that the edge devices can handle and is typically quantified using metrics such as the number of CPU cores, memory size, and floating-point arithmetic capabilities. Available computing power metrics are the specific quantification results of the edge node's computing power, determined through evaluation, and used to set constraints for model compression, such as the maximum allowable computational complexity or the upper limit on the number of model parameters.

[0079] Model compression constraint parameters are set based on available computing power to limit the model size and computational complexity during the compression process, thereby ensuring that the compressed model can run efficiently on edge nodes. The high-efficiency energy-saving control model is the compressed version of the model, designed to retain the energy-saving control performance of the full-capability model while meeting the computing power limitations of edge nodes, so as to achieve rapid local energy-saving control.

[0080] First, a comprehensive assessment of the computing power of the cluster edge nodes is required, including key indicators such as the number of CPU cores, clock speed, memory size, floating-point operation capability, and storage bandwidth, to determine their available computing power. Next, based on the assessed available computing power, constraint parameters for model compression are set. For example, in a typical scenario, the available memory of the edge node is 512MB, and the maximum acceptable inference latency is 80ms. Based on this, the system sets the model compression constraint parameters as follows: model size ≤ 400MB, inference time ≤ 70ms (70ms is lower than the maximum acceptable inference latency of 80ms to ensure that the model meets performance requirements while having a certain amount of redundancy), and FLOPs ≤ 5G.

[0081] Based on the set compression constraints, the cloud-trained full-capacity energy-saving control model is compressed. This compression includes pruning (removing weight connections that have little impact on the prediction results), quantization (quantizing high-precision weight values ​​into low-precision representations, such as quantizing 32-bit floating-point values ​​into 8-bit integers), and knowledge distillation (distilling knowledge from the large model into the small model). After compression, the critical path of the model is preserved, while non-critical feature channels are removed, thus significantly reducing the model complexity.

[0082] Furthermore, after model compression, a simulation deployment is performed in the simulation environment to verify whether the compressed model meets the available computing power indicators. During the simulation deployment, the computing power consumption of the model during runtime is monitored in real time, including CPU utilization, memory usage, and inference time. If the simulation results show that the model's computing power consumption is within the available computing power indicators, for example, CPU utilization is below 80%, memory usage is below 70%, and the inference time meets the real-time requirements, and if all indicators are within the available range, the model compression is confirmed to be successful, and a high-efficiency energy-saving control model is output and prepared for actual deployment on edge nodes. Conversely, if the simulation results exceed the available computing power indicators, the model compression constraint parameters need to be adjusted and the model compression re-performed until the conditions are met.

[0083] Through the above process, the system can achieve on-demand compression and adaptive deployment of the energy-saving control model at edge nodes, ensuring stable operation of the model under different computing power conditions and avoiding deployment failures or operational lag caused by excessively large models. Simultaneously, this method introduces multi-dimensional resource constraints and performance preservation mechanisms during compression (by setting compression constraint parameters such as model size and inference time, and evaluating computing power consumption during simulated deployment, ensuring that the model still meets core control requirements after compression), maximizing the preservation of the model's core control capabilities, significantly improving the model's response speed and energy efficiency ratio at the edge, and providing a reliable guarantee for real-time energy-saving control of variable frequency wind turbine clusters.

[0084] S3: Obtain real-time task requirements and wind turbine operating status data of the target variable frequency wind turbine cluster, and perform task parsing to obtain a dual-end task set, which includes a cloud task subset and a local task subset.

[0085] Specifically, real-time task requirement information refers to the specific requirements of the production or operation tasks that the target variable frequency fan cluster needs to complete at the current moment or in a short future period, including but not limited to production tasks, required air volume, pressure, and temperature control range. Fan operation status data covers the real-time operating parameters of the fans, such as speed, power consumption, vibration, operating time, wind speed, air volume, load, current, voltage, etc., as well as the health status of the equipment.

[0086] The complexity of control decisions refers to a measure of a composite index, including the intensity of computing resources required to complete a specific wind turbine control task, the depth of model calls, and the degree of dependence on historical data. It is used to assess whether the task is suitable for real-time completion on edge nodes or needs to be sent to the cloud for centralized processing. The cloud task subset includes complex tasks that require powerful computing capabilities and big data support, which are processed in the cloud; the on-site task subset includes simple tasks that require fast response and high real-time requirements, which are processed on local edge nodes.

[0087] By acquiring real-time task requirements and turbine operating status data from the target variable frequency wind turbine cluster, and performing task parsing to obtain dual-end task sets, the system can achieve reasonable task allocation and efficient processing. On one hand, the cloud can utilize its powerful computing capabilities to process complex task subsets, perform big data analysis and complex model predictions, and optimize control strategies; on the other hand, local edge nodes can quickly respond to simple task subsets, achieve real-time control, and reduce data transmission latency and bandwidth consumption.

[0088] In some embodiments, step S3 includes:

[0089] The real-time task requirements information is decomposed based on the wind turbine ID to obtain a real-time task set.

[0090] Using wind turbine operating status data as an index, a mapping relationship between real-time task sets and historical wind turbine control records is established. Based on the mapping relationship, historical wind turbine control records are called for statistical analysis to determine the control decision complexity of each real-time task.

[0091] Based on the complexity of control decisions and the requirements for real-time performance, the real-time task set is divided into a cloud task subset and a local task subset, forming a dual-end task set.

[0092] Specifically, the process begins by acquiring real-time task requirements information for the target variable frequency wind turbine cluster and the operating status data of each turbine. Real-time task requirements information may include temperature control targets, energy-saving strategies, and user-defined values. Subsequently, using the turbine ID as an index, the real-time task requirements information is decomposed into tasks, generating a more granular set of real-time tasks, with each task corresponding to a specific control objective for one or a type of wind turbine.

[0093] Next, the current wind turbine operating status data is matched with historical wind turbine control records to establish a mapping relationship between real-time tasks and historical control behaviors. For example, if a wind turbine is currently under high load and high temperature conditions, the system will retrieve historical control strategies and execution results under similar conditions, perform statistical analysis, assess the computational complexity, model call path, and response requirements required for the task, and thus determine the control decision complexity of the task.

[0094] Furthermore, considering the complexity of the task's control and decision-making processes and the pre-defined real-time requirements, a pre-defined task partitioning strategy (such as based on resource thresholds or latency tolerance) is adopted to divide the real-time task set into: a cloud task subset (suitable for complex computation, high-dimensional model inference, global optimization scheduling, and other tasks) and a local task subset (suitable for rapid response, state closed-loop control, local model inference, and other tasks), thereby forming the final dual-end task set, providing a task foundation for subsequent collaborative control.

[0095] Through the above process, the system achieves intelligent parsing and dual-end assignment of variable frequency wind turbine cluster control tasks, significantly improving the execution efficiency and resource utilization of control tasks, avoiding communication delays caused by all tasks being uploaded to the cloud, and preventing resource overload on edge nodes. Simultaneously, the mapping analysis mechanism based on historical wind turbine control records enables experience-driven task assignment, dynamically judging task complexity and resource matching, thereby achieving high-level cloud-edge collaborative task scheduling.

[0096] In some embodiments, step S3 further includes:

[0097] Based on the model parameters and operating status data of the variable frequency fan, cluster analysis is performed on the real-time task set, and a simplified task set is generated according to the cluster analysis results. The simplified task set is then traversed to parse the tasks, forming a dual-end task set.

[0098] Specifically, model parameters include the fan's rated power, rated speed, maximum air volume, and motor type. These parameters determine the fan's basic performance and operating range. The simplified task set refers to a representative task set generated by clustering the original real-time task set according to the similarity between the fan model parameters and operating status. It contains multiple task clusters with similar characteristics, each representing a similar type of task requirement. This is used to reduce the number of tasks, decrease computational redundancy, and improve task scheduling efficiency without affecting control accuracy.

[0099] After generating the initial real-time task set, the task set is further feature-encoded based on the model parameters and current operating status information of each wind turbine. For example, the model parameters are standardized into category vectors, and the operating status parameters are normalized and combined into feature vectors to form the feature representation of each task. Subsequently, clustering algorithms (such as K-means, DBSCAN, or hierarchical clustering based on cosine similarity) are used to perform cluster analysis on the above task features, resulting in several task clusters. Each task cluster represents a class of wind turbine tasks with similar control requirements and execution conditions.

[0100] Furthermore, for each task cluster, a central task or representative task is selected to form a simplified task set to replace the original task set for subsequent parsing, thus forming a dual-end task set.

[0101] The above process can significantly reduce the number of tasks that need to be parsed independently, reduce the system's computing burden, make task parsing more efficient, and enable tasks to be quickly distributed to the cloud or local for processing.

[0102] S4: Combine the dual-end task set, the full-capacity energy-saving control model and the high-efficiency energy-saving control model to make cloud-based collaborative control scheme decisions and obtain the local control scheme decision results and the cloud-based control scheme decision results.

[0103] Specifically, after dividing the task sets into two parts, the corresponding control models are invoked to calculate energy-saving solutions for the cloud task subset and the on-site task subset, respectively.

[0104] For a subset of cloud-based tasks, the system invokes a full-capacity energy-saving control model deployed on the cloud platform. This model can generate control schemes for complex tasks based on a global perspective, considering factors such as the overall load distribution of the wind turbine group, historical control effects, and external environment predictions. It features high precision and high complexity.

[0105] For the local task subset, the system calls the high-efficiency energy-saving control model deployed on the edge nodes. This model uses local state awareness and lightweight model inference to quickly generate local control schemes that are highly adaptable and efficient, meeting the real-time response requirements.

[0106] Optionally, the two control schemes can be integrated or optimized based on factors such as task priority, model confidence, and execution cost to further improve the overall energy-saving effect and response efficiency of the system.

[0107] Through the above process, the system realizes a collaborative generation mechanism for model-level control schemes based on dual-end task sets. This fully leverages the complementary advantages of cloud and edge models in terms of computing power, response speed, and global perspective, while balancing system performance, energy consumption optimization, and response timeliness. Specifically, the cloud-based full-capability model provides high accuracy and global optimization capabilities, suitable for handling complex, nonlinear, and multi-objective control tasks; while the edge-based high-efficiency model provides rapid response capabilities, suitable for tasks with high real-time requirements and demanding local optimization needs.

[0108] In some embodiments, step S4 includes:

[0109] The high-efficiency energy-saving control model is deployed locally to the cluster edge nodes of the target variable frequency wind turbine cluster.

[0110] Transmit a subset of local tasks to the high-efficiency energy-saving control model, make local control scheme decisions, and generate local control scheme decision results.

[0111] The full-capacity energy-saving control model synchronously uploads a subset of tasks from the cloud to the cloud node, executes cloud control scheme decisions, and generates cloud control scheme decision results.

[0112] Specifically, cluster edge nodes are edge computing devices deployed in the local network of the target variable frequency wind turbine cluster. They possess data access, model inference, and local control capabilities, and are used to support high-efficiency energy-saving control models, enabling low-latency on-site task response. The on-site task subset consists of local control tasks suitable for execution via edge nodes within the dual-end task set, characterized by short response times, low data dependencies, and fast control loops. The cloud task subset comprises complex control tasks suitable for unified scheduling and optimization by the cloud, such as cross-wind turbine collaboration and global energy efficiency assessment.

[0113] First, the high-efficiency energy-saving control model is deployed and initialized locally on the edge nodes of the target variable frequency wind turbine cluster, enabling rapid loading and inference capabilities at the edge. Then, a subset of local tasks is transmitted to the deployed high-efficiency energy-saving control model, where edge nodes perform local inference and control strategy generation based on real-time status data, outputting the local control scheme decision results. Simultaneously, a subset of cloud tasks is uploaded to the cloud nodes, invoking the full-capacity energy-saving control model for strategy deduction, generating cloud-based control scheme decision results. By combining the dual-end task sets, the full-capacity energy-saving control model, and the high-efficiency energy-saving control model for cloud-based collaborative control scheme decision-making, the system fully leverages the advantages of cloud and edge computing to achieve optimal energy-saving control results.

[0114] In some embodiments, step S4 further includes:

[0115] Based on a preset crossover ratio, a subset of local tasks is randomly selected to obtain an uplink crossover task set.

[0116] Based on a preset cross ratio, a subset of cloud tasks is randomly selected to obtain a downlink cross task set.

[0117] The uplink cross-task set and the downlink cross-task set are added to the cloud task subset and the local task subset respectively, and cloud-based collaborative control scheme decision is made.

[0118] Based on the uplink cross-task set and the downlink cross-task set, cross-validate the decision results of the local control scheme and the decision results of the cloud control scheme. If the cross-validation passes, output the decision results of the local control scheme and the decision results of the cloud control scheme.

[0119] Specifically, the uplink cross-task set is a subset of tasks randomly selected from the local task subset based on a preset cross-task ratio. This subset of tasks is transmitted uplink to the cloud node and processed by the full-capability energy-saving control model to cross-validate the control results of the edge model.

[0120] The downlink cross-task set is another set of tasks randomly selected from the cloud task subset based on a preset cross-task ratio. This set of tasks is transmitted downlink to the edge node and processed by the high-efficiency energy-saving control model to cross-validate the control results of the cloud model.

[0121] Cross-validation refers to the process of sharing a subset of tasks between cloud and edge models and comparing and evaluating the consistency of their control results in order to discover potential biases between models and thereby improve the stability and reliability of the overall control scheme.

[0122] Specifically, to improve the accuracy and consistency of coordinated control, firstly, based on preset crossover ratios (e.g., 10%, 20%), tasks are randomly selected from both the local task subset and the cloud task subset to form an uplink crossover task set and a downlink crossover task set. For example, if the local task subset has 100 tasks and the cloud task subset has 200 tasks, selecting tasks at a 10% ratio results in an uplink crossover task set of 10 tasks and a downlink crossover task set of 20 tasks. Next, the uplink crossover task set is added to the existing cloud task subset and processed by the full-capacity energy-saving control model; the downlink crossover task set is added to the existing local task subset and processed by the high-efficiency energy-saving control model.

[0123] Furthermore, the updated local control scheme decision results and cloud control scheme decision results are obtained separately. The decision results of the uplink cross task set generated in the cloud are compared with the decision results of the corresponding tasks generated locally. Similarly, the decision results of the downlink cross task set generated locally are compared with the decision results of the corresponding tasks generated in the cloud. Difference analysis and consistency assessment are performed, such as calculating indicators like control parameter deviation, behavior consistency rate, or energy efficiency gap.

[0124] Furthermore, if cross-validation passes (e.g., error less than 5%), the current model's output control scheme is confirmed to be effective, and the decision results for both the local and cloud-based control schemes are output. If cross-validation fails, adjustment mechanisms such as model retraining, task reallocation, or control scheme fusion can be triggered. Model retraining is triggered when cross-validation shows that the model's prediction accuracy is lower than a set threshold; task reallocation is triggered when system resource allocation is unreasonable or task execution efficiency is low; and control scheme fusion is triggered when multiple alternative schemes each have their advantages, but a single scheme cannot meet all the requirements.

[0125] The specific integration of the control scheme is as follows: First, a two-way task exchange strategy is adopted. Through stratified random sampling, tasks are stratified according to their energy consumption levels, and random sampling is performed within each stratum to ensure that the coverage of high-energy-consuming tasks in cross-validation is not lower than a set threshold. Based on this, a task similarity matching mechanism is introduced, using cosine similarity metric to ensure that the complexity of the exchanged tasks matches. The calculation formula is: sim(T) i ,T j ) = (E i ·E j ) / (||E i ||·||E j ||), where sim(T) i ,T j ) represents task T i And Task T j Similarity score between them, E i For task Ti The energy consumption feature vector, E j For task T j The energy consumption feature vector.

[0126] To ensure the accuracy and performance of task execution results, the solution further constructs a three-level verification system: the first level is real-time result comparison, which requires result verification to be completed within 100 milliseconds after task execution; the second level is energy consumption simulation verification, which uses a digital twin model to predict and compare task energy consumption; and the third level is actual load testing, which selects 5% of typical nodes for load verification in a real environment to evaluate the consistency between the theoretical model and actual execution.

[0127] If energy consumption discrepancies are detected during the verification process, the system will initiate a fault handling workflow and take tiered measures based on the cross-validation energy consumption discrepancy rate ΔE: when the discrepancy rate is less than 5%, local model fine-tuning will be performed; if the discrepancy rate is between 5% and 15%, task redistribution and scheduling optimization will be performed; and when the discrepancy rate exceeds 15%, the global model retraining process will be triggered to fundamentally adjust the system strategy.

[0128] Ultimately, through the aforementioned multi-strategy collaboration and multi-layer verification mechanism, the system outputs hybrid decisions, forming a comprehensive task scheduling and energy consumption control scheme, thereby ensuring the efficient and stable operation of the system while meeting the requirements for energy consumption control and reliability.

[0129] The cloud-edge collaborative verification mechanism based on task set cross-validation introduced through the above process not only improves the reliability of the control results but also enables dynamic benchmarking and self-calibration between models. This mechanism can effectively address the accuracy degradation of edge models under specific operating conditions and can also be used to discover the adaptive deviations of cloud models in actual deployment scenarios, further enhancing the intelligence and adaptability of the entire energy-saving control system.

[0130] S5: Based on the decision results of the local control scheme and the decision results of the cloud control scheme, perform energy-saving control of the target variable frequency wind turbine cluster.

[0131] Specifically, energy-saving control of the target variable frequency wind turbine cluster is performed based on the decisions made by the local control scheme and the cloud-based control scheme: First, the control scheme decisions generated in the cloud and those generated locally are integrated. Then, the integrated control commands are sent to the corresponding IDs of the wind turbines in the target variable frequency wind turbine cluster through the automated control system. Each wind turbine adjusts its own operating parameters according to the received control commands, such as adjusting the motor frequency to change the speed or adjusting the power output to meet new task requirements.

[0132] Optionally, the system can monitor the operating status and task execution of the wind turbines in real time during the control process to dynamically adjust the control strategy. If a deviation is found between the actual operating effect of a wind turbine and the control target during the control process, the system will use this feedback information as new input data to regenerate the control plan and continue energy-saving control.

[0133] In some embodiments, step S5 further includes:

[0134] Based on a preset constraint period, the decision results of the local control scheme and the corresponding local tasks for this period are randomly selected and uploaded to the cloud node. The cloud node then makes a verification control decision based on the full-capacity energy-saving control model. Combining the decision verification constraints, the verification decision results are compared with the decision results of the local control scheme to perform decision verification, and the high-efficiency energy-saving control model is optimized accordingly based on the decision verification results.

[0135] Specifically, the constraint period refers to the time window or task round set by the system for periodically performing model verification and optimization, which is usually set according to the wind turbine operating cycle, the frequency of environmental changes, or the model update strategy.

[0136] Verification-based control decisions refer to the control results obtained by cloud nodes through re-simulating and reasoning about historical tasks that have already been controlled by edge nodes, based on a full-capacity energy-saving control model. These results are used to evaluate the accuracy of the edge model's decisions. Decision verification constraints are a set of rules used to determine the reliability of the edge model's output results. These include error thresholds, policy consistency indicators, and energy efficiency deviation rates, and are used to support triggering judgments for model updates or parameter optimization.

[0137] Specifically, after completing the energy-saving control of the target variable frequency wind turbine cluster based on the decision results of the local control scheme and the decision results of the cloud-based control scheme, the following periodic verification and optimization process is executed:

[0138] First, within each preset constraint period, a subset of samples is randomly selected from the local control scheme decision results of the current period, and the corresponding local tasks are packaged and uploaded to the cloud node to form a verification sample set. Upon receiving the verification sample set, the cloud node invokes the full-capacity energy-saving control model to re-infer these tasks, generating corresponding verification control decision results as standard reference results. Then, based on the set decision verification constraints, the control results output by the edge model are compared and analyzed with the verification control results of the cloud model to evaluate their consistency in terms of strategy parameters, target deviation, and energy efficiency indicators. If the verification results show that the deviation exceeds the allowable range, corresponding optimization of the edge-side high-efficiency energy-saving control model is triggered, including but not limited to model retraining, parameter fine-tuning, or strategy updates.

[0139] Through the above process, continuous accuracy maintenance and adaptive performance updates of the edge model can be achieved during long-term operation. This mechanism not only improves the stability and generalization ability of the edge model under changing operating conditions, but also effectively reduces the risk of control failure caused by model aging or data distribution drift, further enhancing the intelligence level and operation and maintenance controllability of the entire wind turbine cluster energy-saving system.

[0140] In some implementations, the corresponding high-efficiency energy-saving control model is optimized based on the decision verification results. The execution steps also include:

[0141] Calculate the decision residuals based on the decision verification results and construct a corresponding reinforcement training dataset; perform reinforcement learning training on the high-efficiency energy-saving control model based on the reinforcement training dataset; optimize, verify, and iterate the high-efficiency energy-saving control model after reinforcement learning until the optimization verification results meet the decision verification constraints, then update and deploy the high-efficiency energy-saving control model.

[0142] Specifically, decision residuals refer to the difference between the on-site control scheme decision results and the cloud-based confirmatory control decision results, reflecting the prediction bias of the current high-efficiency energy-saving control model in a specific task scenario. The reinforcement training dataset is a set of training samples constructed using decision residuals as the core label and combined with original task features, used for reinforcement learning training of the high-efficiency energy-saving control model.

[0143] Optimization verification is the process of evaluating the predictive performance of a model on a representative task set after reinforcement learning training. It is used to determine whether the high-efficiency energy-saving control model after reinforcement learning meets the preset decision verification constraints. These constraints include, but are not limited to, residual thresholds, energy efficiency indicators, and response time.

[0144] Specifically, after completing the decision verification, if a significant deviation is found in the high-efficiency energy-saving control model, the following optimization process is executed: For local tasks uploaded to the cloud node, the local control scheme and the cloud-based confirmatory control results are obtained respectively; the difference between the two is calculated as the decision residual for the task sample. The task samples with residual information are aggregated to construct a reinforcement training dataset, which may include task input features, current model output, cloud reference output, and residual labels. Further, the high-efficiency energy-saving control model is trained using reinforcement learning based on the reinforcement training dataset to optimize the model's policy function or state-action mapping relationship, and the trained model is optimized and verified to evaluate its residual performance and energy efficiency indicators on the validation set; if the optimization verification result does not meet the decision verification constraints, iterative training continues; once the optimization verification result meets the preset standard, the model version is automatically deployed to the local node to replace the old version model.

[0145] Through the above process, a periodic adaptive optimization mechanism for the high-efficiency energy-saving control model is realized, which helps to improve system stability.

[0146] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A cloud-based collaborative method for energy-saving control of variable frequency fan clusters, characterized in that, The cloud-based collaborative variable frequency wind turbine cluster energy-saving control method includes: Collect historical wind turbine control records, which include control parameters and control evaluations; Based on historical wind turbine control records, a full-capacity energy-saving control model is built and trained on cloud nodes, and then deployed on cloud nodes. Based on the computing power of the cluster edge nodes of the target variable frequency wind turbine cluster, the full-capacity energy-saving control model is compressed to obtain a high-efficiency energy-saving control model, which is then deployed on the cluster edge nodes. Obtain real-time task requirements and turbine operating status data for the target variable frequency wind turbine cluster; The real-time task requirement information and wind turbine operation status data are analyzed to obtain a dual-end task set, which includes a cloud task subset and a local task subset. Based on the dual-end task set, the full-capacity energy-saving control model and the high-efficiency energy-saving control model, cloud-based collaborative control scheme decision-making is carried out to obtain the local control scheme decision-making results and the cloud-based control scheme decision-making results. Energy-saving control is implemented for the target variable frequency wind turbine cluster based on the decision results of the local control scheme and the decision results of the cloud-based control scheme. in, The cloud-based collaborative control scheme decision-making based on a dual-end task set, a full-capacity energy-saving control model, and a high-efficiency energy-saving control model yields both local and cloud-based control scheme decision results, including: Based on a preset crossover ratio, the local task subset and the cloud task subset are randomly selected to obtain the uplink crossover task set and the downlink crossover task set. The uplink cross-task set and the downlink cross-task set are added to the cloud task subset and the local task subset respectively, and cloud-based collaborative control scheme decision is made. Based on the uplink cross-task set and the downlink cross-task set, cross-validate the decision results of the local control scheme and the decision results of the cloud control scheme. If the cross-validation passes, the decision results of the local control scheme and the decision results of the cloud-based control scheme will be output. If cross-validation fails, the full-capacity energy-saving control model will be retrained, and tasks will be redistributed to obtain a new dual-end task set or control scheme fusion.

2. The cloud-based collaborative variable frequency wind turbine cluster energy-saving control method according to claim 1, characterized in that, The process of building and training a full-capacity energy-saving control model on cloud nodes based on historical wind turbine control records includes: Historical wind turbine control records are filtered based on preset control evaluation thresholds; The screening results are preprocessed and uploaded to the cloud node; Using the preprocessed results as training data, a full-capacity energy-saving regulation model is constructed and trained on cloud nodes.

3. The cloud-based collaborative variable frequency wind turbine cluster energy-saving control method according to claim 1, characterized in that, The computing power of the cluster edge nodes based on the target variable frequency wind turbine cluster is used to compress the full-capacity energy-saving control model to obtain a high-efficiency energy-saving control model, including: The computing power of the cluster edge nodes is evaluated to obtain the available computing power index; Set model compression constraint parameters based on available computing power indicators; The full-capacity energy-saving control model is compressed based on the compression constraint parameters to obtain a high-efficiency energy-saving control model. Based on compression constraint parameters, a high-efficiency energy-saving control model is simulated and deployed, and the simulated computing power consumption is obtained. If the simulated computing power consumption meets the available computing power index, then a high-efficiency energy-saving control model is output. If the simulated computing power consumption does not meet the available computing power index, the constraint parameters are adjusted and the model is compressed again until the simulated computing power consumption does not meet the available computing power index.

4. The cloud-based collaborative variable frequency wind turbine cluster energy-saving control method according to claim 1, characterized in that, The process of parsing real-time task requirements and wind turbine operating status data to obtain a dual-end task set includes: The real-time task requirements information is decomposed to obtain a real-time task set. Using the wind turbine ID as an index, establish a mapping relationship between real-time task sets and historical wind turbine control records; The complexity of the control decision for each real-time task is obtained by calling historical wind turbine control records based on the mapping relationship. Based on the complexity of control decisions and the preset real-time requirements, the real-time task set is divided into a cloud task subset and a local task subset, resulting in a dual-end task set.

5. The cloud-based collaborative variable frequency wind turbine cluster energy-saving control method according to claim 4, characterized in that, The process of parsing real-time task requirements information and wind turbine operating status data to obtain a dual-end task set also includes: After obtaining the real-time task set, cluster analysis is performed on the real-time task set based on the model parameters of the variable frequency fan and the fan operating status data to generate a simplified task set; The simplified task set will be used as the new real-time task set.

6. The cloud-based collaborative variable frequency wind turbine cluster energy-saving control method according to claim 1, characterized in that, The cloud-based collaborative control scheme decision-making based on a dual-end task set, a full-capacity energy-saving control model, and a high-efficiency energy-saving control model yields both local and cloud-based control scheme decision results, including: A subset of local tasks will be transmitted to a high-efficiency energy-saving control model to make local control scheme decisions, thereby generating local control scheme decision results; A subset of cloud tasks is synchronously uploaded to the full-capacity energy-saving control model of the cloud node to execute cloud control scheme decisions, thereby generating the cloud control scheme decision results.

7. The cloud-based collaborative variable frequency wind turbine cluster energy-saving control method according to claim 1, characterized in that, The energy-saving control of the target variable frequency wind turbine cluster based on the decision results of the local control scheme and the decision results of the cloud-based control scheme includes: The decision-making results of local control plans will be integrated with those of cloud-based control plans. The integrated control commands are sent to the corresponding ID of the fan in the target variable frequency fan cluster; Each wind turbine adjusts its own operating parameters according to the control instructions it receives.

8. The cloud-based collaborative variable frequency wind turbine cluster energy-saving control method according to claim 1, characterized in that, Also includes: Based on a preset constraint period, the decision results of the local control scheme and the corresponding local tasks for the current constraint period are randomly extracted and uploaded to the cloud node. Verification and control decisions are made by cloud nodes based on a full-capacity energy-saving control model; Based on the preset decision verification constraints, the decision verification results are compared with the decision results of the local control scheme to obtain the decision verification results; Optimize the corresponding high-efficiency energy-saving control model based on the decision verification results.

9. The cloud-based collaborative variable frequency wind turbine cluster energy-saving control method according to claim 8, characterized in that, The optimization of the corresponding high-efficiency energy-saving control model based on the decision verification results includes: Calculate the decision residuals based on the decision verification results, and construct a reinforcement training dataset based on the decision residuals; The high-efficiency energy-saving control model is trained using a reinforcement training dataset; The high-efficiency energy-saving control model after reinforcement learning is optimized, verified, and iterated until the optimization and verification results meet the decision verification constraints. Update and deploy high-efficiency energy-saving control models.

Images